H-GRPO: Permutation-Invariant Reinforcement Learning for Grounded Visual Reasoning

2026-06-29

AI Summary

This work addresses the tendency of existing vision-language models to generate hallucinations or rely on superficial shortcuts during reasoning, which undermines their interpretability and ability to perform structured inference grounded in genuine visual evidence. To mitigate these issues, the authors propose a decompositional evidence localization framework that breaks complex questions into atomic sub-questions, each explicitly tied to localized visual regions in the image and resolved with unambiguous answers, thereby constructing human-like logical reasoning chains. The approach integrates a decomposition-based reasoning architecture, fine-grained bounding box annotations for local evidence, and a reinforcement learning–driven sequential decision policy to ensure strict alignment between reasoning steps and visual content. Experiments demonstrate that this framework substantially improves accuracy and robustness on challenging visual question answering tasks while producing interpretable and verifiable reasoning trajectories, effectively alleviating hallucination and dataset bias.

Abstract

Vision-Language Models (VLMs) often achieve high performance on benchmarks while remaining "black boxes", yet they remain prone to hallucination or rely on superficial shortcuts. In this work, we propose a framework designed to enhance both performance and interpretability through De-compositional Evidence Grounding. Unlike monolithic inference approaches, our approach forces the model to decompose a global query into a sequence of atomic sub-questions, each requiring an explicit sub-answer and critically a localized evidence bounding box. By grounding intermediate logical steps (e.g. identifying a container, analyzing liquid properties, and assessing environmental context) in specific visual regions, we construct a structured reasoning path that mirrors human-like deduction. This allows the final answer to emerge as a logical consequence of verified visual facts rather than a statistical guess.

Problem

• Vision-Language Models
• hallucination
• interpretability
• visual reasoning
• evidence grounding

Innovation

• Permutation-Invariant Reinforcement Learning
• Decompositional Evidence Grounding
• Visual Reasoning
• Vision-Language Models
• Interpretable AI