aiqbench analysis
This research introduces Stepwise Marginal Information Gain (MIG), an intrinsic reward mechanism derived directly from the policy to better credit intermediate reasoning steps in large language and vision-language models. By avoiding reliance on external annotations or auxiliary models, this method demonstrates significant accuracy improvements over standard binary-reward training across multiple benchmarks. It represents a meaningful advancement in reinforcement learning paradigms by enabling more efficient, annotation-free optimization of model reasoning capabilities.
Stepwise Intrinsic Rewards for Reasoning in Large Language Models
arXiv:2602.01034v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) has become a widely used paradigm for improving the reasoning abilities of large language models (LLMs) and Vision-language models (VLMs). Sparse binary outcome rewards, however, score only final correctness and cannot identify which intermediate steps contributed to it; in multimodal tasks, they may also reward answers driven by linguistic priors rather than visual evidence. Process reward models (PRMs) densify supervision but usually require process annotations, auxiliary models, or inference-time search. In this paper, we introduce Stepwise Marginal Information Gain (MIG), an intrinsic process reward computed from the policy itself. MIG measures how each structured reasoning prefix changes the length-normalized, teacher-forced log-likelihood of the reference answer. A monotonic historical watermark rewards only new likelihood maxima, avoiding duplicate credit after sub-record detours. We combine this signal with outcome and format rewards and a gated self-distillation objective that retains only structurally valid and correct trajectories. For VLMs, a real-versus-blank likelihood gate down-weights rewards when answers remain predictable without the image. Across eight task-specific benchmarks, the full method exceeds outcome-only GRPO in every single-run comparison. In broad-data transfer, it improves average accuracy by up to 4.8 points over binary-reward training and gains 12.6 points on MathVerse. At 7B, it exceeds an external PRM-BoN@16 baseline by 12.9 points on vision-language transfer without inference-time reranking. These results support policy-derived stepwise credit as an annotation-free alternative to explicit process reward modeling.
Abstract:Reinforcement learning (RL) has become a widely used paradigm for improving the reasoning abilities of large language models (LLMs) and Vision-language models (VLMs). Sparse binary outcome rewards, however, score only final correctness and cannot identify which intermediate steps contributed to it; in multimodal tasks, they may also reward answers driven by linguistic priors rather than visual evidence. Process reward models (PRMs) densify supervision but usually require process annotations, auxiliary models, or inference-time search. In this paper, we introduce Stepwise Marginal Information Gain (MIG), an intrinsic process reward computed from the policy itself. MIG measures how each structured reasoning prefix changes the length-normalized, teacher-forced log-likelihood of the reference answer. A monotonic historical watermark rewards only new likelihood maxima, avoiding duplicate credit after sub-record detours. We combine this signal with outcome and format rewards and a gated self-distillation objective that retains only structurally valid and correct trajectories. For VLMs, a real-versus-blank likelihood gate down-weights rewards when answers remain predictable without the image. Across eight task-specific benchmarks, the full method exceeds outcome-only GRPO in every single-run comparison. In broad-data transfer, it improves average accuracy by up to 4.8 points over binary-reward training and gains 12.6 points on MathVerse. At 7B, it exceeds an external PRM-BoN@16 baseline by 12.9 points on vision-language transfer without inference-time reranking. These results support policy-derived stepwise credit as an annotation-free alternative to explicit process reward modeling.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2602.01034 [cs.AI]
(or arXiv:2602.01034v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2602.01034 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 323 words
Related coverage
RL & search is a terrifying way to build AGI (an FAQ)
AI Alignment Forum · Jul 27, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026
Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
arXiv · Sep 16, 2026
OpenAI's myopia just keeps causing alignment problems
LessWrong · Jul 27, 2026