SGPO extracts strategies from strong-model responses, builds autonomous and guided trajectories, and applies token-level forward-KL distillation with adaptive weighting to outperform SFT and RL baselines by 2.2 points on math benchmarks.
hub Canonical reference
Learning to Reason under Off-Policy Guidance
Canonical reference. 89% of citing Pith papers cite this work as background.
55
Pith papers citing it
Background
89% of classified citations
abstract
Recent advances in large reasoning models (LRMs) demonstrate that sophisticated behaviors such as multi-step reasoning and self-reflection can emerge via reinforcement learning with verifiable rewards~(\textit{RLVR}). However, existing \textit{RLVR} approaches are inherently ``on-policy'', limiting learning to a model's own outputs and failing to acquire reasoning abilities beyond its initial capabilities. To address this issue, we introduce \textbf{LUFFY} (\textbf{L}earning to reason \textbf{U}nder o\textbf{FF}-polic\textbf{Y} guidance), a framework that augments \textit{RLVR} with off-policy reasoning traces. LUFFY dynamically balances imitation and exploration by combining off-policy demonstrations with on-policy rollouts during training. Specifically, LUFFY combines the Mixed-Policy GRPO framework, which has a theoretically guaranteed convergence rate, alongside policy shaping via regularized importance sampling to avoid superficial and rigid imitation during mixed-policy training. Compared with previous RLVR methods, LUFFY achieves an over \textbf{+6.4} average gain across six math benchmarks and an advantage of over \textbf{+6.2} points in out-of-distribution tasks. Most significantly, we show that LUFFY successfully trains weak models in scenarios where on-policy RLVR completely fails. These results provide compelling evidence that LUFFY transcends the fundamental limitations of on-policy RLVR and demonstrates the great potential of utilizing off-policy guidance in RLVR.
hub tools
citation-role summary
background 8
baseline 1
citation-polarity summary
representative citing papers
SWITCH uses explicit <swi> and </swi> boundary tokens to make latent chain-of-thought compatible with on-policy RL (GRPO) and open to causal mechanistic probing, outperforming prior hidden-state recurrence methods.
ATLAS traces RLVR data to 20 atomic sources, most datasets are variants, and DAPO++ curated with SCA improves RLVR performance while Q predicts training effectiveness.
ActGuide-RL uses human action data as plan-style guidance in mixed-policy RL to overcome exploration barriers in LLM agents, matching SFT+RL performance on search benchmarks without cold-start training.
The cumulative token IS ratio gives unbiased prefix correction and lower variance than full-sequence ratios for token-level gradients in LLM policy optimization, enabling CTPO to outperform GRPO and GSPO baselines on mathematical reasoning tasks.
NPO uses a policy's own near-future checkpoint as auxiliary trajectories to maximize effective learning signal S = Q/V, improving performance from 57.88 to 63.15 on Qwen3-VL-8B-Instruct with GRPO while accelerating convergence.
NExt accelerates RLVR training for LLMs by nonlinearly extrapolating low-rank parameter trajectories extracted from LoRA runs.
OPD for LLMs suffers length inflation and repetition collapse; StableOPD uses reference divergence and rollout mixing to prevent it and improve math reasoning performance by 7.2% on average.
GPS trains a small model on optimization history to predict prompt difficulty and select intermediate-difficulty diverse batches, yielding better training efficiency, final performance, and test-time allocation than baselines on reasoning benchmarks.
Prefix-RFT blends SFT and RFT via prefix sampling from demonstrations to outperform standalone SFT, RFT, and mixed-policy baselines on math reasoning problems.
AdaPrefix-GRPO treats solution-prefix length as a feedback controller targeting 50% rollout success rate during GRPO training, then anneals to zero prefix, yielding 1.6–2.1× accuracy gains over vanilla GRPO at matched compute on hard math.
Pairing outcome rewards with verifiable per-action path penalties reduces constraint violations nearly sixfold at equal task success, while a progress potential accelerates learning only where partial progress is reachable.
Theoretical analysis of RLVR update dynamics leads to ACPO, an adaptive clipping method that outperforms DAPO and CISPO on reasoning benchmarks with 3B and 7B models.
Comparative weight-space analysis finds SFT/RFT/RIFT colinear with similar accuracy, DFT more divergent, GRPO partially orthogonal, and DPO near-orthogonal with highest GSM8K/AIME accuracy but using 10x smaller learning rate.
Boundary-aware Curriculum RL raises average pass@256 by 9.8 points over base models and 10.3 points over vanilla RLVR on Qwen, Llama, and DeepSeek families.
RASFT is an adaptive SFT method that strengthens or relaxes expert imitation per problem based on on-policy rollout solvability and adds clipped reference-policy ratio to limit drift, reporting better results than standard SFT and RL on math and code benchmarks.
Experiments indicate RL applied early in pre-training often matches full SFT-then-RL performance, targeted data composition outweighs scale for RL success, and averaging RL and SFT objectives outperforms sequential or single methods.
HDPO distills 'propose-select-think' reasoning into LLMs and uses GRPO-style RL with scheduled diversity and entropy-based reliability rewards, improving average benchmark scores by about 7 points over GRPO on 4B–8B models.
DelTA estimates token coefficients to amplify discriminative directions in token-gradient vectors, reweighting the RLVR surrogate to produce more contrastive side-wise centroids and yielding 3.26 and 2.62 point gains on math benchmarks for 8B and 14B Qwen3 models.
DMPO approximates forward KL minimization in on-policy RL by aligning the policy to a group-level reward-proportional target distribution, yielding 9-12% relative gains over GRPO on NP-Bench and smaller gains on math reasoning.
HASP upgrades textual skills into executable Program Functions that intervene in LLM agent loops at inference, post-training, or self-evolution, delivering 25% gains over ReAct and 30.4% over Search-R1 on reasoning benchmarks.
VSPO samples rollouts at varying steering intensities to improve behavioral control in LLMs while preserving task accuracy.
STRIDE co-trains generator and verifier on outcome rewards alone to deliver learnable stepwise language feedback that redirects LLM reasoning trajectories and outperforms scalar-reward baselines.
TGPO improves on-policy reasoning distillation in LLMs by using teacher-guided token generation conditioned on student contexts together with trajectory rewards, outperforming RKL-based methods under large policy divergence.
citing papers explorer
-
Beyond Trajectory Imitation: Strategy-Guided Policy Optimization for LLM Reasoning
SGPO extracts strategies from strong-model responses, builds autonomous and guided trajectories, and applies token-level forward-KL distillation with adaptive weighting to outperform SFT and RL baselines by 2.2 points on math benchmarks.
-
Demystifying Hidden-State Recurrence: Switchable Latent Reasoning with On-Policy Reinforcement Learning
SWITCH uses explicit <swi> and </swi> boundary tokens to make latent chain-of-thought compatible with on-policy RL (GRPO) and open to causal mechanistic probing, outperforming prior hidden-state recurrence methods.
-
RLVR Datasets and Where to Find Them: Tracing Data Lineage for Better Training Data
ATLAS traces RLVR data to 20 atomic sources, most datasets are variants, and DAPO++ curated with SCA improves RLVR performance while Q predicts training effectiveness.
-
Learning Agentic Policy from Action Guidance
ActGuide-RL uses human action data as plan-style guidance in mixed-policy RL to overcome exploration barriers in LLM agents, matching SFT+RL performance on search benchmarks without cold-start training.
-
Rethinking Importance Sampling in LLM Policy Optimization: A Cumulative Token Perspective
The cumulative token IS ratio gives unbiased prefix correction and lower variance than full-sequence ratios for token-level gradients in LLM policy optimization, enabling CTPO to outperform GRPO and GSPO baselines on mathematical reasoning tasks.
-
Near-Future Policy Optimization
NPO uses a policy's own near-future checkpoint as auxiliary trajectories to maximize effective learning signal S = Q/V, improving performance from 57.88 to 63.15 on Qwen3-VL-8B-Instruct with GRPO while accelerating convergence.
-
Low-rank Optimization Trajectories Modeling for LLM RLVR Acceleration
NExt accelerates RLVR training for LLMs by nonlinearly extrapolating low-rank parameter trajectories extracted from LoRA runs.
-
Demystifying OPD: Length Inflation and Stabilization Strategies for Large Language Models
OPD for LLMs suffers length inflation and repetition collapse; StableOPD uses reference divergence and rollout mixing to prevent it and improve math reasoning performance by 7.2% on average.
-
Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models
GPS trains a small model on optimization history to predict prompt difficulty and select intermediate-difficulty diverse batches, yielding better training efficiency, final performance, and test-time allocation than baselines on reasoning benchmarks.
-
Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling
Prefix-RFT blends SFT and RFT via prefix sampling from demonstrations to outperform standalone SFT, RFT, and mixed-policy baselines on math reasoning problems.
-
Max Out GRPO Signal: Adaptive Trace Prefix Control for Hard Reasoning Problems
AdaPrefix-GRPO treats solution-prefix length as a feedback controller targeting 50% rollout success rate during GRPO training, then anneals to zero prefix, yielding 1.6–2.1× accuracy gains over vanilla GRPO at matched compute on hard math.
-
RLVP: Penalize the Path, Reward the Outcome
Pairing outcome rewards with verifiable per-action path penalties reduces constraint violations nearly sixfold at equal task success, while a progress potential accelerates learning only where partial progress is reachable.
-
What are Key Factors for Updates in RL for LLM Reasoning?
Theoretical analysis of RLVR update dynamics leads to ACPO, an adaptive clipping method that outperforms DAPO and CISPO on reasoning benchmarks with 3B and 7B models.
-
Weight-Space Geometry of Offline Reasoning Training
Comparative weight-space analysis finds SFT/RFT/RIFT colinear with similar accuracy, DFT more divergent, GRPO partially orthogonal, and DPO near-orthogonal with highest GSM8K/AIME accuracy but using 10x smaller learning rate.
-
Curriculum Reinforcement Learning Can Incentivize Reasoning Capacity in LLMs Beyond the Base Model
Boundary-aware Curriculum RL raises average pass@256 by 9.8 points over base models and 10.3 points over vanilla RLVR on Qwen, Llama, and DeepSeek families.
-
RASFT: Rollout-Adaptive Supervised Fine-Tuning for Reasoning
RASFT is an adaptive SFT method that strengthens or relaxes expert imitation per problem based on on-policy rollout solvability and adds clipped reference-policy ratio to limit drift, reporting better results than standard SFT and RL on math and code benchmarks.
-
RL Excursions during Pre-Training: Re-examining Policy Optimization for LLM training
Experiments indicate RL applied early in pre-training often matches full SFT-then-RL performance, targeted data composition outweighs scale for RL success, and averaging RL and SFT objectives outperforms sequential or single methods.
-
Hint-Guided Diversified Policy Optimization for LLM Reasoning
HDPO distills 'propose-select-think' reasoning into LLMs and uses GRPO-style RL with scheduled diversity and entropy-based reliability rewards, improving average benchmark scores by about 7 points over GRPO on 4B–8B models.
-
DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards
DelTA estimates token coefficients to amplify discriminative directions in token-gradient vectors, reweighting the RLVR surrogate to produce more contrastive side-wise centroids and yielding 3.26 and 2.62 point gains on math benchmarks for 8B and 14B Qwen3 models.
-
Beyond Mode Collapse: Distribution Matching for Diverse Reasoning
DMPO approximates forward KL minimization in on-policy RL by aligning the policy to a group-level reward-proportional target distribution, yielding 9-12% relative gains over GRPO on NP-Bench and smaller gains on math reasoning.
-
Harnessing LLM Agents with Skill Programs
HASP upgrades textual skills into executable Program Functions that intervene in LLM agent loops at inference, post-training, or self-evolution, delivering 25% gains over ReAct and 30.4% over Search-R1 on reasoning benchmarks.
-
VSPO: Vector-Steered Policy Optimization for Behavioral Control
VSPO samples rollouts at varying steering intensities to improve behavioral control in LLMs while preserving task accuracy.
-
STRIDE: Learnable Stepwise Language Feedback for LLM Reasoning
STRIDE co-trains generator and verifier on outcome rewards alone to deliver learnable stepwise language feedback that redirects LLM reasoning trajectories and outperforms scalar-reward baselines.
-
Teacher-Guided Policy Optimization for On-Policy Reasoning Distillation under Large Policy Divergence
TGPO improves on-policy reasoning distillation in LLMs by using teacher-guided token generation conditioned on student contexts together with trajectory rewards, outperforming RKL-based methods under large policy divergence.
-
Multi-Rollout On-Policy Distillation via Peer Successes and Failures
MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.
-
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.
-
Hidden States Know Where Reasoning Diverges: Credit Assignment via Span-Level Wasserstein Distance
Span-level Wasserstein distances between hidden-state distributions of correct and incorrect rollouts provide a self-supervised signal to reweight advantages in GRPO, improving fine-grained credit assignment on math and code tasks.
-
SPS: Steering Probability Squeezing for Better Exploration in Reinforcement Learning for Large Language Models
SPS interleaves RL and IRL to counteract probability squeezing in LLM reasoning trajectories, improving Pass@k on five benchmarks while identifying an empirical upper bound on multi-sample performance.
-
From $P(y|x)$ to $P(y)$: Investigating Reinforcement Learning in Pre-train Space
PreRL applies reward-driven updates to P(y) in pre-train space, uses Negative Sample Reinforcement to prune bad reasoning paths and boost reflection, and combines with standard RL in Dual Space RL to outperform baselines on reasoning tasks.
-
Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation
Lightning OPD is an offline on-policy distillation method that matches standard OPD performance at 4x efficiency by enforcing teacher consistency between SFT and distillation phases.
-
TensorHub: Scalable and Elastic Weight Transfer for LLM RL Training
TensorHub uses Reference-Oriented Storage to enable scalable weight transfer in LLM RL training by referencing replicated GPU weights, achieving up to 19x reduction in cross-datacenter stall time.
-
Can LLMs Learn to Reason Robustly under Noisy Supervision?
Online Label Refinement lets LLMs learn robust reasoning from noisy supervision by correcting labels when majority answers show rising rollout success and stable history, delivering 3-4% gains on math and reasoning benchmarks even at high noise levels.
-
Hindsight-Anchored Policy Optimization: Turning Failure into Feedback in Sparse Reward Settings
HAPO adds a hindsight-anchored SSI operator with Thompson gating to GRPO-style RLVR, achieving asymptotic consistency that recovers unbiased on-policy gradients as the policy improves.
-
No More Stale Feedback: Co-Evolving Critics for Open-World Agent Learning
ECHO jointly optimizes policy and critic via co-evolution, cascaded rollouts, and saturation-aware shaping to deliver non-stale feedback and higher success in open-world LLM agent RL.
-
SPHINX: A Synthetic Environment for Visual Perception and Reasoning
SPHINX generates synthetic visual puzzles for benchmarking LVLMs, where GPT-5 scores 51.1% and RLVR training improves both in-domain and external visual reasoning performance.
-
InternBootcamp Technical Report: Boosting LLM Reasoning with Verifiable Task Scaling
InternBootcamp supplies 1000+ verifiable, auto-generated task environments across domains that enable task scaling to improve LLM reasoning, producing a 32B model with state-of-the-art results on the new Bootcamp-EVAL benchmark.
-
EvoCoT: Overcoming the Exploration Bottleneck in Reinforcement Learning
EvoCoT uses self-generated and verified CoT trajectories in a two-stage curriculum to let LLMs learn from initially unsolved hard problems in RLVR settings.
-
RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization
RL-PLUS is a hybrid RL approach for LLMs that combines internal exploitation with external data via importance sampling and exploration advantages to prevent capability boundary collapse and achieve gains on math and OOD reasoning benchmarks.
-
Modularized Reinforcement Learning on LLMs: From MDP Creation to Exploration and Learning
Survey mapping RL techniques onto LLM training and highlighting gaps in value-based, off-policy, and bootstrapping methods.
-
GeoMin: Data-Efficient Semi-Supervised RLVR via Geometric Distribution Modeling
GeoMin uses geometric distribution modeling on labeled data to assess self-reward reliability, enabling better performance in semi-supervised RLVR with only 10% of typical annotations.
-
RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning
POPO uses recency-based prioritized group replay and decoupled off-policy optimization to avoid zero-variance ineffective samples in RLVR, accelerating LLM reasoning finetuning with fewer rollouts.
-
Trust Region On-Policy Distillation
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
-
GAC: Noise-Aware Adaptive Mixing for Hybrid SFT-RL Post-Training
GAC derives adaptive mixing weights for SFT-RL hybrid post-training from online gradient variance and signal disagreement estimates, improving benchmark performance over fixed schedules with under 1% overhead.
-
Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning
MARS introduces mono-anchored advantage normalization to quantify information gain from multi-source integration in RLVR, yielding 3.2% and 4.9% gains on GRPO and DAPO.
-
Hide to Guide: Learning via Semantic Masking
SMEPO applies fine-grained semantic masking to expert guidance in RLVR, turning hard problems into fill-in-the-blank tasks while preserving structure, yielding up to 3.2 point accuracy gains and 4.2x faster training.
-
Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning
Sequential SFT followed by RL, guided by the Plasticity-Ceiling Framework, achieves higher performance ceilings in LLM mathematical reasoning than synchronized methods by optimizing data scale and transition timing.
-
Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Exploration
DARS adaptively increases rollouts on hard problems in RLVR to improve Pass@K, and when paired with batch scaling for breadth, achieves gains in both Pass@K and Pass@1 by treating depth and breadth as complementary exploration dimensions.
-
Sample-efficient LLM Optimization with Reset Replay
LoRR augments preference optimization methods like DPO with high-replay training, periodic resets to initial data/policy, and a hybrid objective to improve sample efficiency and reduce primacy bias on math and reasoning tasks.
-
BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards
BV-Blend blends prompt-local and semantic-cluster historical reward statistics via SEM-derived weights to stabilize critic-free RL advantage estimation.
-
Smart Picks in the Dark: Towards Efficient RLVR for Reasoning via Tracing Metacognitive Pivots
PivotTrace selects unlabeled data for RLVR by quantifying uncertainty via pivot density from attention dynamics, outperforming full supervision using only 29.3% annotations and converging 2.75 times faster.