A classifier trained only on transformer fine-tuning data detects an invariant memorization signature that transfers to Mamba, RWKV-4, and RecurrentGemma with AUCs of 0.963, 0.972, and 0.936.
super hub Mixed citations
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
Mixed citation behavior. Most common role is background (67%).
167
Pith papers citing it
Background
67% of classified citations
abstract
While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper we introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form, allowing us to solve the standard RLHF problem with only a simple classification loss. The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant, and computationally lightweight, eliminating the need for sampling from the LM during fine-tuning or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds PPO-based RLHF in ability to control sentiment of generations, and matches or improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.
hub tools
citation-role summary
background 25
method 9
baseline 1
dataset 1
citation-polarity summary
claims ledger
- abstract While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then
authors
co-cited works
representative citing papers
ORPO performs preference alignment during supervised fine-tuning via a monolithic odds ratio penalty, allowing 7B models to outperform larger state-of-the-art models on alignment benchmarks.
Controlled student-teacher experiments across four benchmarks show interactive gains are driven more by the student's ability to use feedback than by teacher quality, with self-feedback adding little beyond unguided retries.
PrincipalBench exposes a sharp split in frontier LLMs between selective and over-refusing behavior on multi-party loyalty, with prompt scaffolding and KL distillation reducing harm rates but only along an existing leak/over-refusal trade-off.
Flow models reach 99.2% Sudoku accuracy in 7 passes and 96.1% on out-of-distribution Sudoku-Extreme by selecting dynamically stable candidates and training with self-conditioning plus DPO to avoid failed outputs.
LOGICA adds context to pretrained biological LMs via logit-space contrastive alignment with gated adapters, improving AUC on held-out drug-resistance mutation ranking from ~0.55 to ~0.65 while preserving token likelihoods.
TimeROME-DLM enables training-free knowledge editing in masked diffusion language models via temporal causal tracing and low-rank residual edit memory applied at inference time.
Alignment defenses adapted from DPO and GRPO mitigate property inference attacks on LLMs while preserving utility.
LLM judges exhibit high stability under neutral re-evaluation but substantial reversibility under targeted post-decision challenges, quantified via a new Evaluation Robustness Score (ERS).
PInVerify is a new offline embodied benchmark for active instance verification that supplies multi-view captures and 6-sector navigation topology, with MLLM baselines reaching 85.6% after fine-tuning but showing no reliable benefit from tested next-best-view strategies.
A hybrid first-order then zeroth-order optimization approach improves robustness of safety-aligned LLMs while preserving utility, with layer-wise sensitivity estimation for efficiency.
Self-evolving rubric with anti-gaming fitness reveals that objective capability scaling fails to transfer to subjective LLM behaviors, with advice-restraint as the universal lowest dimension that can regress.
Formalizes Reasoning Portability (RP) and proposes RDB-CL to modulate per-sample KL regularization in RLVR for MLLM continual learning, achieving +12.0% Last accuracy over vanilla RLVR baseline by preserving reusable reasoning on high-RP samples.
Fast-Slow Training uses context optimization as fast weights alongside parameter updates as slow weights to achieve up to 3x better sample efficiency, higher performance, and less catastrophic forgetting than standard RL in continual LLM learning.
Introduces TBPO, which derives a Bregman-divergence density-ratio matching objective for token-level preference optimization that generalizes DPO while preserving the induced optimal policy.
Introduces Block-R1 benchmark, Block-R1-41K dataset, and a conflict score to handle domain-specific optimal block sizes in RL post-training of diffusion LLMs.
Optimistic bilevel optimization with manifold lower-level minimizers is differentiable if the optimistic selection is unique, yielding a pseudoinverse hyper-gradient and a convergent HG-MS algorithm whose rate depends on intrinsic manifold dimension.
POISE trains a lightweight probe on the actor's internal states to predict expected rewards for RLVR, matching DAPO performance on math benchmarks with lower compute by avoiding extra rollouts or critic models.
Instruction tuning makes late-layer computation depend more on the model's own post-trained upstream state than on base-model upstream state, producing a consistent +1.68 logit interaction effect across five model families.
Topology-enhanced alignment via persistent homology on trajectories outperforms standard SFT and DPO baselines on preference metrics for LLMs.
Full development of 7B and 32B Olmo 3 models used 12.3 GWh datacenter energy and emitted 4,251 tCO2eq, with development overheads accounting for 82% of compute and reasoning models costing 17x more to post-train than instruction-tuned ones.
PEO optimizes original prompt embeddings continuously over adaptive rounds to jailbreak aligned LLMs, preserving the exact visible prompt text and outperforming discrete suffix, appended embedding, and search-based white-box attacks on harmful-behavior benchmarks.
Small VLMs show higher sycophancy (22.3% for 450M model) than larger ones (6.0% for 7B) when scoring image-text alignment on 173k fantasy portraits, quantified via a new Bluffing Coefficient metric.
A new 7x4 taxonomy organizes agentic AI security threats by architectural layer and persistence timescale, revealing under-explored upper layers and missing defenses after surveying 116 papers.
citing papers explorer
-
Select-then-differentiate: Solving Bilevel Optimization with Manifold Lower-level Solution Sets
Optimistic bilevel optimization with manifold lower-level minimizers is differentiable if the optimistic selection is unique, yielding a pseudoinverse hyper-gradient and a convergent HG-MS algorithm whose rate depends on intrinsic manifold dimension.