Rethinking Importance Sampling in LLM Policy Optimization: A Cumulative Token Perspective

• URL: http://arxiv.org/abs/2605.07331v1
• Date: Fri, 08 May 2026 06:35:02 GMT
• Title: Rethinking Importance Sampling in LLM Policy Optimization: A Cumulative Token Perspective
• Abstract summary: 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.
• Score: 22.848847562976633
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reinforcement learning, including reinforcement learning with verifiable rewards (RLVR), has emerged as a powerful approach for LLM post-training. Central to these approaches is the design of the importance sampling (IS) ratio used in off-policy policy-gradient estimation. Existing methods face a fundamental bias-variance dilemma: token-level IS ratios, as adopted by PPO (Schulman et al., 2017) and GRPO (Shao et al., 2024), introduce bias by ignoring prefix state distribution mismatch; full sequence ratios provide exact trajectory-level correction but suffer from high variance due to the multiplicative accumulation of per-token ratios, while GSPO (Zheng et al., 2025) improves numerical stability via length normalization at the cost of deviating from the exact full-sequence IS correction. In this work, we identify the cumulative token IS ratio, the product of per-token ratios up to position $t$, as a theoretically principled solution to this dilemma. We prove that, under the token-level policy-gradient formulation, this ratio provides an unbiased prefix correction for each token-level gradient term and has strictly lower variance than the full sequence ratio. Building on this insight, we propose CTPO (Cumulative Token Policy Optimization), which combines the cumulative token IS ratio with position-adaptive clipping that scales log-space clip bounds according to the natural $\sqrt{t}$ growth of the cumulative log-ratio. This yields more consistent regularization across token positions. We implement and evaluate CTPO in the tool-integrated reasoning setting on several challenging mathematical reasoning benchmarks, achieving the best average performance across both model scales compared with strong GRPO and GSPO baselines. Code will be available at https://github.com/horizon-llm/CTPO.

Related papers

• Holder Policy Optimisation [26.521180498291717]
  textbfHlderPO is a generalised policy optimisation framework.<n>It unifies token-level probability aggregation via the Hlder mean.<n>It achieves a state-of-the-art average accuracy of $54.9%$ across multiple mathematical benchmarks.
  arXiv  Detail & Related papers  (2026-05-12T12:45:03Z)
• Relative Score Policy Optimization for Diffusion Language Models [29.344961499429257]
  Diffusion large language models (dLLMs) offer a promising route to parallel and efficient text generation.<n>The lack of tractable sequence-level log-ratios forces existing methods to rely on high-variance ELBO-based approximations.<n>We propose textbfRelative textbfScore textbfPolicy textbfOptimization (RSPO) to calibrate noisy likelihood estimates in dLLMs.
  arXiv  Detail & Related papers  (2026-05-11T08:58:40Z)
• Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data [55.84428098924793]
  We propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration.<n>We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance.<n> Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO.
  arXiv  Detail & Related papers  (2026-04-20T16:43:28Z)
• Balanced Aggregation: Understanding and Fixing Aggregation Bias in GRPO [70.38763678943648]
  Reinforcement learning with verifiable rewards (RLVR) has become a central paradigm for improving reasoning and code generation in large language models.<n>Standard GRPO uses sequence aggregation, while recent work has advocated token aggregation as a better alternative.<n>We show that these two rules induce different optimization biases; token aggregation introduces sign-length coupling, while sequence aggregation implicitly downweights longer responses.
  arXiv  Detail & Related papers  (2026-04-14T09:48:46Z)
• From $\boldsymbol{\logπ}$ to $\boldsymbolπ$: Taming Divergence in Soft Clipping via Bilateral Decoupled Decay of Probability Gradient Weight [6.07556923512707]
  Reinforcement Learning with Verifiable Rewards (RLVR) has catalyzed a leap in Large Language Model (LLM) reasoning, yet its optimization dynamics remain fragile.<n>Standard algorithms like GRPO enforce stability via hard clipping'', which inadvertently stifles exploration by discarding gradients of tokens outside the trust region.<n>We propose Decoupled Gradient Policy Optimization (DGPO), which employs a decoupled decay mechanism based on importance sampling ratios.
  arXiv  Detail & Related papers  (2026-03-15T14:00:48Z)
• Unbiased Dynamic Pruning for Efficient Group-Based Policy Optimization [60.87651283510059]
  Group Relative Policy Optimization (GRPO) effectively scales LLM reasoning but incurs prohibitive computational costs.<n>We propose Dynamic Pruning Policy Optimization (DPPO), a framework that enables dynamic pruning while preserving unbiased gradient estimation.<n>To mitigate the data sparsity induced by pruning, we introduce Dense Prompt Packing, a window-based greedy strategy.
  arXiv  Detail & Related papers  (2026-03-04T14:48:53Z)
• Stability and Generalization of Push-Sum Based Decentralized Optimization over Directed Graphs [55.77845440440496]
  Push-based decentralized communication enables optimization over communication networks, where information exchange may be asymmetric.<n>We develop a unified uniform-stability framework for the Gradient Push (SGP) algorithm.<n>A key technical ingredient is an imbalance-aware generalization bound through two quantities.
  arXiv  Detail & Related papers  (2026-02-24T05:32:03Z)
• The Optimal Token Baseline: Variance Reduction for Long-Horizon LLM-RL [39.23942538769713]
  Reinforcement Learning for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance.<n>We derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm.<n>Our method achieves training stability and matches the performance of large group sizes with only $N=32$, reducing token consumption by over 65% across single-turn and tool-integrated reasoning tasks.
  arXiv  Detail & Related papers  (2026-02-06T03:16:04Z)
• 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)
• Segmental Advantage Estimation: Enhancing PPO for Long-Context LLM Training [17.530233901658253]
  Segmental Advantage Estimation mitigates the bias that Generalized Advantage Estimation can incur in Reinforcement Learning with Verifiable Rewards.<n> SAE achieves superior performance, with marked improvements in final scores, stability, and sample efficiency.
  arXiv  Detail & Related papers  (2026-01-12T08:41:47Z)
• The Role of Baselines in Policy Gradient Optimization [83.42050606055822]
  We show that the emphstate value baseline allows on-policy. emphnatural policy gradient (NPG) to converge to a globally optimal. policy at an $O (1/t) rate gradient. We find that the primary effect of the value baseline is to textbfreduce the aggressiveness of the updates rather than their variance.
  arXiv  Detail & Related papers  (2023-01-16T06:28:00Z)

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.