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.
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.
132
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
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.
The TAB benchmark reveals that frontier terminal agents achieve high task completion but low selective alignment with relevant environmental cues over distractors, and prompt-injection defenses block both.
Anchor-guided variance-aware reward modeling uses two response-level anchors to resolve non-identifiability in Gaussian models of pluralistic preferences, yielding provable identification, a joint training objective, and improved RLHF performance.
K12-KGraph is a textbook-derived knowledge graph that powers a new benchmark revealing LLMs' poor curriculum cognition and a small training corpus that outperforms general instruction data on educational tasks.
vOPD stabilizes on-policy distillation gradients by subtracting a closed-form per-token negative reverse KL baseline as a detached control variate, preserving unbiasedness while lowering variance and matching expensive full-vocabulary methods.
POPO uses bounded importance sampling on positive rollouts and a siamese policy network to achieve implicit negative gradients and stable optimization, matching or exceeding GRPO on math benchmarks such as 36.67% on AIME 2025.
Reference-sampled weighted SFT with prompt-normalized Boltzmann weights induces the same policy as fixed-reference KL-regularized RLVR, with BOLT as the estimator and a finite one-shot error decomposition separating coverage, variance, and other terms.
Parallel-SFT mixes parallel programs across languages during SFT to produce more transferable RL initializations, yielding better zero-shot generalization to unseen programming languages.
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.
A multi-response discriminative reward model scores N candidates in one pass via concatenation and cross-entropy, achieving SOTA on multimodal benchmarks and improving RL policies over single-response baselines.
ALTO accelerates LoRA tuning up to 13.8x by monitoring loss trajectories for early stopping, using fused grouped GEMM with rank-local adapter parallelism, and combining intra- and inter-task scheduling for heterogeneous workloads without quality loss.
SCRL adds selective positive pseudo-labeling and entropy-gated negative pseudo-labeling to test-time RL, reducing noise from weak consensus and improving LLM reasoning on benchmarks.
Positive-negative prompt pairing with weighted GRPO improves RLVR sample efficiency, raising AIME 2025 Pass@8 from 16.8 to 22.2 on Qwen2.5-Math-7B while matching large-scale training.
HetRL delivers up to 9.17x higher throughput for LLM RL training on heterogeneous GPUs by using hybrid and ILP-based schedulers to solve a joint optimization problem over computation and data dependencies.
ProfBench is a new multi-domain benchmark with human-expert rubrics for judging LLM responses on professional tasks, showing top models reach only 65.9% performance while providing cheap LLM judges that reduce evaluation cost by orders of magnitude.
Defines Decision Potential Surface (DPS) whose zero isohypse equals an LLM decision boundary and supplies a K-sample approximation algorithm with derived upper bounds on absolute, expected, and concentration errors.
Proposes a task taxonomy for functional diversity in LLM outputs, validates it via user study, introduces targeted sampling to boost diversity only where needed, and presents evidence that the diversity-quality tradeoff may be an artifact of task-agnostic measurement.
TokenBuncher constrains response entropy via entropy-as-reward RL and a Token Noiser to stop harmful RL fine-tuning while keeping benign performance intact.
citing papers explorer
-
LazyAttention: Efficient Retrieval-Augmented Generation with Deferred Positional Encoding
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.
-
RECIPE: Procedural Planning via Grounding in Instructional Video
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.
-
A Data-Efficient Path to Multilingual LLMs: Language Expansion via Post-training PARAM$\Delta$ Integration into Upcycled MoE
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.
-
Reasoning Portability: Guiding Continual Learning for MLLMs in the RLVR Era
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.
-
Distributionally Robust Multi-Task Reinforcement Learning via Adaptive Task Sampling
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: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves
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.
-
Learning, Fast and Slow: Towards LLMs That Adapt Continually
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.
-
No More, No Less: Task Alignment in Terminal Agents
The TAB benchmark reveals that frontier terminal agents achieve high task completion but low selective alignment with relevant environmental cues over distractors, and prompt-injection defenses block both.
-
Variance-aware Reward Modeling with Anchor Guidance
Anchor-guided variance-aware reward modeling uses two response-level anchors to resolve non-identifiability in Gaussian models of pluralistic preferences, yielding provable identification, a joint training objective, and improved RLHF performance.
-
K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs
K12-KGraph is a textbook-derived knowledge graph that powers a new benchmark revealing LLMs' poor curriculum cognition and a small training corpus that outperforms general instruction data on educational tasks.
-
KL for a KL: On-Policy Distillation with Control Variate Baseline
vOPD stabilizes on-policy distillation gradients by subtracting a closed-form per-token negative reverse KL baseline as a detached control variate, preserving unbiasedness while lowering variance and matching expensive full-vocabulary methods.
-
Beyond Negative Rollouts: Positive-Only Policy Optimization with Implicit Negative Gradients
POPO uses bounded importance sampling on positive rollouts and a siamese policy network to achieve implicit negative gradients and stable optimization, matching or exceeding GRPO on math benchmarks such as 36.67% on AIME 2025.
-
Reference-Sampled Boltzmann Projection for KL-Regularized RLVR: Target-Matched Weighted SFT, Finite One-Shot Gaps, and Policy Mirror Descent
Reference-sampled weighted SFT with prompt-normalized Boltzmann weights induces the same policy as fixed-reference KL-regularized RLVR, with BOLT as the estimator and a finite one-shot error decomposition separating coverage, variance, and other terms.
-
Parallel-SFT: Improving Zero-Shot Cross-Programming-Language Transfer for Code RL
Parallel-SFT mixes parallel programs across languages during SFT to produce more transferable RL initializations, yielding better zero-shot generalization to unseen programming languages.
-
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.
-
You Only Judge Once: Multi-response Reward Modeling in a Single Forward Pass
A multi-response discriminative reward model scores N candidates in one pass via concatenation and cross-entropy, achieving SOTA on multimodal benchmarks and improving RL policies over single-response baselines.
-
ALTO: Adaptive LoRA Tuning and Orchestration for Heterogeneous LoRA Training Workloads
ALTO accelerates LoRA tuning up to 13.8x by monitoring loss trajectories for early stopping, using fused grouped GEMM with rank-local adapter parallelism, and combining intra- and inter-task scheduling for heterogeneous workloads without quality loss.
-
What If Consensus Lies? Selective-Complementary Reinforcement Learning at Test Time
SCRL adds selective positive pseudo-labeling and entropy-gated negative pseudo-labeling to test-time RL, reducing noise from weak consensus and improving LLM reasoning on benchmarks.
-
Beyond Variance: Prompt-Efficient RLVR via Rare-Event Amplification and Bidirectional Pairing
Positive-negative prompt pairing with weighted GRPO improves RLVR sample efficiency, raising AIME 2025 Pass@8 from 16.8 to 22.2 on Qwen2.5-Math-7B while matching large-scale training.
-
HetRL: Efficient Reinforcement Learning for LLMs in Heterogeneous Environments
HetRL delivers up to 9.17x higher throughput for LLM RL training on heterogeneous GPUs by using hybrid and ILP-based schedulers to solve a joint optimization problem over computation and data dependencies.
-
ProfBench: Multi-Domain Rubrics requiring Professional Knowledge to Answer and Judge
ProfBench is a new multi-domain benchmark with human-expert rubrics for judging LLM responses on professional tasks, showing top models reach only 65.9% performance while providing cheap LLM judges that reduce evaluation cost by orders of magnitude.
-
Decision Potential Surface: A Theoretical and Practical Approximation of Large Language Model Decision Boundary
Defines Decision Potential Surface (DPS) whose zero isohypse equals an LLM decision boundary and supplies a K-sample approximation algorithm with derived upper bounds on absolute, expected, and concentration errors.
-
Task-Dependent Evaluation of LLM Output Homogenization: A Taxonomy-Guided Framework
Proposes a task taxonomy for functional diversity in LLM outputs, validates it via user study, introduces targeted sampling to boost diversity only where needed, and presents evidence that the diversity-quality tradeoff may be an artifact of task-agnostic measurement.
-
Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning
TokenBuncher constrains response entropy via entropy-as-reward RL and a Token Noiser to stop harmful RL fine-tuning while keeping benign performance intact.
-
Reinforcement Learning for Reasoning in Large Language Models with One Training Example
One training example via RLVR boosts LLM math reasoning from 17.6% to 35.7% average across six benchmarks.
-
Enhancing LLM Metacognition via Cognitive Pairwise Training
CPT is introduced as a pairwise reasoning-trace comparison stage that improves the reasoning-metacognition trade-off over standard SFT+RL pipelines across model scales.
-
Decomposed On-Policy Distillation for Vision-Language Reasoning: Steering Gradients for Visual Grounding
Decomposes VLM distillation loss into orthogonal language and visual components and introduces Visual Gradient Steering to prioritize visual grounding over standard monolithic optimization.
-
Combinatorial Synthesis: Scaling Code RLVR via Atomic Decomposition and Recombination
ADR generates novel verifiable code tasks via atomic decomposition and recombination, outperforming heuristic baselines in originality, difficulty, and downstream RLVR gains across coding domains.
-
VeriGate: Verifier-Gated Step-Level Supervision for GRPO
VeriGate adds verifier-gated step-level supervision to GRPO via cumulated PRM rewards and group-normalized token advantages, raising accuracy 20% and 12% on 1.5B and 7B models on MATH and six benchmarks.
-
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.
-
TimeSRL: Generalizable Time-Series Behavioral Modeling via Semantic RL-Tuned LLMs -- A Case Study in Mental Health
TimeSRL uses semantic abstractions from time-series data optimized via reinforcement learning to achieve better cross-dataset generalization than standard ML or LLM baselines in mental health prediction.
-
Rewarding Beliefs, Not Actions: Consistency-Guided Credit Assignment for Long-Horizon Agents
ReBel uses belief-consistency supervision and belief-aware grouping to improve credit assignment in long-horizon RL for LLM agents, achieving up to 20.4 percentage points higher success and 2.1x better sample efficiency than GRPO on ALFWorld and WebShop.
-
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.
-
Self-Supervised On-Policy Distillation for Reasoning Language Models
SSOPD converts intra-group correct-wrong contrast into process supervision by distilling a teacher distribution from the shortest correct completion into prefixes of the longest wrong completion, improving GRPO on AIME and HMMT benchmarks.
-
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.
-
SAGE: Shaping Anchors for Guided Exploration in RLVR of LLMs
SAGE reshapes the reverse-KL anchor via guide function q(x,y) for controllable empirical support expansion, yielding gains in both pass@1 and pass@k on math reasoning 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.
-
PreFT: Prefill-only finetuning for efficient inference
Prefill-only adaptation of LLMs yields 1.9x higher throughput for 512 adapters on Llama 3.1 70B with near-parity performance on RL tasks and recoverable loss on SFT.
-
N-vium: Mixture-of-Exits Transformer for Accelerated Exact Generation
N-vium achieves 57.9% wall-clock speedup over matched standard transformers at no perplexity cost by mixing exact predictions from multiple model depths.
-
Bayesian Model Merging
Bayesian Model Merging introduces a bi-level optimization framework that merges task-specific models via closed-form Bayesian regression with an anchor prior and global hyperparameter search, outperforming baselines and nearly matching expert averages on up to 20-task vision and 5-task language Merg
-
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.
-
Understanding and Preventing Entropy Collapse in RLVR with On-Policy Entropy Flow Optimization
OPEFO prevents entropy collapse in RLVR by rescaling token updates according to their entropy change contributions, yielding more stable optimization and better results on math benchmarks.
-
Annotations Mitigate Post-Training Mode Collapse
Annotation-anchored training reduces semantic diversity collapse in post-trained language models by a factor of six compared to standard supervised fine-tuning while preserving instruction-following and improving with scale.
-
dFlowGRPO: Rate-Aware Policy Optimization for Discrete Flow Models
dFlowGRPO is a new rate-aware RL method for discrete flow models that outperforms prior GRPO approaches on image generation and matches continuous flow models while supporting broad probability paths.
-
DeltaRubric: Generative Multimodal Reward Modeling via Joint Planning and Verification
DeltaRubric decomposes multimodal preference evaluation into self-generated planning and verification steps within a single model, producing large accuracy improvements on VL-RewardBench via multi-role reinforcement learning.
-
Reinforcing Multimodal Reasoning Against Visual Degradation
ROMA improves MLLM robustness to seen and unseen visual corruptions by +2.3-2.4% over GRPO on seven reasoning benchmarks while matching clean accuracy.
-
Dr. Post-Training: A Data Regularization Perspective on LLM Post-Training
Dr. Post-Training reframes general data as a data-induced regularizer for LLM post-training updates, yielding a family of methods that outperform data-selection baselines on SFT, RLHF, and RLVR tasks.
-
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.