DataComp-VLM benchmark shows instruction-heavy data mixing outperforms filtering for VLM training, with DCVLM-Baseline achieving 63.6% on 33 tasks for 8B models (+5.4pp over FineVision).
super hub Canonical reference
Tulu 3: Pushing Frontiers in Open Language Model Post-Training
Canonical reference. 77% of citing Pith papers cite this work as background.
177
Pith papers citing it
Background
77% of classified citations
abstract
Language model post-training is applied to refine behaviors and unlock new skills across a wide range of recent language models, but open recipes for applying these techniques lag behind proprietary ones. The underlying training data and recipes for post-training are simultaneously the most important pieces of the puzzle and the portion with the least transparency. To bridge this gap, we introduce Tulu 3, a family of fully-open state-of-the-art post-trained models, alongside its data, code, and training recipes, serving as a comprehensive guide for modern post-training techniques. Tulu 3, which builds on Llama 3.1 base models, achieves results surpassing the instruct versions of Llama 3.1, Qwen 2.5, Mistral, and even closed models such as GPT-4o-mini and Claude 3.5-Haiku. The training algorithms for our models include supervised finetuning (SFT), Direct Preference Optimization (DPO), and a novel method we call Reinforcement Learning with Verifiable Rewards (RLVR). With Tulu 3, we introduce a multi-task evaluation scheme for post-training recipes with development and unseen evaluations, standard benchmark implementations, and substantial decontamination of existing open datasets on said benchmarks. We conclude with analysis and discussion of training methods that did not reliably improve performance. In addition to the Tulu 3 model weights and demo, we release the complete recipe -- including datasets for diverse core skills, a robust toolkit for data curation and evaluation, the training code and infrastructure, and, most importantly, a detailed report for reproducing and further adapting the Tulu 3 approach to more domains.
hub tools
citation-role summary
background 20
method 3
dataset 2
baseline 1
citation-polarity summary
claims ledger
- abstract Language model post-training is applied to refine behaviors and unlock new skills across a wide range of recent language models, but open recipes for applying these techniques lag behind proprietary ones. The underlying training data and recipes for post-training are simultaneously the most important pieces of the puzzle and the portion with the least transparency. To bridge this gap, we introduce Tulu 3, a family of fully-open state-of-the-art post-trained models, alongside its data, code, and training recipes, serving as a comprehensive guide for modern post-training techniques. Tulu 3, whic
authors
co-cited works
representative citing papers
Derives an exact telescoping decomposition of the naive RLVR reward-design estimator into null, elicitation, and reward-design terms on a tabular-GRPO simulator, measures the components across prior strengths, and validates via pre-registered factorial experiments plus re-audits of prior papers.
Masked diffusion LMs can use continuous x-prediction flow with token-wise asynchronous updates and an RL policy network to reach 97% performance on HumanEval using only 25% of the usual decoding budget.
PACE is a clipped per-coordinate controller added to AdamW that improves the limiting error of the returned iterate average in both quadratic analysis and LM experiments.
Verifiable search procedures cannot be learned as forward chain-of-thought by language models; they instead learn memorization, verification, or require precomputed catalogs.
Distance-Adaptive Representation (DAR) keeps full KV dimensionality inside a local window and reduces it to 1/4 outside, matching full-dimensional baselines on pretraining (70M-410M) and 1B-scale fine-tuning while uniform reduction performs worse.
Behavior Forecasters trained on LRM trajectories outperform larger models in predicting repeatability and input sensitivity at low cost.
Trajectories from a Bittensor ShoppingBench subnet arena, filtered to retain only agentic tool-calling behavior, enable SFT+GRPO post-training of Qwen3-4B to 42.7% ASR on leak-guarded held-out tests, nearly matching synthetic-data baselines with a fraction of a day's data.
CapRL++ applies reinforcement learning with verifiable rewards to dense image and video captioning by scoring captions via the accuracy of a vision-free LLM answering MCQs from the caption alone.
PRISM is a contrastive, policy-aware training framework for process reward models that reduces false positives by 22% on PRMBench and boosts downstream accuracy up to 33% in Best-of-N selection by learning reliable relative comparisons instead of pointwise labels.
HARP is a train-based data selector for LLM finetuning that uses hierarchical active region pruning and empirical Bayes posteriors to achieve up to 8.9 point gains with roughly 7 times fewer training examples.
LazyAttention kernelizes deferred positional encoding to enable zero-copy, position-agnostic KV cache reuse, delivering 1.37× lower TTFT and 1.40× higher throughput than Block-Attention under skewed document distributions while preserving output quality.
RogueMerge is a unified attack method that jointly optimizes task vectors to succeed after merging, using stochastic min-max simulation for unknown merging settings and a Taylor-approximated DRO for prompt generalization on generative LLMs.
MENTIS applies layerwise covariance torsion (T1), spectral torsion (T2), and ERA localization to paired IT/PA 7-8B models, finding selective larger shifts for normative concepts, negative correlation with entropy, and mid-to-late layer peaks.
Observational study of 20,574 sessions identifies seven misalignment forms where 90.5% cause effort/trust costs and 91.5% require explicit user correction, varying by interface and over time.
CAI Dataset is presented as the largest described corpus of LLM-driven hacker trajectories, with the claim that operator data concentration in frontier-model providers creates a major security risk best addressed by on-premise specialized LLMs.
Introduces a state-aligned latent actor-critic framework that lets diffusion models act as their own timestep-conditioned value functions for trajectory-level RL post-training and inference steering.
Infilling extraction on diffusion language models extracts up to three times more verbatim sequences than prefix methods and achieves higher recall on redacted emails than autoregressive models.
RECIPE improves visual procedural planners by rewarding plans according to their grounding quality in ASR transcripts via GRPO, yielding +7–8 in-domain and up to +16 zero-shot macro-accuracy gains over base models and outperforming supervised fine-tuning on seven benchmarks.
PARAMΔ upcycles dense models to MoE for per-language experts and grafts post-training deltas to enable data-efficient language expansion while preserving original capabilities.
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.
DRATS derives a minimax objective from a feasibility formulation of MTRL to adaptively sample tasks with the largest return gaps, leading to better worst-task performance on MetaWorld benchmarks.
CurveBench is a new benchmark for recovering rooted containment trees from images of nested Jordan curves, where the strongest model reaches only 19.1% accuracy on hard cases and fine-tuning lifts an open model to 33.3% on easy cases.
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.
citing papers explorer
-
A Pre-Registered Causal Partition of Self-Consistency Elicitation and Reward Design in RLVR
Derives an exact telescoping decomposition of the naive RLVR reward-design estimator into null, elicitation, and reward-design terms on a tabular-GRPO simulator, measures the components across prior strengths, and validates via pre-registered factorial experiments plus re-audits of prior papers.
-
Forecasting Future Behavior as a Learning Task
Behavior Forecasters trained on LRM trajectories outperform larger models in predicting repeatability and input sensitivity at low cost.
-
Multi-modal Reasoning with LLMs for Visual Semantic Arithmetic
SAri-RFT applies GRPO-based reinforcement fine-tuning to LVLMs on novel two-term and three-term visual semantic arithmetic tasks, reaching SOTA on the new IRPD dataset and Visual7W-Telling.
-
SKILL.nb: Selective Formalization and Gated Execution for Durable Agent Workflows
SKILL.nb uses selective formalization and gate-conditioned execution in auditable notebooks to improve durability of agent workflows, achieving 53.7% success on WebArena-Verified with 91.7% retention across re-executions.
-
The Reliability Gap in Benchmark Auditing: Distribution Shift and Scale as Failure Modes of Contamination Detection
Empirical evaluation across 25 LLMs shows contamination detection methods achieve correct outcomes in only 201 of 335 cases, exposing failure modes from distribution shift and benchmark scale.
-
TRON: Targeted Rule-Verifiable Online Environments for Visual Reasoning RL
TRON supplies 520 rule-verifiable online visual reasoning environments across five ability buckets that generate unlimited training instances for RL post-training, yielding consistent gains on ten external multimodal benchmarks for three vision-language models.
-
In LLM Reasoning, there is Irrationality on top of Value Misalignment
LLMs display widespread rational value risk in reasoning that value alignment reduces but does not remove, with risk sensitive to inference strategy and showing diminishing returns from longer reasoning.
-
EVE-Agent: Evidence-Verifiable Self-Evolving Agents
EVE-Agent adds an evidence verifier to the proposer-solver loop that rewards spans by marginal accuracy gain, producing self-generated but inspectable training examples for search agents.
-
Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs
Introduces MOOD benchmark for OOD LLM alignment failures and shows guard models plus Mahalanobis and perplexity OOD detectors improve recall from 39% to 45% with positive scaling.
-
SAPO: Step-Aligned Policy Optimization for Reasoning-Based Generative Recommendation
SAPO computes per-reasoning-step group-relative advantages in RL to improve credit assignment for structured generation of semantic identifiers in recommendation systems.
-
PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-Play
PopuLoRA shows that co-evolving populations of LoRA adapters through cross-evaluated self-play can outperform compute-matched single-agent baselines on multiple code and math reasoning benchmarks.
-
Nudging Beyond the Comfort Zone: Efficient Strategy-Guided Exploration for RLVR
NudgeRL conditions RLVR rollouts on strategy-level contexts to drive diverse trajectories and applies an inter/intra-context reward decomposition plus distillation objective, outperforming GRPO and oracle baselines on math benchmarks.
-
BEAM: Binary Expert Activation Masking for Dynamic Routing in MoE
BEAM uses binary expert activation masks trained end-to-end to achieve dynamic sparsity in MoE models, cutting FLOPs by 85% with over 98% performance retention.
-
Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion
MORA breaks the safety-helpfulness ceiling in LLMs by pre-sampling single-reward prompts and rewriting them to incorporate multi-dimensional intents, delivering 5-12.4% gains in sequential alignment and 4.6% overall improvement in simultaneous alignment.
-
Seir\^enes: Adversarial Self-Play with Evolving Distractions for LLM Reasoning
Seirênes trains LLMs via adversarial self-play to generate and overcome evolving distractions, producing gains of 7-10 points on math reasoning benchmarks and exposing blind spots in larger models.
-
Shallow Prefill, Deep Decoding: Efficient Long-Context Inference via Layer-Asymmetric KV Visibility
SPEED uses layer-asymmetric KV visibility to process non-anchor prompt tokens only in lower layers during prefill, achieving near-baseline quality on Llama-3.1-8B with 33% better TTFT and 25% lower active KV memory at 128K context.
-
ZAYA1-8B Technical Report
ZAYA1-8B is a reasoning MoE model with 700M active parameters that matches larger models on math and coding benchmarks and reaches 91.9% on AIME'25 via Markovian RSA test-time compute.
-
Characterizing Model-Native Skills
Recovering an orthogonal basis from model activations yields a model-native skill characterization that improves reasoning Pass@1 by up to 41% via targeted data selection and supports inference steering, outperforming human-characterized alternatives.
-
The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
LRMs exhibit complete accuracy collapse beyond certain puzzle complexities, with reasoning effort rising then declining, outperforming standard LLMs only on medium-complexity tasks.
-
ComplexConstraints and Beyond: Expert Rubrics for RLVR
Expert-curated rubrics in the new ComplexConstraints dataset improve LLM instruction following by 12-15% when used as RL training signals, with gains transferring to out-of-distribution agentic benchmarks.
-
Capability Self-Assessment: Teaching LLMs to Know Their Limits
Reinforcement learning teaches LLMs to assess their own capabilities more effectively than supervised fine-tuning, preserves original skills, generalizes out of distribution, and aids local-cloud routing and data selection.
-
Mechanistically Interpreting the Role of Sample Difficulty in RLVR for LLMs
Sample difficulty in RLVR shows non-monotonic effects on LLM reasoning, with easy/medium problems strengthening computation and reasoning features while hard problems often yield weak or harmful signals.
-
Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments
NoisyAgent trains LLM agents with controlled user and tool noise to improve robustness in stochastic environments while also boosting clean-benchmark performance.
-
Mid-Training with Self-Generated Data Improves Reinforcement Learning in Language Models
Mid-training LLMs on self-generated diverse reasoning paths improves subsequent RL performance on mathematical benchmarks and OOD tasks.
-
Rethinking Math Reasoning Evaluation: A Robust LLM-as-a-Judge Framework Beyond Symbolic Rigidity
An LLM-as-a-judge evaluation framework for math reasoning outperforms symbolic methods by accurately assessing diverse answer representations and formats.
-
Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations
CRAFT uses contrastive representation learning and RL on hidden states to align reasoning models for improved safety against jailbreaks, reporting 79% and 87.7% gains over base models.
-
GENIUS: An Agentic AI Framework for Autonomous Design and Execution of Simulation Protocols
GENIUS is an agentic AI framework that automates generation, validation, and repair of Quantum ESPRESSO DFT input files, succeeding on ~80% of 295 benchmarks with 76% autonomous repairs and lower cost than LLM-only baselines.
-
Beyond Next-Token Prediction: An RLVR Proof of Concept for Tool-Use Agents on Atlassian Workflows
RLVR training on five synthetic Atlassian API environments raises average tool-use reward for Qwen models from 0.35-0.92 to 0.95-1.00 on four non-degenerate scenarios.
-
Entropy-KL Divergence-based Token Masking: A Novel Approach for Selective Fine-tuning of Large Language Models
EKSFT masks high-entropy or high-KL tokens in low-data SFT to preserve pre-trained distribution and improve downstream RL performance on math reasoning tasks.
-
Efficient Post-training of LLMs for Code Generation With Offline Reinforcement Learning
Offline RL post-training boosts code generation performance in LLMs, with larger gains for small models and hard problems, using pre-collected datasets.
-
Distilling Game Code World Model Generation into Lightweight Large Language Models
SFT followed by RLVR on Qwen2.5-3B-Instruct raises syntactic and execution correctness when generating Game Code World Models across 30 games.
-
Text-Driven 3D Indoor Scene Synthesis in Non-Manhattan Environments
SPG-Layout combines statistical object priors with hierarchical large-object-first placement to produce physically plausible text-driven 3D scenes in non-Manhattan rooms and outperforms baselines on a new 500-scene benchmark.
-
The Hitchhiker's Guide to Agentic AI: From Foundations to Systems
A comprehensive reference book organizing existing techniques for agentic AI systems across LLM substrate, reasoning, agent design patterns, inter-agent coordination, and production deployment.
- SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions
- OASES: Outcome-Aligned Search-Evaluation Co-Training for Agentic Search