Large reasoning models reduce unnecessary steps using causal intervention
Learning to Reason Efficiently with Interventional Step Credit
Computation and LanguageMachine Learning
Summary
Long reasoning chains in AI often contain useful and unnecessary steps mixed together. The authors introduce a new way to judge the value of each step by measuring how changing it affects the final answer, rather than just looking at the step itself. Using this method, they propose a way to train models that cut out redundant steps while keeping important ones, making reasoning shorter and better. Their approach was tested on many tasks and showed significant step reduction with improved performance.
What this means in practice
- •For ai model developers: Train reasoning models that automatically skip unneeded steps to produce answers more efficiently while maintaining accuracy.
- •For natural language processing engineers: Optimize chain-of-thought outputs in NLP systems to speed up processing and reduce computational costs.
Authors
Guanda Li, Xuefeng Bai, Yulong Chen, Zhiyong Wang, Kehai Chen, Min Zhang
Abstract
Large reasoning models often produce lengthy chain-of-thought (CoT) traces mixing useful intermediate steps with unnecessary ones. Recent methods explore step-level signals such as attention mass and step entropy to identify and suppress unnecessary steps. Despite their effectiveness, such signals reflect the model's behavior at a step rather than that step's contribution to the answer. To address this, we propose the interventional step credit (ISC), which takes a causal-intervention view and assesses step importance to the reasoning through targeted parameter-level interventions. Building on ISC, we further propose IGPO, which uses ISC to guide fine-grained advantage rescaling in reinforcement learning, reducing redundant steps while preserving the important ones. Experiments across two model scales and nine benchmarks show that IGPO achieves up to 62% average token reduction and better overall performance among compared methods.