New method enhances LLM reasoning with intrinsic rewards

Researchers have developed a new method called Stepwise Marginal Information Gain (MIG) to improve the reasoning capabilities of large language models (LLMs) and vision-language models (VLMs). MIG calculates an intrinsic reward based on how each step in a reasoning process increases the likelihood of the correct answer, offering a more granular feedback mechanism than traditional sparse binary rewards. This approach has demonstrated significant improvements across various benchmarks, outperforming outcome-only rewards and even surpassing external Process Reward Models (PRMs) in certain vision-language tasks without requiring additional inference-time processing. AI

IMPACT This new intrinsic reward method could lead to more efficient training of LLMs and VLMs, improving their reasoning capabilities without extensive manual annotation.

RANK_REASON The cluster describes a new method published in an academic paper for improving LLM reasoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• GRPO
• large language models
• MathVerse
• PRM-BoN
• Process reward models
• Stepwise Marginal Information Gain
• vision-language models
• Xiangwei Wang

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →