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.
hub
Tulu 3: Pushing Frontiers in Open Language Model Post-Training
62 Pith papers cite this work. Polarity classification is still indexing.
62
Pith papers citing it
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 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
co-cited works
roles
background 1polarities
background 1representative citing papers
dGRPO merges outcome-based policy optimization with dense teacher guidance from on-policy distillation, yielding more stable long-context reasoning on the new LongBlocks synthetic dataset.
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.
CoT-PoT ensembling achieves self-consistency accuracy in LLMs with only two samples for 78.6% of tasks, reducing computation by 9.3x compared to standard methods.
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.
SUPERNOVA adapts instruction-tuning data for RLVR and achieves up to 52.8% relative gains on general reasoning benchmarks like BBEH through targeted task selection and mixing.
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.
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 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
Fast-Slow Training combines slow parameter updates with fast context optimization to achieve up to 3x better sample efficiency, higher performance, less forgetting, and preserved plasticity in continual LLM learning.
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.
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.
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.
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 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 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.
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 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.
citing papers explorer
-
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.
-
Combining On-Policy Optimization and Distillation for Long-Context Reasoning in Large Language Models
dGRPO merges outcome-based policy optimization with dense teacher guidance from on-policy distillation, yielding more stable long-context reasoning on the new LongBlocks synthetic dataset.
-
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.
-
Self-Consistency from Only Two Samples: CoT-PoT Ensembling for Efficient LLM Reasoning
CoT-PoT ensembling achieves self-consistency accuracy in LLMs with only two samples for 78.6% of tasks, reducing computation by 9.3x compared to standard methods.
-
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.
-
SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions
SUPERNOVA adapts instruction-tuning data for RLVR and achieves up to 52.8% relative gains on general reasoning benchmarks like BBEH through targeted task selection and mixing.
-
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.
-
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
-
Learning, Fast and Slow: Towards LLMs That Adapt Continually
Fast-Slow Training combines slow parameter updates with fast context optimization to achieve up to 3x better sample efficiency, higher performance, less forgetting, and preserved plasticity in continual LLM learning.
-
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.
-
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.
-
Power Distribution Bridges Sampling, Self-Reward RL, and Self-Distillation
The power distribution is the target of power sampling, the closed-form solution to self-reward KL-regularized RL, and the basis for power self-distillation that matches sampling performance at lower cost.
-
MolmoAct2: Action Reasoning Models for Real-world Deployment
MolmoAct2 is an open VLA model that outperforms baselines like Pi-05 on 7 benchmarks and whose backbone surpasses GPT-5 on 13 embodied-reasoning tasks through new datasets, specialized training, and architecture changes for lower latency.
-
Diversity in Large Language Models under Supervised Fine-Tuning
TOFU loss mitigates the narrowing of generative diversity in LLMs after supervised fine-tuning by addressing neglect of low-frequency patterns and forgetting of prior knowledge.
-
Kernelized Advantage Estimation: From Nonparametric Statistics to LLM Reasoning
Kernel smoothing yields accurate value and gradient estimates for low-variance policy learning in LLM reasoning under tight per-prompt sampling budgets.
-
When Errors Can Be Beneficial: A Categorization of Imperfect Rewards for Policy Gradient
Certain errors in proxy rewards for policy gradient methods can be benign or beneficial by preventing policies from stalling on outputs with mediocre ground truth rewards, enabling improved RLHF metrics and reward design insights.
-
What Makes Good Instruction-Tuning Data? An In-Context Learning Perspective
A weighted in-context influence metric selects effective instruction-tuning data, outperforming baselines while showing that harder samples have lower influence.
-
CharTide: Data-Centric Chart-to-Code Generation via Tri-Perspective Tuning and Inquiry-Driven Evolution
CharTide decouples chart-to-code data into three perspectives and uses inquiry-driven RL with atomic QA verification to let smaller VLMs surpass GPT-4o on chart-to-code tasks.
-
Understanding and Mitigating Spurious Signal Amplification in Test-Time Reinforcement Learning for Math Reasoning
DDRL reduces spurious reward noise in test-time RL for math by excluding ambiguous samples, using fixed advantages, and adding consensus-based updates, outperforming prior TTRL methods on math benchmarks.
-
SSL-R1: Self-Supervised Visual Reinforcement Post-Training for Multimodal Large Language Models
SSL-R1 reformulates visual SSL tasks into verifiable puzzles to supply rewards for RL post-training of MLLMs, yielding gains on multimodal benchmarks without external supervision.
-
Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts
BAR trains independent domain experts via separate mid-training, SFT, and RL pipelines then composes them with a MoE router to match monolithic retraining performance at lower cost and without catastrophic forgetting.
-
HEALing Entropy Collapse: Enhancing Exploration in Few-Shot RLVR via Hybrid-Domain Entropy Dynamics Alignment
HEAL mitigates entropy collapse in few-shot RLVR by selectively adding general-domain data and aligning trajectory-level entropy dynamics, matching full-shot performance with 32 target samples.
-
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.
-
GroupDPO: Memory efficient Group-wise Direct Preference Optimization
GroupDPO decouples group-wise preference optimization during backpropagation to cut peak memory while keeping the same gradients, allowing larger groups and consistent gains over single-pair DPO plus an NLL term on positives.
-
Eliciting Medical Reasoning with Knowledge-enhanced Data Synthesis: A Semi-Supervised Reinforcement Learning Approach
MedSSR improves LLM medical reasoning on rare diseases by up to 5.93% through knowledge-enhanced question synthesis and semi-supervised RL with self-generated pseudo-labels.
-
The Art of (Mis)alignment: How Fine-Tuning Methods Effectively Misalign and Realign LLMs in Post-Training
ORPO is most effective at misaligning LLMs while DPO excels at realigning them, though it reduces utility, revealing an asymmetry between attack and defense methods.
-
Target Policy Optimization
TPO constructs a target distribution q proportional to the old policy times exp(utility) and trains the policy to match it via cross-entropy, matching or beating PPO and GRPO especially under sparse rewards.
-
OASES: Outcome-Aligned Search-Evaluation Co-Training for Agentic Search
OASES co-trains search policies and evaluators to generate outcome-aligned process rewards, outperforming standard RL baselines on five multi-hop QA benchmarks.
-
Memory in the Age of AI Agents
The paper maps agent memory research via three forms (token-level, parametric, latent), three functions (factual, experiential, working), and dynamics of formation/evolution/retrieval, plus benchmarks and future directions.
-
Dream 7B: Diffusion Large Language Models
Dream 7B is a 7B diffusion LLM that refines sequences in parallel via denoising and outperforms prior diffusion models on general, mathematical, and coding benchmarks with added flexibility in generation order and quality-speed tradeoffs.
-
Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains
RaR uses aggregated rubric feedback as rewards in on-policy RL, delivering up to 31% relative gains on HealthBench and 7% on GPQA-Diamond versus direct Likert LLM-as-judge baselines.
-
Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning
High-entropy minority tokens drive RLVR gains, so restricting gradients to the top 20% maintains or improves performance over full updates on Qwen3 models, especially larger ones.
-
Visual-RFT: Visual Reinforcement Fine-Tuning
Visual-RFT applies reinforcement learning with verifiable perception rewards to improve large vision-language models on fine-grained classification, few-shot detection, and grounding tasks.
-
Process Reinforcement through Implicit Rewards
PRIME enables online process reward model updates in LLM RL using implicit rewards from rollouts and outcome labels, yielding 15.1% average gains on reasoning benchmarks and surpassing a stronger instruct model with 10% of the data.
-
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.