cs.LG — Pith

0

cs.LG 2026-05-13 2 theorems

Orthogonal transformations keep weight singular values fixed

Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation

Pion replaces additive steps with left and right rotations that fix spectral norm while still allowing matrix geometry to change during LLM

abstract click to expand

We introduce Pion, a spectrum-preserving optimizer for large language model (LLM) training based on orthogonal equivalence transformation. Unlike additive optimizers such as Adam and Muon, Pion updates each weight matrix through left and right orthogonal transformations, preserving its singular values throughout training. This yields an optimization mechanism that modulates the geometry of weight matrices while keeping their spectral norm fixed. We derive the Pion update rule, systematically examine its design choices, and analyze its convergence behavior along with several key properties. Empirical results show that Pion offers a stable and competitive alternative to standard optimizers for both LLM pretraining and finetuning.

0

cs.LG 2026-05-13 2 theorems

Fast-slow training triples sample efficiency for LLM adaptation

Learning, Fast and Slow: Towards LLMs That Adapt Continually

Treating context as temporary fast weights keeps parameters stable, cuts forgetting, and lets models keep learning when tasks change.

abstract click to expand

Large language models (LLMs) are trained for downstream tasks by updating their parameters (e.g., via RL). However, updating parameters forces them to absorb task-specific information, which can result in catastrophic forgetting and loss of plasticity. In contrast, in-context learning with fixed LLM parameters can cheaply and rapidly adapt to task-specific requirements (e.g., prompt optimization), but cannot by itself typically match the performance gains available through updating LLM parameters. There is no good reason for restricting learning to being in-context or in-weights. Moreover, humans also likely learn at different time scales (e.g., System 1 vs 2). To this end, we introduce a fast-slow learning framework for LLMs, with model parameters as "slow" weights and optimized context as "fast" weights. These fast "weights" can learn from textual feedback to absorb the task-specific information, while allowing slow weights to stay closer to the base model and persist general reasoning behaviors. Fast-Slow Training (FST) is up to 3x more sample-efficient than only slow learning (RL) across reasoning tasks, while consistently reaching a higher performance asymptote. Moreover, FST-trained models remain closer to the base LLM (up to 70% less KL divergence), resulting in less catastrophic forgetting than RL-training. This reduced drift also preserves plasticity: after training on one task, FST trained models adapt more effectively to a subsequent task than parameter-only trained models. In continual learning scenarios, where task domains change on the fly, FST continues to acquire each new task while parameter-only RL stalls.

0

cs.LG 2026-05-13 Recognition

RL on bigger teachers then dense distillation beats direct GRPO on small students

Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training

With scarce verifiable data, apply sparse rewards upstream for discovery and dense transfer downstream for compression to improve 1.7B math

abstract click to expand

In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated carefully. The standard practice is to use this data directly on the model that will be deployed, for example by running GRPO on the deployment student. We argue that this is often an inefficient allocation because it overlooks a reward-density principle: sparse sequence-level reward should train models where exploration is productive, while dense token-level teacher reward should be used where the aim is to compress behavior into a smaller model. In this view, GRPO-style sparse RL and OPD-style dense teacher supervision are not separate recipes; they are different reward-density regimes. The allocation rule is simple: use scarce labeled training data upstream on the strongest model that can turn it into reward-shaped behavior, then transfer that behavior downstream as dense supervision. We evaluate this rule on verifiable math with Qwen3 and Llama models. At fixed Qwen3-1.7B deployment-student size, an RL-improved 8B teacher distilled through the dense bridge outperforms direct GRPO on the same student, while transfer from the same teacher before RL underperforms. The bridge is important: a forward-KL warmup on teacher rollouts followed by OPD on student rollouts is consistently strongest on MATH before any post-bridge student-side sparse RL, and also gives the best pre-Stage~3 AIME endpoints for the canonical 8B/14B teachers. The bridge also makes later student-side sparse RL effective: GRPO that is weak on a cold student lifts MATH from $75.4\%$ to $78.5\%$ after the bridge and outperforms a matched replay control by $2.8$ points. The operational principal is to avoid using scarce labeled data on the least prepared policy: use sparse reward for teacher-side discovery, dense transfer for student compression, and student-side sparse reward only after the bridge.

0

cs.LG 2026-05-13 Recognition

LLM memory systems fail dependency reasoning across evolving entities

MEME: Multi-entity & Evolving Memory Evaluation

All six tested systems score 3% and 1% on Cascade and Absence tasks despite adequate static retrieval, and fixes cost 70x more.

abstract click to expand

LLM-based agents increasingly operate in persistent environments where they must store, update, and reason over information across many sessions. While prior benchmarks evaluate only single-entity updates, MEME defines six tasks spanning the full space defined by the multi-entity and evolving axes, including three not scored by prior work: Cascade and Absence (dependency reasoning) and Deletion (post-removal state). Evaluating six memory systems spanning three memory paradigms on 100 controlled episodes, we find that all systems collapse on dependency reasoning under the default configuration (Cascade: 3%, Absence: 1% in average accuracy) despite adequate static retrieval performance. Prompt optimization, deeper retrieval, reduced filler noise, and most stronger LLMs fail to close this gap. Only a file-based agent paired with Claude Opus 4.7 as its internal LLM partially closes the gap, but at ~70x the baseline cost, indicating closure currently depends on configurations that are not practical at scale. Code and data are available on the project page: https://seokwonjung-jay.github.io/meme-eval/.

0

cs.LG 2026-05-13 2 theorems

Routers align geometrically with experts they activate

Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts

This shared gradient direction supports better load balance than auxiliary losses, shown by a K-means router with small perplexity trade-off

abstract click to expand

Sparse Mixture-of-Experts (SMoE) models enable scaling language models efficiently, but training them remains challenging, as routing can collapse onto few experts and auxiliary load-balancing losses can reduce specialization. Motivated by these hurdles, we study how routing decisions in SMoEs are formed mechanistically. First, we reveal a geometric coupling between routers and their corresponding experts. For a given token, the router weights for the selected expert and the expert weights processing it receive gradients along the same input direction, differing only in scalar coefficients. Thus, matched router--expert directions accumulate the same routed token history. This theoretical coupling also appears empirically in routing dynamics. In a $1$B SMoE trained from scratch, higher router scores predict stronger expert neuron activations, showing that routing decisions are mirrored inside the selected expert. Next, we analyze the effects of auxiliary load balancing on the router--expert geometric coupling, showing that such losses break this structure by spreading input-directed gradients across router weights, making distinct router directions nearly three times more similar to each other. Last, we demonstrate the centrality of geometric coupling for effective routing with a parameter-free online K-Means router, in which each expert maintains a running average of the hidden states routed to it and tokens are assigned based on cosine similarity. Compared with auxiliary-loss and loss-free balancing, this router achieves the lowest load imbalance with only a modest perplexity increase, indicating that geometric coupling captures a substantial part of what the router learns. Overall, our results explain how routers form assignment geometry that supports an effective division of labor.

0

cs.LG 2026-05-13 Recognition

Pretrained transformers handle 128K contexts via KV-cache folding

KV-Fold: One-Step KV-Cache Recurrence for Long-Context Inference

One-step cache updates across chunks preserve exact long-range retrieval on a single GPU without retraining

abstract click to expand

We introduce KV-Fold, a simple, training-free long-context inference protocol that treats the key-value (KV) cache as the accumulator in a left fold over sequence chunks. At each step, the model processes the next chunk conditioned on the accumulated cache, appends the newly produced keys and values, and passes the enlarged cache forward; the same one-step update is applied repeatedly, analogous to foldl in functional programming. Building on the KV cache concatenation primitive introduced for latent multi-agent communication, we repurpose it as a chunk-to-chunk recurrence for long-context inference. When processing chunk t, the model attends to the KV cache carried from earlier chunks as a prefix, reusing its internal state across segments without modifying or retraining the model. Despite its simplicity, the induced recurrence is stable: per-step drift rises briefly and then saturates into a flat plateau that persists across deep chains. This plateau is insensitive to a 10,000x change in numerical precision, robust across chunk sizes, and consistent across model families. At the task level, KV-Fold preserves exact information over long distances. On a needle-in-a-haystack benchmark, it achieves 100% exact-match retrieval across 152 trials spanning contexts from 16K to 128K tokens and chain depths up to 511 on Llama-3.1-8B, while remaining within the memory limits of a single 40GB GPU. Compared to streaming methods, which trade fidelity for bounded memory, KV-Fold maintains long-range retrieval while operating as a sequence of tractable forward passes. Overall, our results show that frozen pretrained transformers already support a stable form of KV-cache recurrence, providing a practical route to long-context inference without architectural changes or training.

0

cs.LG 2026-05-13 2 theorems

Attractor models beat larger transformers on language and puzzles

Solve the Loop: Attractor Models for Language and Reasoning

Fixed-point solving keeps training memory constant while letting small models reach over 90 percent accuracy on hard reasoning tasks.

abstract click to expand

Looped Transformers offer a promising alternative to purely feed-forward computation by iteratively refining latent representations, improving language modeling and reasoning. Yet recurrent architectures remain unstable to train, costly to optimize and deploy, and constrained to small, fixed recurrence depths. We introduce Attractor Models, in which a backbone module first proposes output embeddings, then an attractor module refines them by solving for the fixed point, with gradients obtained through implicit differentiation. Thus, training memory remains constant in effective depth, and iterations are chosen adaptively by convergence. Empirically, Attractor Models outperform existing models across two regimes, large-scale language-model pretraining and reasoning with tiny models. In language modeling, Attractor Models deliver a Pareto improvement over standard Transformers and stable looped models across sizes, improving perplexity by up to 46.6% and downstream accuracy by up to 19.7% while reducing training cost. Notably, a 770M Attractor Model outperforms a 1.3B Transformer trained on twice as many tokens. On challenging reasoning tasks, we show that our model with only 27M parameters and approximately 1000 examples achieves 91.4% accuracy on Sudoku-Extreme and 93.1% on Maze-Hard, scaling favorably where frontier models like Claude and GPT o3, fail completely, and specialized recursive reasoners collapse at larger sizes. Lastly, we show that Attractor Models exhibit a novel phenomenon, which we call equilibrium internalization: fixed-point training places the model's initial output embedding near equilibrium, allowing the solver to be removed at inference time with little degradation. Together, these results suggest that Attractor Models make iterative refinement scalable by turning recurrence into a computation the model can learn to internalize.

0

cs.LG 2026-05-13 Recognition

High-arity compression implies PAC learnability

High-arity Sample Compression

A non-trivial-quality high-arity sample compression scheme guarantees that the concept class is PAC learnable over product spaces.

abstract click to expand

Recently, a series of works have started studying variations of concepts from learning theory for product spaces, which can be collected under the name high-arity learning theory. In this work, we consider a high-arity variant of sample compression schemes and we prove that the existence of a high-arity sample compression scheme of non-trivial quality implies high-arity PAC learnability.

0

cs.LG 2026-05-13 Recognition

Block-scale search cuts quantization error 27% in BFP

Search Your Block Floating Point Scales!

Fine-grained search using mantissa bits improves PTQ accuracy and enables near-lossless low-precision attention.

abstract click to expand

Quantization has emerged as a standard technique for accelerating inference for generative models by enabling faster low-precision computations and reduced memory transfers. Recently, GPU accelerators have added first-class support for microscaling Block Floating Point (BFP) formats. Standard BFP algorithms use a fixed scale based on the maximum magnitude of the block. We observe that this scale choice can be suboptimal with respect to quantization errors. In this work, we propose ScaleSearch, an alternative strategy for selecting these scale factors: using a fine-grained search leveraging the mantissa bits in microscaling formats to minimize the quantization error for the given distribution. ScaleSearch can be integrated with existing quantization methods such as Post Training Quantization and low-precision attention, and is shown to improve their performance. Additionally, we introduce ScaleSearchAttention, an accelerated NVFP4-based attention algorithm, which uses ScaleSearch and adapted prior techniques to ensure near-0 performance loss for causal language modeling. Experiments show that ScaleSearch reduces quantization error by 27% for NVFP4 and improves language model PTQ by up to 15 points for MATH500 (Qwen3-8B), while ScaleSearchAttention improves Wikitext-2 PPL by upto 0.77 points for Llama 3.1 70B. The proposed methods closely match baseline performance while providing quantization accuracy improvements.

0

cs.LG 2026-05-13 2 theorems

Parallel streams let models read while writing

Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs

Training language models on multiple simultaneous streams removes the single-sequence limit that blocks agents from thinking and acting at 0

abstract click to expand

The continued improvements in language model capability have unlocked their widespread use as drivers of autonomous agents, for example in coding or computer use applications. However, the core of these systems has not changed much since early instruction-tuned models like ChatGPT. Even advanced AI agents function on message exchange formats, successively exchanging messages with users, systems, with itself (i.e. chain-of-thought) and tools in a single stream of computation. This bottleneck to a single stream in chat models leads to a number of limitations: the agent cannot act (generate output) while reading, and in reverse, cannot react to new information while writing. Similarly, the agent cannot act while thinking and cannot think while reading or acting on information. In this work, we show that models can be unblocked by switching from instruction-tuning for sequential message formats to instruction-tuning for multiple, parallel streams of computation, splitting each role into a separate stream. Every forward pass of the language model then simultaneously reads from multiple input streams and generates tokens in multiple output streams, all of which causally depend on earlier timesteps. We argue that this data-driven change remedies a number of usability limitations as outlined above, improves model efficiency through parallelization, improves model security through better separation of concerns and can further improve model monitorability.

0

cs.LG 2026-05-13 2 theorems

Decoupled method aligns verbalized confidence in LLMs

ORCE: Order-Aware Alignment of Verbalized Confidence in Large Language Models

New framework improves calibration and failure detection on reasoning tasks without hurting answer accuracy.

abstract click to expand

Large language models (LLMs) often produce answers with high certainty even when they are incorrect, making reliable confidence estimation essential for deployment in real-world scenarios. Verbalized confidence, where models explicitly state their confidence in natural language, provides a flexible and user-facing uncertainty signal that can be applied even when token logits are unavailable. However, existing verbalized-confidence methods often optimize answer generation and confidence generation jointly, which can cause confidence-alignment objectives to interfere with answer accuracy. In this work, we propose a decoupled and order-aware framework for verbalized confidence calibration. Our method first generates an answer and then estimates confidence conditioned on the fixed question--answer pair, allowing confidence optimization without directly perturbing the answer-generation process. To align confidence with correctness likelihood, we construct a sampling-based surrogate from multiple model completions and optimize rank-based reinforcement learning objectives that encourage responses with higher estimated correctness likelihood to receive higher verbalized confidence. Experiments on reasoning and knowledge-intensive benchmarks show that our method improves calibration and failure prediction performance while largely preserving answer accuracy. These results demonstrate that verbalized confidence can be more reliably aligned by decoupling confidence estimation from answer generation and optimizing the relative ordering of confidence across responses.

0

cs.LG 2026-05-13 2 theorems

Adaptive optimization reaches 0.73 AUC on shifted wildfire data

Environment-Adaptive Preference Optimization for Wildfire Prediction

Aligning datasets to target environments and emphasizing rare events improves detection reliability under changing conditions.

abstract click to expand

Predicting rare extreme events such as wildfires from meteorological data requires models that remain reliable under evolving environmental conditions. This problem is inherently long-tailed: wildfire events are rare but high-impact, while most observations correspond to non-fire conditions, causing standard learning objectives to underemphasize the minority class (fire) that matters most. In addition, models trained on historical distributions often fail under distribution shifts, exhibiting degraded performance in new environments. To this end, we propose Environment-Adaptive Preference Optimization (EAPO), a framework that adapts prediction to the target environment with long-tail distribution. Given a new input distribution, we first construct distribution-aligned datasets via $k$-nearest neighbor retrieval. We then perform a hybrid fine-tuning procedure on this local manifold, combining supervised learning with preference optimization, as well as emphasizing on rare extreme events. EAPO refines decision boundaries while avoiding conflicting signals from heterogeneous training data. We evaluate EAPO on a real-world wildfire prediction task with environmental shifts. EAPO achieves robust performance (ROC-AUC 0.7310) and improves detection in extreme regimes, demonstrating its effectiveness in dynamic wildfire prediction systems.

0

cs.LG 2026-05-13 Recognition

RL discovers rigid graphs with record realization counts

Learning Minimally Rigid Graphs with High Realization Counts

A neural policy trained on Henneberg extensions matches planar optima and sets new spherical records.

abstract click to expand

For minimally rigid graphs, the same edge-length data can admit multiple realizations (up to translations and rotations). Finding graphs with exceptionally many realizations is an extremal problem in rigidity theory, but exhaustive search quickly becomes infeasible due to the super-exponential growth of the number of candidate graphs and the high cost of realization-count evaluation. We propose a reinforcement-learning approach that constructs minimally rigid graphs via 0- and 1-extensions, also known as Henneberg moves. We optimize realization-count invariants using the Deep Cross-Entropy Method with a policy parameterized by a Graph Isomorphism Network encoder and a permutation-equivariant extension-level action head. Empirically, our method matches the known optima for planar realization counts and improves the best known bounds for spherical realization counts, yielding new record graphs.

0

cs.LG 2026-05-13 2 theorems

Closed-form Q-guidance aligns flow map policies for online RL

Aligning Flow Map Policies with Optimal Q-Guidance

Deriving an exact optimal target lets offline flow policies improve Q-values while jumping directly across generative dynamics.

abstract click to expand

Generative policies based on expressive model classes, such as diffusion and flow matching, are well-suited to complex control problems with highly multimodal action distributions. Their expressivity, however, comes at a significant inference cost: generating each action typically requires simulating many steps of the generative process, compounding latency across sequential decision-making rollouts. We introduce flow map policies, a novel class of generative policies designed for fast action generation by learning to take arbitrary-size jumps including one-step jumps-across the generative dynamics of existing flow-based policies. We instantiate flow map policies for offline-to-online reinforcement learning (RL) and formulate online adaptation as a trust-region optimization problem that improves the critic's Q-value while remaining close to the offline policy. We theoretically derive FLOW MAP Q-GUIDANCE (FMQ), a principled closed-form learning target that is optimal for adapting offline flow map policies under a critic-guided trust-region constraint. We further introduce Q-GUIDED BEAM SEARCH (QGBS), a stochastic flow-map sampler that combines renoising with beam search to enable iterative inference-time refinement. Across 12 challenging robotic manipulation and locomotion tasks from OGBench and RoboMimic, FMQ achieves state-of-the-art performance in offline-to-online RL, outperforming the previous one-step policy MVP by a relative improvement of 21.3% on the average success rate.

0

cs.LG 2026-05-13 Recognition

Tabular model predicts AI agents' moves from 16 past games

Predicting Decisions of AI Agents from Limited Interaction through Text-Tabular Modeling

Hidden LLM states improve accuracy over direct prompting in controlled bargaining and negotiation tasks.

abstract click to expand

AI agents negotiate and transact in natural language with unfamiliar counterparts: a buyer bot facing an unknown seller, or a procurement assistant negotiating with a supplier. In such interactions, the counterpart's LLM, prompts, control logic, and rule-based fallbacks are hidden, while each decision can have monetary consequences. We ask whether an agent can predict an unfamiliar counterpart's next decision from a few interactions. To avoid real-world logging confounds, we study this problem in controlled bargaining and negotiation games, formulating it as target-adaptive text-tabular prediction: each decision point is a table row combining structured game state, offer history, and dialogue, while $K$ previous games of the same target agent, i.e., the counterpart being modeled, are provided in the prompt as labeled adaptation examples. Our model is built on a tabular foundation model that represents rows using game-state features and LLM-based text representations, and adds LLM-as-Observer as an additional representation: a small frozen LLM reads the decision-time state and dialogue; its answer is discarded, and its hidden state becomes a decision-oriented feature, making the LLM an encoder rather than a direct few-shot predictor. Training on 13 frontier-LLM agents and testing on 91 held-out scaffolded agents, the full model outperforms direct LLM-as-Predictor prompting and game+text features baselines. Within this tabular model, Observer features contribute beyond the other feature schemes: at $K=16$, they improve response-prediction AUC by about 4 points across both tasks and reduce bargaining offer-prediction error by 14%. These results show that formulating counterpart prediction as a target-adaptive text-tabular task enables effective adaptation, and that hidden LLM representations expose decision-relevant signals that direct prompting does not surface.

0

cs.LG 2026-05-13 2 theorems

Outcome rewards steer logits to cut bias in LLM self-distillation

OGLS-SD: On-Policy Self-Distillation with Outcome-Guided Logit Steering for LLM Reasoning

Contrasting successful and failed trajectories calibrates teacher signals and raises reasoning accuracy over plain on-policy distillation.

abstract click to expand

We study {on-policy self-distillation} (OPSD), where a language model improves its reasoning ability by distilling privileged teacher distributions along its own on-policy trajectories. Despite the performance gains of OPSD, we identify a common but often overlooked mismatch between teacher and student responses: self-reflected teacher responses can be shifted by reflection-induced bias and response templates, leading to miscalibrated token-level supervision. To mitigate this issue, we propose \methodname, an outcome-guided logit-steering framework that leverages verifiable outcome rewards to contrast successful and failed on-policy trajectories and calibrate teacher logits. By combining outcome-level correctness with dense token-level guidance through logit steering, \methodname stabilizes self-distillation and improves reasoning performance over standard OPSD and other variants across diverse benchmarks.

0

cs.LG 2026-05-13 Recognition

Correlation Traps in weights mark start of overfitting in grokking

Detecting overfitting in Neural Networks during long-horizon grokking using Random Matrix Theory

Random-matrix outliers grow as test accuracy falls while train accuracy stays high, and appear in some large language models too.

abstract click to expand

Training Neural Networks (NNs) without overfitting is difficult; detecting that overfitting is difficult as well. We present a novel Random Matrix Theory method that detects the onset of overfitting in deep learning models without access to train or test data. For each model layer, we randomize each weight matrix element-wise, $\mathbf{W} \to \mathbf{W}_{\mathrm{rand}}$, fit the randomized empirical spectral distribution with a Marchenko-Pastur distribution, and identify large outliers that violate self-averaging. We call these outliers Correlation Traps. During the onset of overfitting, which we call the "anti-grokking" phase in long-horizon grokking, Correlation Traps form and grow in number and scale as test accuracy decreases while train accuracy remains high. Traps may be benign or may harm generalization; we provide an empirical approach to distinguish between them by passing random data through the trained model and evaluating the JS divergence of output logits. Our findings show that anti-grokking is an additional grokking phase with high train accuracy and decreasing test accuracy, structurally distinct from pre-grokking through its Correlation Traps. More broadly, we find that some foundation-scale LLMs exhibit the same Correlation Traps, indicating potentially harmful overfitting.

0

cs.LG 2026-05-13 2 theorems

Batch statistic replaces fixed clipping in RL policy training

Trust the Batch, On- or Off-Policy: Adaptive Policy Optimization for RL Post-Training

The normalized effective sample size of policy ratios in each batch sets its own trust-region and off-policy controls automatically.

abstract click to expand

Reinforcement learning is structurally harder than supervised learning because the policy changes the data distribution it learns from. The resulting fragility is especially visible in large-model training, where the training and rollout systems differ in numerical precision, sampling, and other implementation details. Existing methods manage this fragility by adding hyper-parameters to the training objective, which makes the algorithm more sensitive to its configuration and requires retuning whenever the task, model scale, or distribution mismatch changes. This fragility traces to two concerns that current objectives entangle through hyper-parameters set before training begins: a trust-region concern, that updates should not move the policy too far from its current value, and an off-policy concern, that data from older or different behavior policies should influence the update only to the extent that it remains reliable. Neither concern is a constant to set in advance, and their severity is reflected in the policy-ratio distribution of the current batch. We present a simple yet effective batch-adaptive objective that replaces fixed clipping with the normalized effective sample size of the policy ratios. The same statistic caps the score-function weight and sets the strength of an off-policy regularizer, so the update stays close to the usual on-policy score-function update when ratios are nearly uniform, and tightens automatically when stale or mismatched data cause ratio concentration, while retaining a nonzero learning signal on high-ratio tokens. Experiments across a wide range of settings show that our method matches or exceeds tuned baselines, introducing no new objective hyper-parameters and removing several existing ones. The code is available at https://github.com/FeynRL-project/FeynRL.

0

cs.LG 2026-05-13 2 theorems

Path-space penalty keeps RL policies stable from offline to online

Discrete Flow Matching for Offline-to-Online Reinforcement Learning

DRIFT combines discrete flow matching with trajectory regularization and candidate sampling to improve without losing prior gains.

abstract click to expand

Many reinforcement learning (RL) tasks have discrete action spaces, but most generative policy methods based on diffusion and flow matching are designed for continuous control. Meanwhile, generative policies usually rely heavily on offline datasets and offline-to-online RL is itself challenging, as the policy must improve from new interaction without losing useful behavior learned from static data. To address those challenges, we introduce DRIFT, an online fine-tuning method that updates an offline pretrained continuous-time Markov chain (CTMC) policy with an advantage-weighted discrete flow matching loss. To preserve useful pretrained knowledge, we add a path-space penalty that regularizes the full CTMC trajectory distribution, rather than only the final action distribution. For large discrete action spaces, we introduce a candidate-set approximation that updates the actor over a small subset of actions sampled from reference-policy rollouts and uniform exploration. Our theoretical analysis shows that the candidate-set error is controlled by missing target probability mass, and the induced CTMC generator error decreases as the candidate set covers more high-probability actions. Experiments on prevailing discrete action RL task show that our method provides stable offline-to-online improvement across all tasks, achieving the highest average score on Jericho with a simple GRU encoder while outperforming methods that use pretrained language models. Controlled experiments further confirm that the path-space penalty remains bounded during fine-tuning and that the CTMC generator adapts to shifted rewards faster than deterministic baselines. The candidate-set mechanism is supported by a stability analysis showing that the generator error decreases exponentially with candidate coverage.

0

cs.LG 2026-05-13 2 theorems

LLM agent cuts strawberry yield forecast errors by up to 56 percent

Agent-Based Post-Hoc Correction of Agricultural Yield Forecasts

Post-hoc tools for phase detection and bias learning improve XGBoost and other models on standard farm records for both strawberries and US-

abstract click to expand

Accurate crop yield forecasting in commercial soft fruit production is constrained by the data available in typical commercial farm records, which lack the sensor networks, satellite imagery, and high-resolution meteorological inputs that most state-of-the-art approaches assume. We propose a structured LLM agent framework that performs post-hoc correction of existing model predictions, encoding agricultural domain knowledge across tools for phase detection, bias learning, and range validation. Evaluated on a proprietary strawberry yield dataset and a public USDA corn harvest dataset, agent refinement of XGBoost reduced MAE by 20% and MASE by 56% on strawberry, with consistent improvements across Moirai2 (MAE 24%, MASE 22%) and Random Forest (MAE 28%, MASE 66%) baselines. Using Llama 3.1 8B as the agent produced the strongest corrections across all configurations; LLaVA 13B showed inconsistent gains, highlighting sensitivity to the choice of refinement model.

0

cs.LG 2026-05-13 2 theorems

Meta-trained network solves PDEs with one linear solve

MetaColloc: Optimization-Free PDE Solving via Meta-Learned Basis Functions

Basis functions are learned once offline on random fields and then frozen, turning each new equation into a fast least-squares problem.

abstract click to expand

Solving partial differential equations (PDEs) with machine learning typically requires training a new neural network for every new equation. This optimization is slow. We introduce MetaColloc. It is an optimization-free and data-free framework that removes this bottleneck completely. We decouple basis discovery from the solving process. We meta-train a dual-branch neural network on diverse Gaussian Random Fields. This offline process creates a universal dictionary of neural basis functions. At test time, we freeze the network. We solve the PDE by assembling a collocation matrix. We find the solution through a single linear least squares step. For non-linear PDEs, we apply the Newton-Raphson method to achieve fast quadratic convergence. Our experiments across six 2D and 3D PDEs show massive improvements. MetaColloc reaches state-of-the-art accuracy on smooth and non-linear problems. It also reduces test-time computation by several orders of magnitude. Finally, we provide a detailed frequency sweep analysis. This analysis reveals a critical mismatch between function approximation and operator stability at extremely high frequencies. This profound finding opens a clear path toward future operator-aware meta-learning.

0

cs.LG 2026-05-13 Recognition

Sequence models recast graphs by separating processing from propagation depth

From Message-Passing to Linearized Graph Sequence Models

Core graph decisions become sequence modeling choices, improving long-range information flow while keeping structural bias.

abstract click to expand

Message-passing based approaches form the default backbone of most learning architectures on graph-structured data. However, the rapid progress of modern deep learning architectures in other domains, particularly sequence modeling, raises the question of how graph learning can benefit from these advances. We introduce Linearized Graph Sequence Models, a framework that recasts message-passing graph computation from the perspective of sequence modeling to simplify architectural choices. Our approach systematically separates the computational processing depth from the information propagation depth, allowing core graph architectural decisions to be treated as sequence modeling choices. Specifically, we analyze, both empirically and theoretically, what sequence properties make methods effective for learning and preserving the graph inductive bias. In particular, we validate our findings, demonstrating improved performance on long-range information tasks in graphs. Our findings provide a principled way to integrate modern sequence modeling advances into message-passing based graph learning. Beyond this, our work demonstrates how the separation of processing and information depth can recast central architectural questions as input modeling choices.

0

cs.LG 2026-05-13 Recognition

The paper proposes FAMeX, a new explainable AI algorithm that builds a Feature…

A New Technique for AI Explainability using Feature Association Map

FAMeX introduces a graph-theoretic Feature Association Map to explain feature importance in AI classification models and outperforms PFI…

abstract click to expand

Lack of transparency in AI systems poses challenges in critical real-life applications. It is important to be able to explain the decisions of an AI system to ensure trust on the system. Explainable AI (XAI) algorithms play a vital role in achieving this objective. In this paper, we are proposing a new algorithm for Explaining AI systems, FAMeX (Feature Association Map based eXplainability). The proposed algorithm is based on a graph-theoretic formulation of the feature set termed as Feature Association Map (FAM). The foundation of the modelling is based on association between features. The proposed FAMeX algorithm has been found to be better than the competing XAI algorithms - Permutation Feature Importance (PFI) and SHapley Additive exPlanations (SHAP). Experiments conducted with eight benchmark algorithms show that FAMeX is able to gauge feature importance in the context of classification better than the competing algorithms. This definitely shows that FAMeX is a promising algorithm in explaining the predictions from an AI system

0

cs.LG 2026-05-13 Recognition

Local 3x3x3 neural patches plus Schwarz iteration solve large 3D PDEs

Neural-Schwarz Tiling for Geometry-Universal PDE Solving at Scale

The same trained local solver works on new domains and scales by delegating global consistency to classical domain decomposition.

abstract click to expand

Most learned PDE solvers follow a global-surrogate paradigm: a neural operator is trained to map full problem descriptions to full solution fields for a prescribed distribution of geometries, boundary conditions, and coefficients. This has enabled fast inference within fixed problem families, but limits reuse across new domains and makes large-scale deployment dependent on expensive problem-specific data generation. We introduce $\textbf{NEST}$ ($\textbf{Ne}$ural-$\textbf{S}$chwarz $\textbf{T}$iling), a local-to-global framework that shifts learning from full-domain solution operators to reusable local physical solvers. The central premise is that, although global PDE solutions depend on geometry, scale, and boundary conditions, the physical response on small neighborhoods can be learned locally and composed into global solutions through classical domain decomposition. NEST learns a neural operator on minimal voxel patches ($3 \times 3 \times 3$) with diverse local geometries and boundary/interface data. At inference time, an unseen voxelized domain is tiled into overlapping patches, the learned local solver is applied patchwise, and global consistency is enforced through iterative Schwarz coupling with partition-of-unity assembly. In this way, generalization is shifted from a monolithic neural model to the combination of local physics learning and algorithmic global assembly. We instantiate NEST on nonlinear static equilibrium in compressible neo-Hookean solids and evaluate it on large, geometrically complex 3D domains far outside the scale of the training patches. Our results show that local neural building blocks, coupled through Schwarz iteration, offer a reusable local-training path toward scalable learned PDE solvers that generalize across domain size, shape, and boundary-condition configurations.

0

cs.LG 2026-05-13 2 theorems

Safety alignment reduces to density ratio matching

BSO: Safety Alignment Is Density Ratio Matching

A closed-form ratio decomposition yields single-stage losses that recover the optimal safe policy without auxiliary models.

abstract click to expand

Aligning language models for both helpfulness and safety typically requires complex pipelines-separate reward and cost models, online reinforcement learning, and primal-dual updates. Recent direct preference optimization approaches simplify training but incorporate safety through ad-hoc modifications such as multi-stage procedures or heuristic margin terms, lacking a principled derivation. We show that the likelihood ratio of the optimal safe policy admits a closed-form decomposition that reduces safety alignment to a density ratio matching problem. Minimizing Bregman divergences between the data and model ratios yields Bregman Safety Optimization (BSO), a family of single-stage loss functions, each induced by a convex generator, that provably recover the optimal safe policy. BSO is both general and simple: it requires no auxiliary models, introduces only one hyperparameter beyond standard preference optimization, and recovers existing safety-aware methods as special cases. Experiments across safety alignment benchmarks show that BSO consistently improves the safety-helpfulness trade-off.

0

cs.LG 2026-05-13 2 theorems

Entropy-maximizing resampling mixes across disconnected manifolds

Manifold Sampling via Entropy Maximization

MASEM decreases KL divergence exponentially in the mean field and improves Sinkhorn distance by an order of magnitude on benchmarks.

abstract click to expand

Sampling from constrained distributions has a wide range of applications, including in Bayesian optimization and robotics. Prior work establishes convergence and feasibility guarantees for constrained sampling, but assumes that the feasible set is connected. However, in practice, the feasible set often decomposes into multiple disconnected components, which makes efficient sampling under constraints challenging. In this paper, we propose MAnifold Sampling via Entropy Maximization (MASEM) for sampling on a manifold with an unknown number of disconnected components, implicitly defined by smooth equality and inequality constraints. The presented method uses a resampling scheme to maximize the entropy of the empirical distribution based on k-nearest neighbor density estimation. We show that, in the mean field, MASEM decreases the KL-divergence between the empirical distribution and the maximum-entropy target exponentially in the number of resampling steps. We instantiate MASEM with multiple local samplers and demonstrate its versatility and efficiency on synthetic and robotics-based benchmarks. MASEM enables fast and scalable mixing across a range of constrained sampling problems, improving over alternatives by an order of magnitude in Sinkhorn distance with competitive runtime.

0

cs.LG 2026-05-13 2 theorems

Multiple grids per group boost 4-bit LLM accuracy

Grid Games: The Power of Multiple Grids for Quantizing Large Language Models

Choosing the best 4-bit grid from a small set for each weight group cuts quantization error compared with fixed grids in both training and 4

abstract click to expand

A major recent advance in quantization is given by microscaled 4-bit formats such as NVFP4 and MXFP4, quantizing values into small groups sharing a scale, assuming a fixed floating-point grid. In this paper, we study the following natural extension: assume that, for each group of values, we are free to select the "better" among two or more 4-bit grids marked by one or more bits in the scale value. We formalize the power-of-two-grids (PO2) problem, and provide theoretical results showing that practical small-group formats such as MXFP or NVFP can benefit significantly from PO2 grids, while the advantage vanishes for very large groups. On the practical side, we instantiate several grid families, including 1) PO2(NF4), which pairs the standard NF4 normal grid with a learned grid, 2) MPO2, a grid pair that is fully learned over real weights and activations, 3) PO2(Split87), an explicit-zero asymmetric grid and 4) SFP4, a TensorCore-implementable triple which pairs NVFP4 with two shifted variants. Results for post-training quantization of standard open models and pre-training of Llama-like models show that adaptive grids consistently improve accuracy vs single-grid FP4 under both weight-only and weight+activation. Source code is available at https://github.com/IST-DASLab/GridGames.

0

cs.LG 2026-05-13 Recognition

Joint KL keeps approximation error horizon-free but forces linear estimation growth

Autoregressive Learning in Joint KL: Sharp Oracle Bounds and Lower Bounds

Matching upper and lower bounds show that sequence length affects only the statistical term in autoregressive learning under misspecfit.

abstract click to expand

We study the fundamental and timely problem of learning long sequences in autoregressive modeling and next-token prediction under model misspecification, measured by the joint Kullback--Leibler (KL) divergence. Our goal is to characterize how the sequence horizon $H$ affects both approximation and estimation errors in this joint-distribution, sequence-level regime. By establishing matching upper and lower bounds, we provide, to our knowledge, the first complete characterization of long-horizon error behavior under the natural joint KL objective, with improved rates and optimality justification relative to existing work. On the approximation side, we show that joint KL admits a horizon-free approximation factor, in sharp contrast to Hellinger-based analyses that exhibit an $\Omega(H)$ dependence for computationally efficient methods; this isolates the choice of divergence as the source of approximation amplification. On the estimation side, we prove a fundamental information-theoretic lower bound of order $\Omega(H)$ that holds for both decomposable policy classes and fully shared policies, matching the $\widetilde O(H)$ upper bounds achieved by computationally efficient algorithms. Our analysis clarifies the landscape of recent autoregressive learning results by aligning the log-loss training objective, the sequence-level evaluation metric, and the approximation metric {\color{black}through a sharp joint-KL oracle theory}. We further show that these joint-KL guarantees imply policy learning regret bounds at rates matching prior imitation learning literature.

0

cs.LG 2026-05-13 2 theorems

Causal graph modeling lets RL agents handle random delays and transfer to new tasks

Transferable Delay-Aware Reinforcement Learning via Implicit Causal Graph Modeling

Structured latent nodes and message passing recover action effects despite late feedback, speeding adaptation when rewards change.

abstract click to expand

Random delays weaken the temporal correspondence between actions and subsequent state feedback, making it difficult for agents to identify the true propagation process of action effects. In cross-task scenarios, changes in task objectives and reward formulations further reduce the reusability of previously acquired task knowledge. To address this problem, this paper proposes a transferable delay-aware reinforcement learning method based on implicit causal graph modeling. The proposed method uses a field-node encoder to represent high-dimensional observations as latent states with node-level semantics, and employs a message-passing mechanism to characterize dynamic causal dependencies among nodes, thereby learning transferable structured representations and environment dynamics knowledge. On this basis, imagination-driven behavior learning and planning are incorporated to optimize policies in the latent space, enabling cross-task knowledge transfer and rapid adaptation. Experimental results show that the proposed method outperforms baseline methods on DMC continuous control tasks with random delays. Cross-task transfer experiments further demonstrate that the learned structured representations and dynamics knowledge can be effectively transferred to new tasks and significantly accelerate policy adaptation.

0

cs.LG 2026-05-13 2 theorems

Synthetic data trains AI to spot real tipping points

In-context learning to predict critical transitions in dynamical systems

TipPFN uses in-context learning on randomized bifurcation scenarios for detection in unseen regimes and actual observations.

abstract click to expand

Critical transitions - abrupt, often irreversible changes in system dynamics - arise across human and natural systems, often with catastrophic consequences. Real-world observations of such shifts remain scarce, preventing the development of reliable early warning systems. Conventional statistical and spectral indicators, such as increasing variance, tend to fail under realistic conditions of limited data and correlated noise, whereas existing deep learning classifiers do not extrapolate beyond their training data distribution. In this work, we introduce TipPFN, an in-context learning (ICL) framework that uses a prior-data fitted network to infer a system's proximity to a critical transition. Trained on our novel synthetic data generator, which is based on canonical bifurcation scenarios coupled to diverse, randomized stochastic dynamics, TipPFN flexibly capitalizes on contexts of various sizes, complexity and dimensionalities. We demonstrate robust, state-of-the-art early detection of critical transitions in previously unseen tipping regimes, sim-to-real examples, and real-world observations in both ICL and zero-shot settings.

0

cs.LG 2026-05-13 Recognition

KAN heads cut forgetting by 88-93% on Split-CIFAR tasks

KAN-CL: Per-Knot Importance Regularization for Continual Learning with Kolmogorov-Arnold Networks

Per-knot anchoring in the head plus backbone regularization limits interference while keeping accuracy high.

abstract click to expand

Catastrophic forgetting remains the central obstacle in continual learning (CL): parameters shared across tasks interfere with one another, and existing regularization methods such as EWC and SI apply uniform penalties without awareness of which input region a parameter serves. We propose KAN-CL, a continual learning framework that exploits the compact-support spline parameterization of Kolmogorov-Arnold Networks (KANs) to perform importance-weighted anchoring at per-knot granularity. Deployed as a classification head on a convolutional backbone with standard EWC regularization on the backbone (bbEWC) KAN-CL achieves forgetting reductions of 88% and 93% over a head-only KAN baseline on Split-CIFAR-10/5T and Split-CIFAR-100/10T respectively, while matching or exceeding the accuracy of all baselines on both benchmarks. We further provide a Neural Tangent Kernel (NTK) analysis showing that KAN's spline locality induces a structural rank deficit in the cross-task NTK, yielding a forgetting bound that holds even in the feature-learning regime. These results establish that combining an architecture with natural parameter locality (KAN head) with a complementary backbone regularizer (bbEWC) yields a compositional and principled approach to catastrophic forgetting.

0

cs.LG 2026-05-13 Recognition

Monotone operators approximated via graph convergence

Approximation of Maximally Monotone Operators : A Graph Convergence Perspective

Continuous encoder-decoder architectures achieve local graph convergence to any maximal monotone operator while preserving monotonicity.

abstract click to expand

Operator learning has been highly successful for continuous mappings between infinite-dimensional spaces, such as PDE solution operators. However, many operators of interest-including differential operators-are discontinuous or set-valued, and lie outside classical approximation frameworks. We propose a paradigm shift by formulating approximation via graph convergence (Painlev\'e-Kuratowski convergence), which is well-suited for closed operators. We show that uniform and $L^p$ approximation are fundamentally inadequate in this setting. Focusing on maximally monotone operators, we prove that any such operator can be approximated in the sense of local graph convergence by continuous encoder-decoder architectures, and further construct structure-preserving approximations that retain maximal monotonicity via resolvent-based parameterizations.

0

cs.LG 2026-05-13 Recognition

Simple string encodings plus tabular models suffice for most mixed tables

STRABLE: Benchmarking Tabular Machine Learning with Strings

A benchmark of 108 real tables shows categorical strings dominate, making heavy LLM encoders unnecessary except on free-text heavy cases.

abstract click to expand

Benchmarking tabular learning has revealed the benefit of dedicated architectures, pushing the state of the art. But real-world tables often contain string entries, beyond numbers, and these settings have been understudied due to a lack of a solid benchmarking suite. They lead to new research questions: Are dedicated learners needed, with end-to-end modeling of strings and numbers? Or does it suffice to encode strings as numbers, as with a categorical encoding? And if so, do the resulting tables resemble numerical tabular data, calling for the same learners? To enable these studies, we contribute STRABLE, a benchmarking corpus of 108 tables, all real-world learning problems with strings and numbers across diverse application fields. We run the first large-scale empirical study of tabular learning with strings, evaluating 445 pipelines. These pipelines span end-to-end architectures and modular pipelines, where strings are first encoded, then post-processed, and finally passed to a tabular learner. We find that, because most tables in the wild are categorical-dominant, advanced tabular learners paired with simple string embeddings achieve good predictions at low computational cost. On free-text-dominant tables, large LLM encoders become competitive. Their performance also appears sensitive to post-processing, with differences across LLM families. Finally, we show that STRABLE is a good set of tables to study "string tabular" learning as it leads to generalizable pipeline rankings that are close to the oracle rankings. We thus establish STRABLE as a foundation for research on tabular learning with strings, an important yet understudied area.

0

cs.LG 2026-05-13 Recognition

Sparse neuron ablation halves refusal rates in aligned models

Targeted Neuron Modulation via Contrastive Pair Search

Contrastive search locates a fine-tuning-created gate that can be edited without fluency tradeoffs

abstract click to expand

Language models are instruction-tuned to refuse harmful requests, but the mechanisms underlying this behavior remain poorly understood. Popular steering methods operate on the residual stream and degrade output coherence at high intervention strengths, limiting their practical use. We introduce contrastive neuron attribution (CNA), which identifies the 0.1% of MLP neurons whose activations most distinguish harmful from benign prompts, requiring only forward passes with no gradients or auxiliary training. In instruct models, ablating the discovered circuit reduces refusal rates by over 50% on a standard jailbreak benchmark while preserving fluency and non-degeneracy across all steering strengths. Applying CNA to matched base and instruct models across Llama and Qwen architectures (from 1B to 72B parameters), we find that base models contain similar late-layer discrimination structures but steering these neurons produces only content shifts, not behavioral change. These results demonstrate that neuron-level intervention enables reliable behavioral steering without the quality tradeoffs of residual-stream methods. More broadly, our findings suggest that alignment fine-tuning transforms pre-existing discrimination structure into a sparse, targetable refusal gate.

0

cs.LG 2026-05-13 2 theorems

Root injection of LLM priors boosts RL planning efficiency

PriorZero: Bridging Language Priors and World Models for Decision Making

By limiting language knowledge to the search start and using world model values for updates, PriorZero improves performance on games and igr

abstract click to expand

Leveraging the rich world knowledge of Large Language Models (LLMs) to enhance Reinforcement Learning (RL) agents offers a promising path toward general intelligence. However, a fundamental prior-dynamics mismatch hinders existing approaches: static LLM knowledge cannot directly adapt to the complex transition dynamics of long-horizon tasks. Using LLM priors as fixed policies limits exploration diversity, as the prior is blind to environment-specific dynamics; while end-to-end fine-tuning suffers from optimization instability and credit assignment issues. To bridge this gap, we propose PriorZero, a unified framework that integrates LLM-derived conceptual priors into world-model-based planning through a decoupled rollout-training design. During rollout, a novel root-prior injection mechanism incorporates LLM priors exclusively at the root node of Monte Carlo Tree Search (MCTS), focusing search on semantically promising actions while preserving the world model's deep lookahead capability. During training, PriorZero decouples world-model learning from LLM adaptation: the world model is continuously refined on interaction data to jointly improve its dynamics, policy, and value predictions, its value estimates are then leveraged to provide fine-grained credit assignment signals for stable LLM fine-tuning via alternating optimization. Experiments across diverse benchmarks, including text-based adventure games in Jericho and instruction-following gridworld tasks in BabyAI, demonstrate that PriorZero consistently improves both exploration efficiency and asymptotic performance, establishing a promising framework for LLM-empowered decision-making. Our code is available at https://github.com/opendilab/LightZero.

0

cs.LG 2026-05-13 Recognition

Hypernetwork generates classifier weights for any feature subset on demand

Hypernetworks for Dynamic Feature Selection

The resulting dynamic selection model beats prior methods on tabular data and generalizes better to feature combinations never seen in trai

abstract click to expand

Dynamic feature selection (DFS) is a machine learning framework in which features are acquired sequentially for individual samples under budget constraints. The exponential growth in the number of possible feature acquisition paths forces a DFS model to balance fitting specific scenarios against maintaining general performance, even when the feature space is moderate in size. In this paper, we study the structural limitations of existing DFS approaches to achieve an optimal solution. Then, we propose \textsc{Hyper-DFS}, a hypernetwork-based DFS approach that generates feature subset-specific classifier parameters on demand. We show that the use of hypernetworks compared to mask-embedding methods results in a smaller structural complexity bound. We also use a Set Transformer encoding to create a smooth conditioning space for the hypernetwork, so that functionally similar tasks are also geometrically close. In our benchmarks, \textsc{Hyper-DFS} outperforms all state-of-the-art approaches on synthetic and real-life tabular data. It is also competitive or superior across all image datasets tested, and shows substantially stronger zero-shot generalisation to feature subsets never seen during training than existing DFS approaches.

0

cs.LG 2026-05-13 2 theorems

Hierarchical RL learns causal delays for better control under timing noise

Delay-Empowered Causal Hierarchical Reinforcement Learning

DECHRL discovers state-transition causality and delay distributions from observations alone, then uses a delay-aware empowerment objective.

abstract click to expand

Many real-world tasks involve delayed effects, where the outcomes of actions emerge after varying time lags. Existing delay-aware reinforcement learning methods often rely on state augmentation, prior knowledge of delay distributions, or access to non-delayed data, limiting their generalization. Hierarchical reinforcement learning, by contrast, inherently offers advantages in handling delays due to its hierarchical structure, yet existing methods are restricted to fixed delays. To address these limitations, we propose Delay-Empowered Causal Hierarchical Reinforcement Learning (DECHRL). DECHRL explicitly models both the causal structure of state transitions and their associated stochastic delay distributions. These are then incorporated into a delay-aware empowerment objective that drives proactive exploration toward highly controllable states, thereby improving performance under temporal uncertainty. We evaluate DECHRL in modified 2D-Minecraft and MiniGrid environments featuring stochastic delays. Experimental results show that DECHRL effectively models temporal delays and significantly outperforms baselines in decision-making under temporal uncertainty.

0

cs.LG 2026-05-13 2 theorems

Instruction tokens alone flag object hallucinations in MLLMs

Instruction Lens Score: Your Instruction Contributes a Powerful Object Hallucination Detector for Multimodal Large Language Models

The InsLen score uses embedding patterns to detect made-up objects without auxiliary models or retraining and beats prior detectors.

abstract click to expand

Multimodal large language models (MLLMs) have achieved remarkable progress, yet the object hallucination remains a critical challenge for reliable deployment. In this paper, we present an in-depth analysis of instruction token embeddings and reveal that they implicitly encode visual information while effectively filtering erroneous information introduced by misleading visual embeddings. Building on this insight, we propose the Instruction Lens Score (InsLen), which combines a Calibrated Local Score with a Context Consistency Score that measures context consistency of the object tokens. The proposed approach serves as a plug-and-play object hallucination detector without relying on auxiliary models or additional training. Extensive experiments across multiple benchmarks and diverse MLLM architectures demonstrate that InsLen consistently outperforms existing hallucination detection methods, highlighting its effectiveness and robustness. The code is available at https://github.com/Fraserlairh/Instruction-Lens-Score.

0

cs.LG 2026-05-13

SOAR boosts NVFP4 accuracy for LLMs via scale optimization

SOAR: Scale Optimization for Accurate Reconstruction in NVFP4 Quantization

Closed-form joint optimization and decoupled search cut precision loss at fixed 4-bit memory and hardware cost.

abstract click to expand

NVFP4 has recently emerged as an efficient 4-bit microscaling format for large language models (LLMs), offering superior numerical fidelity with native hardware support. However, existing methods often yield suboptimal performance due to inflexible scale selection and the coupled treatment of quantization and dequantization scales. To address these issues, we propose Scale Optimization for Accurate Reconstruction (SOAR), a novel post-training quantization framework that improves the accuracy of NVFP4 quantization. At its core, SOAR features Closed-form Joint Scale Optimization (CJSO), which jointly optimizes global and block-wise scales via analytical solutions derived from reconstruction error minimization. Furthermore, it incorporates Decoupled Scale Search (DSS). DSS decouples the high-precision quantization scale from its constrained dequantization counterpart, and performs discrete search to mitigate precision loss from scale quantization. Extensive experiments across multiple LLMs show that our method consistently outperforms existing NVFP4 quantization baselines, achieving superior accuracy under the same memory footprint with no additional hardware overhead. The code and models will be available at https://github.com/steven-bao1/SOAR.

0

cs.LG 2026-05-13 2 theorems

Agents complete tasks but fail to follow needed cues

No More, No Less: Task Alignment in Terminal Agents

New benchmark of 89 tasks shows frontier agents cannot distinguish relevant instructions from plausible distractors.

abstract click to expand

Terminal agents are increasingly capable of executing complex, long-horizon tasks autonomously from a single user prompt. To do so, they must interpret instructions encountered in the environment (e.g., README files, code comments, stack traces) and determine their relevance to the task. This creates a fundamental challenge: relevant cues must be followed to complete a task, whereas irrelevant or misleading ones must be ignored. Existing benchmarks do not capture this ability. An agent may appear capable by blindly following all instructions, or appear robust by ignoring them altogether. We introduce TAB (Task Alignment Benchmark), a suite of 89 terminal tasks derived from Terminal-Bench 2.1. Each task is intentionally underspecified, with missing information provided as a necessary cue embedded in a natural environmental artifact, alongside a plausible but irrelevant distractor. Solving these tasks requires selectively using the cue while ignoring the distractor. Applying TAB to ten frontier agents reveals a systematic gap between task capability and task alignment. The strongest Terminal-Bench agent achieves high task completion but low task alignment on TAB. Evaluating six prompt-injection defenses further shows that suppressing distractor execution also suppresses the cues required for task completion. These results demonstrate that task-aligned agents require selective use of environmental instructions rather than blanket acceptance or rejection.

0

cs.LG 2026-05-13 Recognition

Vicarious conditioning gives RL agents intrinsic rewards without policy access

Intrinsic Vicarious Conditioning for Deep Reinforcement Learning

Memory implementation of four psychological steps extends episodes in sidewalk and car-racing tasks by avoiding bad terminal states

abstract click to expand

Advancements in reinforcement learning have produced a variety of complex and useful intrinsic driving forces; crucially, these drivers operate under a direct conditioning paradigm. This form of conditioning limits our agents' capacity by restricting how they learn from the environment as well as from others. Off-policy or learn-by-example methods can learn from demonstrators' representations, but they require access to the demonstrating agent's policies or their reward functions. Our work overcomes this direct sampling limitation by introducing vicarious conditioning as an intrinsic reward mechanism. We draw from psychological and biological literature to provide a foundation for vicarious conditioning and use memory-based methods to implement its four steps: attention, retention, reproduction, and reinforcement. Crucially, our vicarious conditioning paradigms support low-shot learning and do not require the demonstrator agent's policy nor its reward functions. We evaluate our approach in the MiniWorld Sidewalk environment, one of the few public environments that features a non-descriptive terminal condition (no reward provided upon agent death), and extend it to Box2D's CarRacing environment. Our results across both environments demonstrate that vicarious conditioning enables longer episode lengths by discouraging the agent from non-descriptive terminal conditions and guiding the agent toward desirable states. Overall, this work emulates a cognitively-plausible learning paradigm better suited to problems such as single-life learning or continual learning.

0

cs.LG 2026-05-13 Recognition

LoRA accuracy depends on which parameters are trained

Not How Many, But Which: Parameter Placement in Low-Rank Adaptation

Random selection fails under GRPO while gradient-informed selection succeeds by picking entries with stable gradient signs.

abstract click to expand

We study the \textit{parameter placement problem}: given a fixed budget of $k$ trainable entries within the B matrix of a LoRA adapter (A frozen), does the choice of which $k$ matter? Under supervised fine-tuning, random and informed subsets achieve comparable performance. Under GRPO on base models, random placement fails to improve over the base model, while gradient-informed placement recovers standard LoRA accuracy. This regime dependence traces to gradient structure: SFT gradients are low-rank and directionally stable, so any subset accumulates coherent updates; GRPO gradients are high-rank and near-orthogonal across steps, so only elements with consistently signed gradients retain the learning signal. Our scoring procedure identifies these critical parameters in under 10 seconds at less than 0.5% of training cost. Selected parameters concentrate on residual-stream-writing projections (V, O, Down), stable across model families and scales (1.5B - 8B).

0

cs.LG 2026-05-13 Recognition

Multistability required for RL policies to generalize across horizons

On the Importance of Multistability for Horizon Generalization in Reinforcement Learning

A policy optimal after short-horizon training stays optimal at long horizons only if its memory network supports multiple stable states; mon

abstract click to expand

In reinforcement learning (RL), agents acting in partially observable Markov decision processes (POMDPs) must rely on memory, typically encoded in a recurrent neural network (RNN), to integrate information from past observations. Long-horizon POMDPs, in which the relevant observation and the optimal action are separated by many time steps (called the horizon), are particularly challenging: training suffers from poor generalization, severe sample inefficiency, and prohibitive exploration costs. Ideally, an agent trained on short horizons would retain optimal behavior at arbitrarily longer ones, but no formal framework currently characterizes when this is achievable. To fill this gap, we formalized temporal horizon generalization, the property that a policy remains optimal for all horizons, derived a necessary and sufficient condition for it, and experimentally evaluated the ability of nonlinear and parallelizable RNN variants to achieve it. This paper presents the resulting theoretical framework, the empirical evaluation, and the dynamical interpretation linking RNN behavior to temporal horizon generalization. Our analyses reveal that multistability is necessary for temporal horizon generalization and, in simple tasks, sufficient; more complex tasks further require transient dynamics. In contrast, modern parallelizable architectures, namely state space models and gated linear RNNs, are monostable by construction and consequently fail to generalize across temporal horizons. We conclude that multistability and transient dynamics are two essential and complementary dynamical regimes for horizon generalization, and that no current parallelizable RNN exhibits both. Designing parallelizable architectures that combine these regimes thus emerges as a key direction for scalable long-horizon RL.

0

cs.LG 2026-05-13 1 theorem

TabPFN-TS beats Chronos-2 on simple covariate tests

Investigating simple target-covariate relationships for Chronos-2 and TabPFN-TS

Controlled synthetic experiments show stronger capture of target relationships especially at short horizons.

abstract click to expand

Time Series Foundation Models (TSFMs) have recently achieved state-of-the-art performance, often outperforming supervised models in zero-shot settings. Recent TSFM architectures, such as Chronos-2 and TabPFN-TS, aim to integrate covariates. In this paper, we design controlled experiments based on simple target-covariate relationships to assess this integration capability. Our results show that TabPFN-TS captures these relationships more effectively than Chronos-2, especially for short horizons, suggesting that the strong benchmark performance of Chronos-2 does not automatically translate into optimal modeling of simple covariate-target dependencies.

0

cs.LG 2026-05-13 Recognition

Misalignment appears after task convergence in fine-tuning

Overtrained, Not Misaligned

Early stopping removes emergent misalignment while preserving 93 percent of the intended task performance

abstract click to expand

Emergent misalignment (EM), where fine-tuning on a narrow task (like insecure code) causes broad misalignment across unrelated domains, was first demonstrated by Betley et al. (2025). We conduct the most comprehensive EM study to date, reproducing the original GPT-4o finding and expanding to 12 open-source models across 4 families (Llama, Qwen, DeepSeek, GPT-OSS) ranging from 8B to 671B parameters, evaluating over one million model responses with multiple random seeds. We find that EM replicates in GPT-4o but is far from universal: only 2 of 12 open-source models (17%) exhibit consistent EM across seeds, with a significant correlation between model size and EM susceptibility. Through checkpoint-level analysis during fine-tuning, we demonstrate that EM emerges late in training, distinct from and subsequent to near convergence of the primary task, suggesting EM emerges from continued training past task convergence. This yields practical mitigations: early stopping eliminates EM while retaining an average of 93% of task performance, and careful learning rate selection further minimizes risk. Cross-domain validation on medical fine-tuning confirms these patterns generalize: the size-EM correlation strengthens (r = 0.90), and overgeneralization to untruthfulness remains avoidable via early stopping in 67% of cases, though semantically proximate training domains produce less separable misalignment. As LLMs become increasingly integrated into real-world systems, fine-tuning and reinforcement learning remain the primary methods for adapting model behavior. Our findings demonstrate that with proper training practices, EM can be avoided, reframing it from an unforeseen fine-tuning risk to an avoidable training artifact.

0

cs.LG 2026-05-13 2 theorems

One GNN encoder unifies graphs across domains inside LLMs

A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning

Adaptive alignment during instruction tuning produces graph tokens that work with diverse structures and tasks.

abstract click to expand

Leveraging Graph Neural Networks (GNNs) as graph encoders and aligning the resulting representations with Large Language Models (LLMs) through alignment instruction tuning has become a mainstream paradigm for constructing Graph Language Models (GLMs), combining the generalization ability of LLMs with the structural modeling capacity of GNNs. However, existing GLMs that adopt GNNs as graph encoders largely overlook the problem of aligning GNN-encoded representations across domains and tasks with the LLM token space to obtain unified graph tokens, thereby limiting their ability to generalize across diverse graph data. To bridge this gap, we aim to incorporate a multi-domain, multi-task GNN encoder into GLMs and align its representations with LLMs to enable multi-domain, multi-task graph alignment instruction tuning. This alignment problem remains underexplored and poses two key challenges: 1) learning GNN-encoded representations that are simultaneously generalizable across domains and tasks and well aligned with textual semantics is difficult, due to substantial variations in graph structures, feature distributions, and supervision signals, together with the lack of textual-semantic alignment guidance in task-specific GNN training; 2) diverse graph data and task-specific instructions can exhibit different degrees of compatibility with the LLM token space during instruction tuning, leading to varying alignment difficulty and rendering a fixed alignment strategy suboptimal. To tackle these challenges, we propose UniGraphLM, a Unified Graph Language Model that incorporates a multi-domain, multi-task GNN encoder to learn generalizable graph representations aligned with textual semantics, and then adaptively aligns these representations with the LLM.

0

cs.LG 2026-05-13 Recognition

Adaptive weather selection lifts wind power forecast accuracy 2-8%

ECTO: Exogenous-Conditioned Temporal Operator for Ultra-Short-Term Wind Power Forecasting

By picking relevant meteorological inputs with physical knowledge and adjusting via learned regime experts, the method cuts errors across 3-

abstract click to expand

Accurate ultra-short-term wind power forecasting is critical for grid dispatch and reserve management, yet remains challenging due to the non-stationary, condition-dependent nature of wind generation. Meteorological exogenous variables carry substantial predictive information, but the most informative variable combination varies across sites, operating conditions, and prediction horizons. Existing deep learning approaches either treat exogenous inputs as generic auxiliary channels through uniform mixing or soft gating, or rely on fixed preprocessing steps such as PCA, without exploiting the physical structure of meteorological variables. We propose ECTO (Exogenous-Conditioned Temporal Operator), a unified framework that decomposes exogenous variable modeling into two complementary modules. Physically-Grounded Variable Selection (PGVS) performs hierarchical, group-aware sparse selection over exogenous variables using a domain-informed physical prior and sparsemax activations, producing a compact, condition-adaptive exogenous context. Exogenous-Conditioned Regime Refinement (ECRR) routes the forecast through learned regime experts that apply gain--bias calibration and horizon-specific corrections via a mixture-of-experts paradigm. Experiments on three wind farms spanning different climates, capacities (66--200 MW), and exogenous dimensions (11--13 variables) demonstrate that ECTO achieves the lowest MSE across all sites, with relative improvements over the strongest baseline ranging from 2.2% to 5.2%, widening to 8.6% at the longer prediction horizon ($H=32$). Ablation analysis confirms that each exogenous-related component contributes positively (PGVS +1.84%, ECRR +2.86%), and interpretability analysis reveals that PGVS learns physically meaningful, site-specific variable selection patterns, while ECRR converges to well-separated calibration strategies consistent across sites.

0

cs.LG 2026-05-13 2 theorems

Learned representations enable fair conformal classification

Fair Conformal Classification via Learning Representation-Based Groups

The method defines adaptive subgroups via nonlinear features to guarantee conditional coverage and reduce bias in prediction sets.

abstract click to expand

Conformal prediction methods provide statistically rigorous marginal coverage guarantees for machine learning models, but such guarantees fail to account for algorithmic biases, thereby undermining fairness and trust. This paper introduces a fair conformal inference framework for classification tasks. The proposed method constructs prediction sets that guarantee conditional coverage on adaptively identified subgroups, which can be implicitly defined through nonlinear feature combinations. By balancing effectiveness and efficiency in producing compact, informative prediction sets and ensuring adaptive equalized coverage across unfairly treated subgroups, our approach paves a practical pathway toward trustworthy machine learning. Extensive experiments on both synthetic and real-world datasets demonstrate the effectiveness of the framework.

0

cs.LG 2026-05-13 Recognition

Low-rank RKHS cuts drifting model training cost

DriftXpress: Faster Drifting Models via Projected RKHS Fields

Approximating kernels in projected feature spaces keeps FID scores while lowering wall-clock training time for one-step generators.

abstract click to expand

Drifting Models have emerged as a new paradigm for one-step generative modeling, achieving strong image quality without iterative inference. The premise is to replace the iterative denoising process in diffusion models with a single evaluation of a generator. However, this creates a different trade-off: drifting reduces inference cost by moving much of the computation into training. We introduce DriftXpress, an accelerated formulation of drifting models based on projected RKHS fields. DriftXpress approximates the drifting kernel in a low-rank feature space. This preserves the attraction-repulsion structure of the original drifting field while reducing the cost of field evaluation. Across image-generation benchmarks, DriftXpress achieves comparable FID to standard drifting while reducing wall-clock training cost. These results show that the training-inference trade-off of drifting models can be pushed further without giving up their one-step inference advantage.

0

cs.LG 2026-05-13 2 theorems

Shared low-rank features cut regret in safe multi-task bandits

Multi-Task Representation Learning for Conservative Linear Bandits

New algorithm recovers constrained representations and proves regret plus sample bounds across multiple linear bandit tasks.

abstract click to expand

This paper presents the Constrained Multi-Task Representation Learning (CMTRL) framework for linear bandits. We consider T linear bandit tasks in a d dimensional space, which share a common low-dimensional representation of dimension r, where r is much smaller than the minimum of d and T. Furthermore, tasks are constrained so that only actions meeting specific safety or performance requirements are allowed, referred to as conservative (safe) bandits. We introduce a novel algorithm, Safe-Alternating projected Gradient Descent and minimization (Safe-AltGDmin), to recover a low-rank feature matrix while satisfying the given constraints. Building on this algorithm, we propose a multi-task representation learning framework for conservative linear bandits and establish theoretical guarantees for its regret and sample complexity bounds. We presented experiments and compared the performance of our algorithm with benchmark algorithms.

0

cs.LG 2026-05-13 2 theorems

Averaging minibatch OT plans yields consistent couplings for flow matching

Expected Batch Optimal Transport Plans and Consequences for Flow Matching

The resulting population coupling produces a regular velocity field that defines unique flows and reduces integration cost in generative sam

abstract click to expand

Solving optimal transport (OT) on random minibatches is a common surrogate for exact OT in large-scale learning. In flow matching (FM), this surrogate is used to obtain OT-like couplings that can straighten probability paths and reduce numerical integration cost. Yet, the population-level coupling induced by repeated minibatch OT remains only partially understood. We formalize this coupling as the expected batch OT plan $\overline{\pi}_{k}$, obtained by averaging empirical OT plans over independent minibatches of size $k$. We then establish its large-batch consistency and, in the semidiscrete case relevant to generative modeling, derive rates for both the transport-cost bias and the convergence of $\overline{\pi}_{k}$ to the OT plan. For FM, this yields a population coupling whose induced velocity field is regular enough to define a unique flow from the source to the discrete target. We finally quantify how OT batch size interacts with numerical integration in a tractable two-atom model and in synthetic and image experiments.

0

cs.LG 2026-05-13 Recognition

One attention layer needs linear heads-degree product for parity

Lower bounds for one-layer transformers that compute parity

Lower bound shows the product must grow with input length to sign-represent parity in single-layer models

abstract click to expand

This note shows that no self-attention layer post-processed by a rational function can sign-represent the parity function unless the product of the number of heads and the degree of the post-processing function grows linearly with the input length. Combining this lower bound with rational approximation of ReLU networks yields a margin-dependent extension for self-attention layers post-processed by ReLU networks.

0

cs.LG 2026-05-13 2 theorems

Low-res data boosts scarce high-res model training

On What We Can Learn from Low-Resolution Data

KL divergence bounds show low-resolution samples retain value, confirmed by gains in vision transformer and CNN experiments.

abstract click to expand

Artificial intelligence systems typically rely on large, centrally collected datasets, a premise that does not hold in many real-world domains such as healthcare and public institutions. In these settings, data sharing is often constrained by storage, privacy, or resource limitations. For example, small wearable devices may lack the bandwidth or energy capacity needed to store and transmit high-resolution data, leading to aggregation during data collection and thus a loss of information. As a result, datasets collected from different sources may consist of a mixture of high- and low-resolution samples. Despite the prevalence of this setting, it remains unclear how informative low-resolution data is when models are ultimately evaluated on high-resolution inputs. We provide a theoretical analysis based on the Kullback-Leibler divergence that characterises how the influence of a datapoint changes with resolution, and derive bounds that relate the relative contribution of high- and low-resolution observations to the information lost under downsampling. To support this analysis, we empirically demonstrate, using both a vision transformer and a convolutional neural network, that adding low-resolution data to the training set consistently improves performance when high-resolution data is scarce.

0

cs.LG 2026-05-13 Recognition

Adaptive weights add feature selection to FGW distances

Fused Gromov-Wasserstein Distance with Feature Selection

Selective suppression of irrelevant features during structure alignment yields more robust and interpretable distances.

abstract click to expand

Fused Gromov-Wasserstein (FGW) distances provide a principled framework for comparing objects by jointly aligning structure and node features. However, existing FGW formulations treat all features uniformly, which limits interpretability and robustness in high-dimensional settings where many features may be irrelevant or noisy. We introduce FGW distances with feature selection, which incorporate adaptive feature suppression weights into the FGW objective to selectively downweight or suppress differentiating features during alignment. We propose two approaches: (1) regularized FGW with Lasso and Ridge penalties, and (2) FGW with simplex-constrained weights, including groupwise extensions. We analyze the resulting models and establish their key theoretical properties, including bounds relative to classical FGW and Gromov-Wasserstein distances, and metric behavior. An efficient alternating minimization algorithm is developed. Experiments illustrate how feature suppression enhances interpretability and reveals task-relevant structure, with a special application to computational redistricting.

0

cs.LG 2026-05-13 Recognition

Contrastive clusters let diffusion models unlearn concepts with less leakage

Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning

SAEParate adds a concept-aware objective and GeLU to sparse autoencoders so one target can be suppressed without harming others.

abstract click to expand

Unlearning specific concepts in text-to-image diffusion models has become increasingly important for preventing undesirable content generation. Among prior approaches, sparse autoencoder (SAE)-based methods have attracted attention due to their ability to suppress target concepts through lightweight manipulation of latent features, without modifying model parameters. However, SAEs trained with sparse reconstruction objectives do not explicitly enforce concept-wise separation, resulting in shared latent features across concepts. To address this, we propose SAEParate, which organizes latent representations into concept-specific clusters via a concept-aware contrastive objective, enabling more precise concept suppression while reducing unintended interference during unlearning. In addition, we enhance the encoder with a GeLU-based nonlinear transformation to increase its expressive capacity under this separation objective, enabling a more discriminative and disentangled latent space. Experiments on UnlearnCanvas demonstrate state-of-the-art performance, with particularly strong gains in joint style-object unlearning, a challenging setting where existing methods suffer from severe interference between target and non-target concepts.

0

cs.LG 2026-05-13 Recognition

Pairwise expert comparisons boost Bayesian optimization efficiency

Elicitation-Augmented Bayesian Optimization

Cost-aware acquisition decides between direct observations and comparisons, improving sample use when expert input is reliable and cheap.

abstract click to expand

Human-in-the-loop Bayesian optimization (HITL BO) methods utilize human expertise to improve the sample-efficiency of BO. Most HITL BO methods assume that a domain expert can quantify their knowledge, for instance by pinpointing query locations or specifying their prior beliefs about the location of the maximum as a probability distribution. However, since human expertise is often tacit and cannot be explicitly quantified, we consider a setting where domain knowledge of an expert is elicited via pairwise comparisons of designs. We interpret the expert's pairwise judgements as noisy evidence about the values of the observable objective function and develop a principled method for combining the information obtained via direct observations and pairwise queries. Specifically, we derive a cost-aware value-of-information acquisition function that balances direct observations against pairwise queries. The proposed method approaches the convex hull of the trajectories of the individual information sources: when pairwise queries are cheap it substantially improves sample-efficiency over observation-only BO, and when pairwise queries are costly or noisy, it recovers the performance of standard BO by relying on direct observations alone.

0

cs.LG 2026-05-13 Recognition

Async RL loses old logits and breaks PPO correction

Missing Old Logits in Asynchronous Agentic RL: Semantic Mismatch and Repair Methods for Off-Policy Correction

Exact recovery strategies and revised PPO-EWMA restore the split between discrepancy and staleness terms, raising training speed and agent性能

abstract click to expand

Asynchronous reinforcement learning improves rollout throughput for large language model agents by decoupling sample generation from policy optimization, but it also introduces a critical failure mode for PPO-style off-policy correction. In heterogeneous training systems, the total importance ratio should ideally be decomposed into two semantically distinct factors: a \emph{training--inference discrepancy term} that aligns inference-side and training-side distributions at the same behavior-policy version, and a \emph{policy-staleness term} that constrains the update from the historical policy to the current policy. We show that practical asynchronous pipelines with delayed updates and partial rollouts often lose the required historical training-side logits, or old logits. This missing-old-logit problem entangles discrepancy repair with staleness correction, breaks the intended semantics of decoupled correction, and makes clipping and masking thresholds interact undesirably. To address this issue, we study both exact and approximate correction routes. We propose three exact old-logit acquisition strategies: snapshot-based version tracking, a dedicated old-logit model, and synchronization via partial rollout interruption, and compare their system trade-offs. From the perspective of approximate correction, we focus on preserving the benefits of decoupled correction through a more appropriate approximate policy when exact old logits cannot be recovered at low cost, without incurring extra system overhead. Following this analysis, we adopt a revised PPO-EWMA method, which achieves significant gains in both training speed and optimization performance. Code at https://github.com/millioniron/ROLL.

0

cs.LG 2026-05-13 Recognition

HölderPO boosts math accuracy to 54.9% over GRPO

H\"older Policy Optimisation

Annealing the Hölder parameter p balances gradient concentration and stability for better policy updates in language models.

abstract click to expand

Group Relative Policy Optimisation (GRPO) enhances large language models by estimating advantages across a group of sampled trajectories. However, mapping these trajectory-level advantages to policy updates requires aggregating token-level probabilities within each sequence. Relying on a fixed aggregation mechanism for this step fundamentally limits the algorithm's adaptability. Empirically, we observe a critical trade-off: certain fixed aggregations frequently suffer from training collapse, while others fail to yield satisfactory performance. To resolve this, we propose \textbf{H\"{o}lderPO}, a generalised policy optimisation framework unifying token-level probability aggregation via the H\"{o}lder mean. By explicitly modulating the parameter $p$, our framework provides continuous control over the trade-off between gradient concentration and variance bounds. Theoretically, we prove that a larger $p$ concentrates the gradient to amplify sparse learning signals, whereas a smaller $p$ strictly bounds gradient variance. Because no static configuration can universally resolve this concentration-stability trade-off, we instantiate the framework with a dynamic annealing algorithm that progressively schedules $p$ across the training lifecycle. Extensive evaluations demonstrate superior stability and convergence over existing baselines. Specifically, our approach achieves a state-of-the-art average accuracy of $54.9\%$ across multiple mathematical benchmarks, yielding a substantial $7.2\%$ relative gain over standard GRPO and secures an exceptional $93.8\%$ success rate on ALFWorld.

0

cs.LG 2026-05-13 Recognition

Direct modeling produces plug-in surrogates that better predict primary effects

Learning plug-in surrogate endpoints for randomized experiments

Learned functions of post-treatment variables outperform prior methods when substituted for long-term outcomes in randomized trials.

abstract click to expand

Surrogate endpoints are used in place of long-term outcomes in randomized experiments when observing the real outcome for a large enough cohort is prohibitively expensive or impractical. A short-term surrogate is good if the result of an experiment using the surrogate is predictive of the result of a hypothetical study using the real outcome. Much attention has been paid to formalizing this property in causal terms, but most criteria are unidentifiable and cannot be turned into practical algorithms for learning surrogate endpoints from data. To address this, we study plug-in composite surrogates, functions of post-treatment variables that may be substituted directly for the primary outcome in a randomized experiment. We propose two methods for learning plug-in surrogates that maximize effect predictiveness, and characterize the possibility of finding endpoints that yield unbiased effect estimates in representative scenarios. Finally, in both synthetic experiments with known effects and in data from a real-world experiment, we find that our method, based on directly modeling the surrogate effect, returns plug-in endpoints more predictive of the primary effect than established methods.

0

cs.LG 2026-05-13 Recognition

Complex neurons beat simple ones under fixed parameters

Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons

Recurrent-network tradeoffs shift the optimum toward higher per-neuron complexity as total budget grows.

abstract click to expand

Cortical neurons are complex, multi-timescale processors wired into recurrent circuits, shaped by long evolutionary pressure under stringent biological constraints. Mainstream machine learning, by contrast, predominantly builds models from extremely simple units, a default inherited from early neural-network theory. We treat this as a normative architectural question. How should one split a fixed parameter budget $P$ between the number of units $N$, per-unit effective complexity $k_e$, and per-unit connectivity $k_c$? What controls the optimal allocation? This calls for a model in which per-unit complexity can be tuned independently of width and connectivity. Accordingly, we introduce the ELM Network, whose recurrent layer is built from Expressive Leaky Memory (ELM) neurons, chosen to mirror functional components of cortical neurons. The architecture allows for individually adjusting $N$, $k_e$, and $k_c$ and trains stably across orders of magnitude in scale. We evaluate the model on two qualitatively different sequence benchmarks: the neuromorphic SHD-Adding task and Enwik8 character-level language modeling. Performance improves monotonically along each of the three axes individually. Under a fixed budget, a clear non-trivial optimum emerges in their tradeoff, and larger budgets favor both more and more complex neurons. A closed-form information-theoretic model captures these tradeoffs and attributes the diminishing returns at two ends to: per-neuron signal-to-noise saturation and across-neuron redundancy. A hyperparameter sweep spanning three orders of magnitude in trainable parameters traces a near-Pareto-frontier scaling law consistent with the framework. This suggests that the simple-unit default in ML is not obviously optimal once this tradeoff surface is probed, and offers a normative lens on cortex's reliance on complex spatio-temporal integrators.

0

cs.LG 2026-05-13 2 theorems

Model maintains diagnostic accuracy with missing medical data

Resilient Vision-Tabular Multimodal Learning under Modality Missingness

Masked attention and dropout let a vision-tabular transformer handle any level of modality absence without imputation.

abstract click to expand

Multimodal deep learning has shown strong potential in medical applications by integrating heterogeneous data sources such as medical images and structured clinical variables. However, most existing approaches implicitly assume complete modality availability, an assumption that rarely holds in real-world clinical settings where entire modalities and individual features are frequently missing. In this work, we propose a multimodal transformer framework for joint vision-tabular learning explicitly designed to operate under pervasive modality missingness, without relying on imputation or heuristic model switching. The architecture integrates three components: a vision, a tabular, and a multimodal fusion encoder. Unimodal representations are weighted through learnable modality tokens and fused via intermediate fusion with masked self-attention, which excludes missing tokens and modalities from information aggregation and gradient propagation. To further enhance resilience, we introduce a modality-dropout regularization strategy that stochastically removes available modalities during training, encouraging the model to exploit complementary information under partial data availability. We evaluate our approach on the MIMIC-CXR dataset paired with structured clinical data from MIMIC-IV for multilabel classification of 14 diagnostic findings with incomplete annotations. Two parallel systematic stress-test protocols progressively increase training and inference missingness in each modality separately, spanning fully multimodal to fully unimodal scenarios. Across all missingness regimes, the proposed method consistently outperforms representative baselines, showing smoother performance degradation and improved robustness. Ablation studies further demonstrate that attention-level masking and intermediate fusion with joint fine-tuning are key to resilient multimodal inference.

0

cs.LG 2026-05-13 Recognition

Laplacian neural operators learn PDE maps with polynomial cost

Approximation Theory of Laplacian-Based Neural Operators for Reaction-Diffusion System

Error bounds depend polynomially on depth, width and spectral rank for the Gierer-Meinhardt system

abstract click to expand

Neural operators provide a framework for learning solution operators of partial differential equations (PDEs), enabling efficient surrogate modeling for complex systems. While universal approximation results are now well understood, approximation analysis specific to nonlinear reaction-diffusion systems remains limited. In this paper, we study neural operators applied to the solution mapping from initial conditions to time-dependent solutions of a generalized Gierer-Meinhardt reaction-diffusion system, a prototypical model of nonlinear pattern formation. Our main results establish explicit approximation error bounds in terms of network depth, width, and spectral rank by exploiting the Laplacian spectral representation of the Green's function underlying the PDE. We show that the required parameter complexity grows at most polynomially with respect to the target accuracy, demonstrating that Laplacian eigenfunction-based neural operator architectures alleviate the curse of parametric complexity encountered in generic operator learning. Numerical experiments on the Gierer-Meinhardt system support the theoretical findings.

0

cs.LG 2026-05-13 2 theorems

LLMs adapt as backbones for activity recognition from sensors

Efficient and Adaptive Human Activity Recognition via LLM Backbones

A convolutional projection plus frozen model and LoRA adapters deliver fast training and cross-dataset robustness with far fewer parameters.

abstract click to expand

Human Activity Recognition (HAR) is a core task in pervasive computing systems, where models must operate under strict computational constraints while remaining robust to heterogeneous and evolving deployment conditions. Recent advances based on Transformer architectures have significantly improved recognition performance, but typically rely on task-specific models trained from scratch, resulting in high training cost, large data requirements, and limited adaptability to domain shifts. In this paper, we propose a paradigm shift that reuses large pretrained language models (LLMs) as generic temporal backbones for sensor-based HAR, instead of designing domain-specific Transformers. To bridge the modality gap between inertial time series and language models, we introduce a structured convolutional projection that maps multivariate accelerometer and gyroscope signals into the latent space of the LLM. The pretrained backbone is kept frozen and adapted using parameter-efficient Low-Rank Adaptation (LoRA), drastically reducing the number of trainable parameters and the overall training cost. Through extensive experiments on standard HAR benchmarks, we show that this approach enables rapid convergence, strong data efficiency, and robust cross-dataset transfer, particularly in low-data and few-shot settings. At the same time, our results highlight the complementary roles of convolutional frontends and LLMs, where local invariances are handled at the signal level while long-range temporal dependencies are captured by the pretrained backbone. Overall, this work demonstrates that LLMs can serve as a practical, frugal, and scalable foundation for adaptive HAR systems, opening new directions for reusing foundation models beyond their original language domain.

0

cs.LG 2026-05-13 Recognition

Reachable subspace dynamics uniquely determined from single trajectory

Limits of Learning Linear Dynamics from Experiments

Linear systems yield unique restricted dynamics on the reachable part even when the full system cannot be identified.

abstract click to expand

Learning governing dynamics from data is a common goal across the sciences, yet it is only well-posed when the underlying mechanisms are identifiable. In practice, many data-driven methods implicitly assume identifiability; when this assumption fails, estimated models can yield spurious predictions and invalid mechanistic conclusions. Classical identifiability guarantees for controlled linear time-invariant (LTI) systems provide sufficient conditions -- controllability and persistent excitation -- but leave open whether identifiability holds when these conditions fail, and which parts of the system remain identifiable without full identifiability. We show that the experimental setup, i.e., the realized initial state and control input, dictates a fundamental limit on the information recoverable from the observed trajectory. We develop a geometric characterization of this limit and derive a closed-form description of all systems consistent with the experimental setup. Crucially, we prove that even when the full system is not identifiable, the restricted dynamics on the subspace reachable by the experiment remain uniquely determined.

0

cs.LG 2026-05-13 2 theorems

Group Lasso in GNN embeddings finds important subgraphs without labels

Estimating Subgraph Importance with Structural Prior Domain Knowledge

A linear regression on pretrained embeddings, grouped by structural priors, recovers subgraph contributions and needs no target values or a

abstract click to expand

We propose a subgraph importance estimation method for pretrained Graph Neural Networks (GNNs) on graph-level tasks, formulated as a linear Group Lasso regression problem in the embedding space. Our method effectively leverages prior domain knowledge of graph substructures, while remaining independent of the specific form of the output layer or readout function used in the GNN architecture, and it does not require access to ground-truth target labels. Experiments on real-world graph datasets demonstrate that our method consistently outperforms existing baselines in subgraph importance estimation. Furthermore, we extend our method to identify important nodes within the graph.

0

cs.LG 2026-05-13 Recognition

MA-BC splits conflicting data to reach Pareto fronts faster

Split the Differences, Pool the Rest: Provably Efficient Multi-Objective Imitation

The algorithm pools consistent state-action pairs across experts and matches the minimax lower bound on sample complexity.

abstract click to expand

This work investigates multi-objective imitation learning: the problem of recovering policies that lie on the Pareto front given demonstrations from multiple Pareto-optimal experts in a Multi-Objective Markov Decision Process (MOMDP). Standard imitation approaches are ill-equipped for this regime, as naively aggregating conflicting expert trajectories can result in dominated policies. To address this, we introduce Multi-Output Augmented Behavioral Cloning (MA-BC), an algorithm that systematically partitions divergent expert data while pooling state-action pairs where no behavior conflict is observed. Theoretically, we prove that MA-BC converges to Pareto-optimal policies at a faster statistical rate than any learner that considers each expert dataset independently. Furthermore, we establish a novel lower bound for multi-objective imitation learning, demonstrating that MA-BC is minimax optimal. Finally, we empirically validate our algorithm across diverse discrete environments and, guided by our theoretical insights, extend and evaluate MA-BC on a continuous Linear Quadratic Regulator (LQR) control task.

0

cs.LG 2026-05-13 2 theorems

Anchor quantization stabilizes Schrödinger bridge couplings

QDSB: Quantized Diffusion Schr\"odinger Bridges

Solving entropic transport only on quantized anchors and lifting the plan cell-wise preserves quality while reducing computation time.

abstract click to expand

Learning generative models in settings where the source and target distributions are only specified through unpaired samples is gaining in importance. Here, one frequently-used model are Schr\"odinger bridges (SB), which represent the most likely evolution between both endpoint distributions. To accelerate training, simulation-free SBs avoid the path simulation of the original SB models. However, learning simulation-free SBs requires paired data; a coupling of the source and target samples is obtained as the solution of the entropic optimal transport (OT) problem. As obtaining the optimal global coupling is infeasible in many practical cases, the entropic OT problem is iteratively solved on minibatches instead. Still, the repeated cost remains substantial and the locality can distort the global transport geometry. We propose quantized diffusion Schr\"odinger bridges (QDSB), which compute the endpoint coupling on anchor-quantized endpoint distributions and lift the resulting plan back to original data points through cell-wise sampling. We show that the regularized optimal coupling is stable w.r.t. anchor quantization, with an error controlled by the quality of the anchor approximation. In real-world experiments, QDSB matches the sample quality of existing baselines, requiring substantially less time. Code and data are available at github.com/mathefuchs/qdsb.

0

cs.LG 2026-05-13 2 theorems

Reach-avoid certificates let RL minimize cost under safety probabilities

Stochastic Minimum-Cost Reach-Avoid Reinforcement Learning

States certified to meet the probabilistic spec are folded into a contracting Bellman surrogate, yielding policies with lower expected cost.

abstract click to expand

We study stochastic minimum-cost reach-avoid reinforcement learning, where an agent must satisfy a reach-avoid specification with probability at least $p$ while minimizing expected cumulative costs in stochastic environments. Existing safe and constrained reinforcement learning methods typically fail to jointly enforce probabilistic reach-avoid constraints and optimize cost in the learning setting in stochastic environments. To address this challenge, we introduce reach-avoid probability certificates (RAPCs), which identify states from which stochastic reach-avoid constraints are satisfiable. Building on RAPCs, we develop a contraction-based Bellman formulation that serves as a principled surrogate for integrating reach-avoid considerations into reinforcement learning, enabling cost optimization under probabilistic constraints. We establish almost sure convergence of the proposed algorithms to locally optimal policies with respect to the resulting objective. Experiments in the MuJoCo simulator demonstrate improved cost performance and consistently higher reach-avoid satisfaction rates.

0

cs.LG 2026-05-13 2 theorems

RL method cuts order bias in LLMs while lifting accuracy

Towards Order Fairness: Mitigating LLMs Order Sensitivity through Dual Group Advantage Optimization

Dual advantage signals reward stable correct outputs across input arrangements on reasoning and retrieval tasks.

abstract click to expand

Large Language Models (LLMs) suffer from order bias, where their performance is affected by the arrangement order of input elements. This unfairness limits the model's applications in scenarios such as in-context learning and Retrieval-Augmented Generation (RAG). Recent studies attempt to obtain optimal or suboptimal arrangements based on statistical results or using dataset-based search, but these methods increase inference overhead while leaving the model's inherent order bias unresolved. Other studies mitigate order sensitivity through supervised fine-tuning using augmented training sets with multiple order variants, but often at the cost of accuracy, trapping the model in consistent yet incorrect hallucinations. In this paper, we propose \textbf{D}ual \textbf{G}roup \textbf{A}dvantage \textbf{O}ptimization (\textbf{DGAO}), which aims to improve model accuracy and order stability simultaneously. DGAO calculates and balances intra-group relative accuracy advantage and inter-group relative stability advantage, rewarding the policy model for generating order-stable and correct outputs while penalizing order-sensitive or incorrect responses. This marks the first time reinforcement learning has been used to mitigate LLMs' order sensitivity. We also propose two new metrics, Consistency Rate and Overconfidence Rate, to reveal the pseudo-stability of previous methods and guide more comprehensive evaluation. Extensive experiments demonstrate that DGAO achieves superior order fairness while improving performance on RAG, mathematical reasoning, and classification tasks. Our code is available at: https://github.com/Hyalinesky/DGAO.

0

cs.LG 2026-05-13 2 theorems

Function embedding cuts local stress to 0.111 on ERA5 data

NOFE -- Neural Operator Function Embedding

By mapping continuous functions instead of points, NOFE keeps fine structure and stays consistent across different samplings.

abstract click to expand

Most dimensionality reduction methods treat data as discrete point clouds, ignoring the continuous domain structure inherent to many real-world processes. To bridge this gap, we introduce Neural Operator Function Embedding (NOFE), a domain-aware framework for continuous dimensionality reduction. NOFE learns function-to-function mappings via a Graph Kernel Operator, enabling mesh-free evaluation at arbitrary query locations independent of input discretization. We establish NOFE as approximation of sheaf-to-sheaf mappings, generalizing Sheaf Neural Networks to continuous domains. We evaluate NOFE across different datasets, comparing it against PCA, t-SNE, and UMAP. Our results demonstrate that NOFE significantly outperforms baselines in local structure preservation, achieving a local Stress of 0.111 compared to 0.398 for PCA, 0.773 for t-SNE, and 0.791 for UMAP for the ERA5 climate reanalysis dataset. NOFE also exhibits robust sampling independence, reducing the Patch Stitching Error by up to $20.0\times$ relative to UMAP (59.0 vs. 267.6 under regional normalization) and ensuring consistency across disjoint domain patches. While maintaining competitive global structure preservation (Stress-1: 0.379 vs. PCA's 0.268), NOFE resolves fine-grained structures and produces smooth, consistent embeddings that generalize across varying sample densities, addressing key limitations of discrete reduction methods.

0

cs.LG 2026-05-13 2 theorems

STAGE aligns multimodal graphs across clients while cutting communication

STAGE: Tackling Semantic Drift in Multimodal Federated Graph Learning

Translation into a shared space plus regulated propagation reduces inconsistency and payload on eight graphs and five tasks.

abstract click to expand

Federated graph learning (FGL) enables collaborative training on graph data across multiple clients. As graph data increasingly contain multimodal node attributes such as text and images, multimodal federated graph learning (MM-FGL) has become an important yet substantially harder setting. The key challenge is that clients from different modality domains may not share a common semantic space: even for the same concept, their local encoders can produce inconsistent representations before collaboration begins. This makes direct parameter coordination unreliable and further causes two downstream problems: forcing heterogeneous client representations into a naively shared semantic space may create false semantic agreement, and graph message passing may amplify residual inconsistency across neighborhoods. To address this issue, we propose \textbf{STAGE}, a protocol-first framework for MM-FGL. Instead of relying on direct parameter averaging, STAGE builds a shared semantic space that first translates heterogeneous multimodal features into comparable representations and then regulates how these representations propagate over local graph structures. In this way, STAGE not only improves cross-client semantic calibration, but also reduces the risk of inconsistency amplification during graph learning. Extensive experiments on 8 multimodal-attributed graphs across 5 graph-centric and modality-centric tasks show that STAGE consistently achieves state-of-the-art performance while reducing per-round communication payload.

0

cs.LG 2026-05-13 1 theorem

Predictive coding aligns updates more closely with errors than backpropagation

Understanding Sample Efficiency in Predictive Coding

Exact formulas for deep linear networks show the efficiency gain grows with depth and pre-training; experiments indicate the pattern holds

abstract click to expand

Predictive Coding (PC) is an influential account of cortical learning. Much of recent work has focused on comparing PC to Backpropagation (BP) to find whether PC offers any advantages. Small scale experiments show that PC enables learning that is more sample efficient and effective in many contexts, though a thorough theoretical understanding of the phenomena remains elusive. To address this, we quantify the efficiency of learning in BP and PC through a metric called ``target alignment'', which measures how closely the change in the output of the network is aligned to the output prediction error. We then derive and empirically validate analytical expressions for target alignment in Deep Linear Networks. We show that learning in PC is more efficient than BP, which is especially pronounced in deep, narrow and pre-trained networks. We also derive exact conditions for guaranteed optimal target alignment in PC and validate our findings through experiments. We study full training trajectories of linear and non-linear models, and find the predicted benefits of PC persist in practice even when some assumptions are violated. Overall, this work provides a mechanistic understanding of the higher learning efficiency observed for PC over BP in previous works, and can guide how PC should be parametrised to learn most effectively.

0

cs.LG 2026-05-13 Recognition

Gated gradients escape bad policy corners at logarithmic speed

Delightful Gradients Accelerate Corner Escape

Multiplying updates by advantage and surprisal removes self-trapping and yields global O(1/t) convergence in bandits and tabular MDPs.

abstract click to expand

Softmax policy gradient converges at $O(1/t)$, but its transient behavior near sub-optimal corners of the simplex can be exponentially slow. The bottleneck is self-trapping: negative-advantage actions reinforce the corner policy and can initially push the optimal action backward. We study \emph{Delightful Policy Gradient} (DG), which gates each policy-gradient term by the product of advantage and action surprisal. For $K$-armed bandits, we prove that the zero-temperature limit of DG removes this corner-trapping mechanism on a quantitative sector near any sub-optimal corner, yielding a first-exit escape bound logarithmic in the initial probability ratio. At every fixed temperature, the same local mechanism persists because harmful actions are polynomially suppressed as they become rare. A key structural insight is that every action better than the corner action is an \emph{ally}: its contribution to escape is non-negative. Combining corner instability with a monotonic value improvement identity, we prove that DG converges globally to the optimal policy in both bandits and tabular MDPs at an asymptotic $O(1/t)$ rate. We also show, via an exact counterexample, that this tabular mechanism can fail under shared function approximation. In MNIST contextual bandits with a shared-parameter neural network, DG nevertheless recovers from bad initializations faster than standard policy gradient, suggesting that the counterexample marks a boundary of the theory rather than a practical prohibition.

0

cs.LG 2026-05-13 Recognition

SFT lifts procedural skills uniformly from 0.8B to 4B

Procedural-skill SFT across capacity tiers: A W-Shaped pre-SFT Trajectory and Regime-Asymmetric Mechanism on 0.8B-4B Qwen3.5 Models

Pre-SFT W-shape dominates final variation, with SFT helping most where bases struggle

abstract click to expand

We measure procedural-skill SFT contribution across three Qwen3.5 dense scales (0.8B, 2B, 4B) on a 200-task / 40-skill holdout, with Claude Haiku 4.5 as a frontier reference. The corpus is 353 rows of (task + procedural-skill block, Opus chain-of-thought, judge-pass) demonstrations. \textbf{Main finding.} Under matched-path LLM-only scoring, the SFT-attributable procedural-$\Delta$ lift is roughly uniform across sizes: $+0.070$ / $+0.040$ / $+0.075$ at 0.8B / 2B / 4B. Variation in post-SFT $\Delta$ ($-0.005$, $+0.100$, $+0.065$) is dominated by a W-shaped pre-SFT base trajectory ($-0.075$, $+0.060$, $-0.010$, Haiku-4-5 at $+0.030$): the 5-step procedure hurts 0.8B and 4B, helps 2B, and helps frontier Haiku modestly. SFT works hardest in absolute terms where the base struggles with the procedure -- a regime-asymmetric pattern with a falsifiable prediction at 8B/14B. \textbf{Methodology.} (i) A bench format-compliance artifact: 83.5\% of the holdout uses a deterministic \texttt{ANSWER}-line extractor that under-counts free-form conclusions; an LLM-only re-judge reveals it was systematically biased against \CU. (ii) A negative-iteration sequence at 0.8B: five recipe variants cluster post-SFT \CU{} pass-rate within a 2\,pp band, constraining the absolute-pass-rate ceiling to base capacity rather than recipe. \textbf{Cross-family validation.} GPT-5.4 via OpenRouter on all 7 configurations (2800 paired episodes) agrees on the direction of every per-student finding: Cohen's $\kappa \geq 0.754$, agreement $\geq 93.25\%$. Earlier ``format-only at 0.8B'' and ``shrinking SFT at 4B'' framings were path-mismatch artifacts; this paper supersedes both (Appendix~\ref{sec:appendix-path}). Single-seed; threats in \S\ref{sec:threats}.

0

cs.LG 2026-05-13 Recognition

Semivalues plus secret validation set enforce truthful data sharing

Incentivizing Truthfulness and Collaborative Fairness in Bayesian Learning

Sources maximize rewards only when they submit datasets that reflect their genuine knowledge in Bayesian collaborative learning.

abstract click to expand

Collaborative machine learning involves training high-quality models using datasets from a number of sources. To incentivize sources to share data, existing data valuation methods fairly reward each source based on its data submitted as is. However, as these methods do not verify nor incentivize data truthfulness, the sources can manipulate their data (e.g., by submitting duplicated or noisy data) to artificially increase their valuations and rewards or prevent others from benefiting. This paper presents the first mechanism that provably ensures (F) collaborative fairness and incentivizes (T) truthfulness at equilibrium for Bayesian models. Our mechanism combines semivalues (e.g., Shapley value), which ensure fairness, and a truthful data valuation function (DVF) based on a validation set that is unknown to the sources. As semivalues are influenced by others' data, we introduce an additional condition to prove that a source can maximize its expected data values in coalitions and semivalues by submitting a dataset that captures its true knowledge. Additionally, we discuss the implications and suitable relaxations of (F) and (T) when the mediator has a limited budget for rewards or lacks a validation set. Our theoretical findings are validated on synthetic and real-world datasets.

0

cs.LG 2026-05-13 Recognition

Sobolev penalty yields global MMD flow convergence

Sobolev Regularized MMD Gradient Flow

Gradient penalty on the witness function removes the need for isoperimetric assumptions and works for both sampling and generation.

abstract click to expand

We propose Sobolev-regularized Maximum Mean Discrepancy (SrMMD) gradient flow, a regularized variant of maximum mean discrepancy (MMD) gradient flow based on a gradient penalty on the witness function. The proposed regularization mitigates the non-convexity of the MMD objective and yields provable \emph{global} convergence guarantees in MMD in both continuous and discrete time. A more surprising appeal is that our convergence analysis does not rely on isoperimetric assumptions on the target distribution. Instead, it is based on a regularity condition on the difference between kernel mean embeddings. A key highlight of the proposed flow is that it is applicable in both sampling (from an unnormalized target distribution) -- using Stein kernels -- and generative modeling settings, unlike previous works, where a gradient flow is suitable for only generative modeling or sampling but not both. The effectiveness of the proposed flow is empirically verified on a broad range of tasks in both generative modelling and sampling.

0

cs.LG 2026-05-13 Recognition

Estimator adapts TD(lambda) for superior multi-agent RL

Adaptive TD-Lambda for Cooperative Multi-agent Reinforcement Learning

By comparing replay buffers from past and current policies, the approach sets lambda per state-action pair based on likelihood and beats fix

abstract click to expand

TD($\lambda$) in value-based MARL algorithms or the Temporal Difference critic learning in Actor-Critic-based (AC-based) algorithms synergistically integrate elements from Monte-Carlo simulation and Q function bootstrapping via dynamic programming, which effectively addresses the inherent bias-variance trade-off in value estimation. Based on that, some recent works link the adaptive $\lambda$ value to the policy distribution in the single-agent reinforcement learning area. However, because of the large joint action space from multiple number of agents, and the limited transition data in Multi-agent Reinforcement Learning, the policy distribution is infeasible to be calculated statistically. To solve the policy distribution calculation problem in MARL settings, we employ a parametric likelihood-free density ratio estimator with two replay buffers instead of calculating statistically. The two replay buffers of different sizes store the historical trajectories that represent the data distribution of the past and current policies correspondingly. Based on the estimator, we assign Adaptive TD($\lambda$), \textbf{ATD($\lambda$)}, values to state-action pairs based on their likelihood under the stationary distribution of the current policy. We apply the proposed method on two competitive baseline methods, QMIX for value-based algorithms, and MAPPO for AC-based algorithms, over SMAC benchmarks and Gfootball academy scenarios, and demonstrate consistently competitive or superior performance compared to other baseline approaches with static $\lambda$ values.

0

cs.LG 2026-05-13 2 theorems

LOFT improves orthogonal fine-tuning via task-aware support selection

LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection

Separating subspace choice from the rotation inside it lets gradient signals pick which directions to adapt, raising accuracy under fixed PE

abstract click to expand

Orthogonal parameter-efficient fine-tuning (PEFT) adapts pretrained weights through structure-preserving multiplicative transformations, but existing methods often conflate two distinct design choices: the subspace in which adaptation occurs and the transformation applied within that subspace. This paper introduces LOFT, a low-rank orthogonal fine-tuning framework that explicitly separates these two components. By viewing orthogonal adaptation as a multiplicative subspace rotation, LOFT provides a unified formulation that recovers representative orthogonal PEFT methods, including coordinate-, butterfly-, Householder-, and principal-subspace-based variants. More importantly, this perspective exposes support selection as a central design axis rather than a byproduct of a particular parameterization. We develop a first-order analysis showing that useful adaptation supports should be informed by the downstream training signal, motivating practical task-aware support selection strategies. Across language understanding, visual transfer, mathematical reasoning, and multilingual out-of-distribution adaptation, LOFT recovers principal-subspace orthogonal adaptation while gradient-informed supports improve the efficiency-performance trade-off under matched parameter, memory, and compute budgets. These results suggest that principled support selection is an important direction for improving orthogonal PEFT.

0

cs.LG 2026-05-13 2 theorems

KL divergence with teacher constraint yields SSL batch centering

Information theoretic underpinning of self-supervised learning by clustering

The derivation shows why distillation and centering are required rather than optional in clustering-based self-supervised learning.

abstract click to expand

Self-supervised learning (SSL) is recognized as an essential tool for building foundation models for Artificial Intelligence applications. The advances in SSL have been made thanks to vigorous arguments about the principles of SSL and through extensive empirical research. The aim of this paper is to contribute to the development of the underpinning theory of SSL, focusing on the deep clustering approach. By analogy to supervised learning, we formulate SSL as K-L divergence optimization. The mode collapse is prevented by imposing an optimisation constraint on the teacher distribution. This leads to normalization using inverse cluster priors. We show that using Jensen inequality this normalization simplifies to the popular batch centering procedure. Distillation and centering are common {heuristics-based} practices in SSL, {but our work underpins them theoretically.} The theoretical model developed not only supports specific existing successful SSL methods, but also suggests directions for future investigations.

0

cs.LG 2026-05-13 2 theorems

Semantic consensus replaces parameter sharing in federated LLM tuning

Beyond Parameter Aggregation: Semantic Consensus for Federated Fine-Tuning of LLMs

Outputs on public prompts yield pseudo-labels whose communication cost stays independent of model size

abstract click to expand

Federated fine-tuning of large language models is commonly formulated as a parameter aggregation problem. However, even parameter-efficient methods require transmitting large collections of trainable weights, assume aligned architectures, and rely on white-box access to model parameters. As model sizes continue to grow and deployments become increasingly heterogeneous, these assumptions become progressively misaligned with practical constraints. We consider an alternative formulation in which collaboration is mediated through model behavior rather than parameters. Clients fine-tune local models on private data and exchange generated outputs on a shared, public prompt set. The server maps these outputs into a semantic representation space, forms a per-prompt semantic consensus, and returns pseudo-labels for further local fine-tuning. This formulation fundamentally changes the communication scaling of federated LLM fine-tuning. The amount of information exchanged depends only on the public prompt budget and the size of the communicated behaviors, independent of model size. As a consequence, the protocol naturally accommodates heterogeneous architectures and applies directly to open-ended text generation. We present a theoretical analysis and empirical results demonstrating that this approach can match strong federated fine-tuning baselines while substantially reducing communication by orders of magnitude (e.g., analytically by a factor of $1006$ for Llama3.1-405B), as well as reductions in runtime and energy consumption. These results suggest that, for generative foundation models, behavior-level consensus provides a more appropriate abstraction for federated adaptation than parameter aggregation.

0