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

TL;DR

This paper introduces Durian, a difficulty-aware group normalization method that improves multimodal reasoning in large language models by reducing sensitivity to extreme samples, leading to better performance on benchmarks.

Contribution

It proposes a novel normalization technique that groups samples by difficulty, enhancing stability and reasoning ability in multimodal large language models.

Findings

Significant performance improvements on multimodal reasoning benchmarks.

Durian reduces sensitivity to extreme samples and stabilizes training.

Effective in preserving intra-group distinctions while improving reasoning.

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.