Enhancing Multi-Modal LLMs Reasoning via Difficulty-Aware Group Normalization

2026-02-25

AI Summary

This work addresses the instability in training multimodal large language models with standard deviation–based reward normalization methods (e.g., GRPO), which is primarily caused by the coupled effects of perception and reasoning errors that amplify the influence of extreme reward samples. To mitigate this issue, the authors propose Durian, a difficulty-aware grouped normalization mechanism that explicitly models sample difficulty for the first time. Specifically, visual entropy quantifies perceptual complexity, while model confidence captures reasoning uncertainty; samples are then grouped according to these metrics, and reward normalization is performed using a shared standard deviation within each group. This approach effectively suppresses the adverse impact of outlier samples, significantly improving performance across multiple multimodal reasoning benchmarks while enhancing both training stability and the model’s reasoning capabilities.

Technology Category

• Machine Learning: Multimodal Learning
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Computer Vision: Diffusion Models for Vision

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO) have significantly advanced the reasoning capabilities of large language models. Extending these methods to multimodal settings, however, faces a critical challenge: the instability of std-based normalization, which is easily distorted by extreme samples with nearly positive or negative rewards. Unlike pure-text LLMs, multimodal models are particularly sensitive to such distortions, as both perceptual and reasoning errors influence their responses. To address this, we characterize each sample by its difficulty, defined through perceptual complexity (measured via visual entropy) and reasoning uncertainty (captured by model confidence). Building on this characterization, we propose difficulty-aware group normalization (Durian), which re-groups samples by difficulty levels and shares the std within each group. Our approach preserves GRPO's intra-group distinctions while eliminating sensitivity to extreme cases, yielding significant performance gains across multiple multimodal reasoning benchmarks.

Problem

• multi-modal LLMs
• reasoning instability
• std-based normalization
• extreme samples
• reward distortion

Innovation

• difficulty-aware group normalization
• multimodal reasoning
• GRPO
• visual entropy
• reinforcement learning