Mitigating Overthinking through Reasoning Shaping

2025-10-10

AI Summary

Large reasoning models (LRMs) trained under reinforcement learning with verifier rewards (RLVR) suffer from “overthinking”—generating lengthy, inefficient reasoning traces that increase computational overhead and degrade performance. Existing token-level penalty methods impair accuracy due to coarse-grained supervision. To address this, we propose **Group Relative Segment Penalization (GRSP)**, a segment-level fine-grained regularization method: it clusters reasoning steps into semantically coherent segments and applies length-aware, dynamically weighted relative penalties at the segment level to enforce structured reasoning. GRSP significantly improves training stability and model scalability. Experiments demonstrate that GRSP reduces inference cost by 32%–47% on complex tasks while preserving or even improving accuracy—marking the first RLVR framework to jointly achieve high efficiency and reliability.

Technology Category

Knowledge Representation and Reasoning: Computational Complexity of ReasoningMachine Learning: Large Multimodal Models (LMMs)Natural Language Processing: Safety and Robustness

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsEconomics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large reasoning models (LRMs) boosted by Reinforcement Learning from Verifier Reward (RLVR) have shown great power in problem solving, yet they often cause overthinking: excessive, meandering reasoning that inflates computational cost. Prior designs of penalization in RLVR manage to reduce token consumption while often harming model performance, which arises from the oversimplicity of token-level supervision. In this paper, we argue that the granularity of supervision plays a crucial role in balancing efficiency and accuracy, and propose Group Relative Segment Penalization (GRSP), a step-level method to regularize reasoning. Since preliminary analyses show that reasoning segments are strongly correlated with token consumption and model performance, we design a length-aware weighting mechanism across segment clusters. Extensive experiments demonstrate that GRSP achieves superior token efficiency without heavily compromising accuracy, especially the advantages with harder problems. Moreover, GRSP stabilizes RL training and scales effectively across model sizes.

Problem

• Mitigating overthinking in large reasoning models
• Reducing excessive reasoning while maintaining accuracy
• Balancing computational efficiency with model performance

Innovation

• Step-level penalization method regularizes reasoning segments
• Length-aware weighting mechanism across segment clusters
• Stabilizes training and scales across model sizes