Turning Off-Policy Tokens On-Policy: A Plug-in Approach for Improving LLM Alignment

• URL: http://arxiv.org/abs/2607.04728v1
• Date: Mon, 06 Jul 2026 06:59:44 GMT
• Title: Turning Off-Policy Tokens On-Policy: A Plug-in Approach for Improving LLM Alignment
• Abstract summary: Selective Importance Sampling (SIS) is inspired by rejection sampling.<n>SIS implements by viewing off-policy model as proposal distribution, and implement a token-level rejection test.<n>SIS consistently improves all objectives, while providing substantially stronger robustness under off-policy data.
• Score: 15.319597516845695
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reinforcement learning (RL) post-training for large language models (LLMs) follows a efficient paradigm of "rollout then update", which inevitably results in off-policy training data. To resolve this, Importance sampling (IS) is proposed, while the token-level ratios compound over long sequences, causing severe variance exploded. A natural idea is "transferring" these off-policy token into on-policy token, so that the importance scores for correction are unnecessary. Following this idea, we propose Selective Importance Sampling (SIS), which is inspired by rejection sampling. Concretely, SIS implements by viewing off-policy model as proposal distribution, and implement a token-level rejection test: accepted tokens are viewed as on-policy, so that receive unit importance score, while rejected tokens retain the standard IS correction. Our proposed SIS is theoretically proved reducing the gap between token-level and sequence-level off-policy gradient estimators. The SIS acts as a plug-in that only modifies the importance ratio in the policy loss, adding negligible wall-clock overhead, and can be combine with a vast vary of RL post-training algorithms. Experiments on dense and MoE LLMs across math and agent benchmarks show that SIS consistently improves all objectives, while providing substantially stronger robustness under off-policy data.

Related papers

• Trust the Batch, On- or Off-Policy: Adaptive Policy Optimization for RL Post-Training [50.86545293331458]
  Reinforcement learning is structurally harder than supervised learning because the policy changes the data distribution it learns from.<n>We present a simple yet effective batch-adaptive objective that replaces fixed clipping with the normalized effective sample size of the policy ratios.
  arXiv  Detail & Related papers  (2026-05-12T16:44:47Z)
• Rethinking Importance Sampling in LLM Policy Optimization: A Cumulative Token Perspective [22.848847562976633]
  Existing methods face a fundamental bias-variance dilemma: token-level IS ratios, as adopted by PPO and GRPO, introduce bias by ignoring prefix state distribution mismatch.<n>We propose CTPO, which combines the cumulative token IS ratio with position-adaptive clipping that scales log-space clip bounds according to the natural $sqrtt$ growth of the cumulative log-ratio.
  arXiv  Detail & Related papers  (2026-05-08T06:35:02Z)
• Unleashing Implicit Rewards: Prefix-Value Learning for Distribution-Level Optimization [74.91418266859297]
  Implicit process reward models (PRMs) provide fine-grained reward signals along the reasoning process.<n>PRMs suffer from a train-inference mismatch: training only constrains a sequence-level aggregate, whereas inference requires token-level scores to reflect local step quality.<n>We propose a novel Implicit Prefix-Value Reward Model (IPVRM), which directly learns a prefix-conditioned value function estimating the probability of eventual correctness.<n>We also propose Distribution-Level RL (DistRL), which computes TD advantages for both sampled tokens and high-probability candidate tokens.
  arXiv  Detail & Related papers  (2026-04-14T18:19:54Z)
• Improving Search Agent with One Line of Code [68.58667107354253]
  Tool-based Agentic Reinforcement Learning (TARL) has emerged as a promising paradigm for training search agents to interact with external tools.<n>We propose textbfSearch textbfAgent textbfPolicy textbfOptimization (textbfSAPO), which stabilizes training via a conditional token-level KL constraint.
  arXiv  Detail & Related papers  (2026-03-10T04:07:39Z)
• Rethinking the Trust Region in LLM Reinforcement Learning [72.25890308541334]
  Proximal Policy Optimization (PPO) serves as the de facto standard algorithm for Large Language Models (LLMs)<n>We propose Divergence Proximal Policy Optimization (DPPO), which substitutes clipping with a more principled constraint.<n>DPPO achieves superior training and efficiency compared to existing methods, offering a more robust foundation for RL-based fine-tuning.
  arXiv  Detail & Related papers  (2026-02-04T18:59:04Z)
• A Step Back: Prefix Importance Ratio Stabilizes Policy Optimization [58.116300485427764]
  Reinforcement learning post-training can elicit reasoning behaviors in large language models.<n> token-level correction often leads to unstable training dynamics when the degree of off-policyness is large.<n>We propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO)
  arXiv  Detail & Related papers  (2026-01-30T08:47:19Z)
• SOUP: Token-level Single-sample Mix-policy Reinforcement Learning for Large Language Models [67.41779761651924]
  SOUP is a framework that unifies off- and on-policy learning within individual samples at the token level.<n>It consistently outperforms standard on-policy training and existing off-policy extensions.
  arXiv  Detail & Related papers  (2026-01-29T09:56:15Z)
• ASPO: Asymmetric Importance Sampling Policy Optimization [31.38346888572171]
  The Importance Sampling (IS) ratios of positive-advantage tokens are mismatched, leading to unbalanced token weighting for positive and negative tokens.<n>This mismatch suppresses the update of low-probability tokens while over-amplifying already high-probability ones.<n>We propose Asymmetric Importance Sampling Policy Optimization (ASPO), which uses a simple yet effective strategy that flips the IS ratios of positive-advantage tokens.
  arXiv  Detail & Related papers  (2025-10-07T15:54:24Z)
• Reinforce-Ada: An Adaptive Sampling Framework for Reinforce-Style LLM Training [47.26632817047513]
  Reinforcement learning applied to large language models (LLMs) for reasoning tasks is often bottlenecked by unstable gradient estimates.<n>We propose Reinforce-Ada, an adaptive sampling framework for online RL post-training of LLMs.<n>Unlike conventional two-stage allocation methods, Reinforce-Ada interleaves estimation and sampling in an online successive elimination process.
  arXiv  Detail & Related papers  (2025-10-06T16:34:09Z)
• Mind the Gap: Data Rewriting for Stable Off-Policy Supervised Fine-Tuning [33.899779762210976]
  Supervised fine-tuning (SFT) of large language models can be viewed as an off-policy learning problem.<n>Existing methods mitigate this issue with KL penalties or clipping, which passively updates rather than actively reducing the gap.<n>We propose a simple yet effective data rewriting framework that proactively shrinks the policy gap before training.
  arXiv  Detail & Related papers  (2025-09-18T17:02:30Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.