arXiv：cs.AI· Kun Zhao, Guodong Liu, Hui Ji, Siyuan Dai, Pan Wang, Jifeng Song, Heng Huang, Chenghua Lin, Liang Zhan, Haoteng Tang·· 20 小时前
HERO：面向放射报告生成的层次化证据推理优化框架
HERO: Hierarchical Evidential Reasoning Optimization for Radiology Report Generation via Reason-then-Summarize
AI 导读
HERO 是一个因子化策略优化框架，通过段落级、token 级和完成级三种粒度，分别优化推理、诊断与推理依据忠实度，以解决 GRPO 在医学监督中均匀分配信用导致的片段干扰与推理-诊断脱节问题。在 MIMIC-CXR 和 IU-Xray 上，HERO 超越强监督和强化学习基线，达到 SOTA 临床效能，并生成更具推理依据支持、think-answer 更一致的报告，显著缓解临床幻觉。
正文
Abstract:Multimodal Large Language Models (MLLMs) have substantially advanced Radiology Report Generation (RRG), yet aligning them through reinforcement learning (RL) remains challenging due to heterogeneous medical supervision. Vanilla Group Relative Policy Optimization (GRPO) assigns uniform credit across the entire generation, leading to segment interference, token dilution, and rationale-diagnosis decoupling, which exacerbates clinical hallucinations. We propose HERO (Hierarchical rationalE Reasoning Optimization), a factorized policy optimization framework that aligns heterogeneous supervision with three optimization granularities. HERO separately optimizes reasoning, diagnosis, and rationale faithfulness through complementary segment-, token-, and completion-level optimization with a heterogeneous reward formulation covering diagnostic accuracy, reasoning quality, and think-answer consistency. Experiments on MIMIC-CXR and IU-Xray show that HERO outperforms strong supervised and reinforcement learning baselines, achieving state-of-the-art clinical efficacy while producing more rationale-supported and think-answer-consistent reports, thereby substantially mitigating clinical hallucinations.
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2601.03321 [cs.LG] |
| (or arXiv:2601.03321v5 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2601.03321 arXiv-issued DOI via DataCite |
Submission history
From: Kun Zhao [view email]
[v1]
Tue, 6 Jan 2026 14:17:44 UTC (2,729 KB)
[v2]
Mon, 12 Jan 2026 05:56:31 UTC (2,729 KB)
[v3]
Mon, 3 Aug 2026 20:54:04 UTC (2,211 KB)
[v4]
Wed, 5 Aug 2026 01:39:10 UTC (2,211 KB)
[v5]
Mon, 12 Oct 2026 19:02:50 UTC (1,536 KB)
来源：arXiv：cs.AI · arxiv.org