DISPO: Enhancing Training Efficiency and Stability in Reinforcement Learning for Large Language Model Mathematical Reasoning

• URL: http://arxiv.org/abs/2602.00983v1
• Date: Sun, 01 Feb 2026 02:45:04 GMT
• Title: DISPO: Enhancing Training Efficiency and Stability in Reinforcement Learning for Large Language Model Mathematical Reasoning
• Abstract summary: DISPO is a simple yet effective REINFORCE-style algorithm that decouples the up-clipping and down-clipping of importance sampling weights for correct and incorrect responses.<n>We show that DISPO achieves 61.04% on AIME'24 (vs. 55.42% CISPO and 50.21% DAPO) with similar gains across various benchmarks and models.
• Score: 31.369103012768964
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Reinforcement learning with verifiable rewards has emerged as a promising paradigm for enhancing the reasoning capabilities of large language models particularly in mathematics. Current approaches in this domain present a clear trade-off: PPO-style methods (e.g., GRPO/DAPO) offer training stability but exhibit slow learning trajectories due to their trust-region constraints on policy updates, while REINFORCE-style approaches (e.g., CISPO) demonstrate improved learning efficiency but suffer from performance instability as they clip importance sampling weights while still permitting non-zero gradients outside the trust-region. To address these limitations, we introduce DISPO, a simple yet effective REINFORCE-style algorithm that decouples the up-clipping and down-clipping of importance sampling weights for correct and incorrect responses, yielding four controllable policy update regimes. Through targeted ablations, we uncover how each regime impacts training: for correct responses, weights >1 increase the average token entropy (i.e., exploration) while weights <1 decrease it (i.e., distillation) -- both beneficial but causing gradual performance degradation when excessive. For incorrect responses, overly restrictive clipping triggers sudden performance collapse through repetitive outputs (when weights >1) or vanishing response lengths (when weights <1). By separately tuning these four clipping parameters, DISPO maintains the exploration-distillation balance while preventing catastrophic failures, achieving 61.04% on AIME'24 (vs. 55.42% CISPO and 50.21% DAPO) with similar gains across various benchmarks and models.

Related papers

• Extreme Region Policy Distillation [36.61472284280031]
  We show that aggressive multi-step optimization brings rapid initial gains, but excessive updates cause trajectory probabilities to deviate and entropy to collapse.<n>This motivates Extreme Region Policy Distillation (ERPD), a two-stage framework that decouples sample efficiency from KL efficiency.
  arXiv  Detail & Related papers  (2026-05-25T08:32:24Z)
• Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping [66.25536973294726]
  Post-training methods for text-to-image (T2I) models are prone to reward hacking.<n>Super-Linear Advantage Shaping (SLAS) reshapes the local policy space.<n>SLAS consistently surpasses the DanceGRPO baseline across multiple backbones and benchmarks.
  arXiv  Detail & Related papers  (2026-05-11T17:59:25Z)
• Learning from Mistakes: Negative Reasoning Samples Enhance Out-of-Domain Generalization [37.59050823977757]
  Supervised fine-tuning (SFT) on chain-of-thought (CoT) trajectories demonstrations is a common approach for enabling reasoning in large language models.<n>Standard practices typically only retain trajectories with correct final answers (positives) while ignoring the rest (negatives)<n>We argue that this paradigm discards substantial supervision and exacerbates overfitting, limiting out-of-domain (OOD) generalization.<n>We propose Gain-based LOss Weighting (GLOW), an adaptive, sample-aware scheme that exploits such distinctive training dynamics by rescaling per-sample loss based on inter-epoch progress
  arXiv  Detail & Related papers  (2026-01-08T14:49:10Z)
• Stabilizing Reinforcement Learning with LLMs: Formulation and Practices [61.361819972410046]
  We show why and under what conditions the true sequence-level reward can be optimized via a surrogate token-level objective in policy gradient methods such as REINFORCE.<n>This insight provides a principled explanation for the crucial role of several widely adopted techniques in stabilizing RL training.
  arXiv  Detail & Related papers  (2025-12-01T07:45:39Z)
• SG-OIF: A Stability-Guided Online Influence Framework for Reliable Vision Data [6.4391040754741296]
  In this paper, we introduce a Stability-Guided Online Influence Framework (SG-OIF) for Approximating training-point influence on test predictions.<n>We show that SG-OIF achieves 91.1% accuracy in the top 1% prediction samples on the CIFAR-10, and 99.8% AUPR score on MNIST.
  arXiv  Detail & Related papers  (2025-11-21T19:58:54Z)
• Continual Learning, Not Training: Online Adaptation For Agents [0.0]
  We introduce our Adaptive Teaching and Learning System (ATLAS), a dual-agent architecture that decouples reasoning (Teacher) from execution (Student)<n>ATLAS achieves gradient-free continual learning, shifting the locus of adaptation from model parameters to system-level orchestration.
  arXiv  Detail & Related papers  (2025-11-02T21:48:31Z)
• Scaf-GRPO: Scaffolded Group Relative Policy Optimization for Enhancing LLM Reasoning [49.290631188365786]
  Scaf-GRPO is a training framework that intervenes when a model's independent learning has plateaued.<n>It boosts the pass@1 score of the Qwen2.5-Math-7B model by a relative 44.3% over a vanilla GRPO baseline.<n>This result demonstrates our framework provides a robust and effective methodology for unlocking a model's ability to solve problems previously beyond its reach.
  arXiv  Detail & Related papers  (2025-10-22T17:41:30Z)
• BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping [69.74252624161652]
  We propose BAlanced Policy Optimization with Adaptive Clipping (BAPO)<n>BAPO dynamically adjusts clipping bounds to adaptively re-balance positive and negative contributions, preserve entropy, and stabilize RL optimization.<n>On AIME 2024 and AIME 2025 benchmarks, our 7B BAPO model surpasses open-source counterparts such as SkyWork-OR1-7B.
  arXiv  Detail & Related papers  (2025-10-21T12:55:04Z)
• 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)
• ResT: Reshaping Token-Level Policy Gradients for Tool-Use Large Language Models [62.82372407840088]
  Large language models (LLMs) transcend passive generation and act as goal-directed agents by invoking external tools.<n>textbfReshaped textbfToken-level policy gradients (textbfResT) for tool-use tasks.<n>textbfResT achieves state-of-the-art results, outperforming prior methods by up to $8.76%$.
  arXiv  Detail & Related papers  (2025-09-26T03:38:27Z)
• Stable Reinforcement Learning for Efficient Reasoning [2.838966689544288]
  GRPO-$lambda$ is an efficient and stabilized variant of GRPO.<n>It dynamically adjusts the reward strategy by monitoring the correctness ratio.<n>It improves average accuracy by 1.48% while reducing CoT sequence length by 47.3%.
  arXiv  Detail & Related papers  (2025-05-23T16:43:03Z)
• NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation [66.36912000442608]
  NoisyRollout is a simple yet effective data augmentation method.<n>It mixes training trajectories from both clean and moderately distorted images.<n>It achieves state-of-the-art performance among open-source RL-tuned models.
  arXiv  Detail & Related papers  (2025-04-17T16:10:13Z)
• Statistically Efficient Variance Reduction with Double Policy Estimation for Off-Policy Evaluation in Sequence-Modeled Reinforcement Learning [53.97273491846883]
  We propose DPE: an RL algorithm that blends offline sequence modeling and offline reinforcement learning with Double Policy Estimation. We validate our method in multiple tasks of OpenAI Gym with D4RL benchmarks.
  arXiv  Detail & Related papers  (2023-08-28T20:46:07Z)

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.