SAT-RRG - LLM-Guided Self-Adaptive Training for Radiology Report Generation - Token-Level Push–Pull Optimization

Sep 14, 2026

Radiology report generators often produce fluent language while missing critical clinical details, creating local semantic conflicts or flipped findings that require stronger, targeted penalties. Cross-entropy increases ground-truth probability yet does not suppress the model’s current incorrect choice and treats all positions uniformly, so corrections are not prioritized. SAT-RRG introduces a self-adaptive optimization framework that uses a frozen LLM referee as weak supervision to localize semantic conflicts via error tags and apply push–pull updates with entropy and confidence signals. Experiments on MIMIC-CXR and IUXray improve language and clinical metrics and remain robust to noisy error cues.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Abstract

Introduction

Related Work

This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. SAT-RRG: LLM-Guided Self-Adaptive Training for Radiology Report Generation with Token-Level Push–Pull Optimization Yunyi Liu 1 * Yingshu Li 1 * Tong Chen 1 Lingqiao Liu3 Lei Wang2 Luping Zhou 1† 1University of Sydney 2University of Wollongong 3University of Adelaide {yunyi.liu1, [yingshu.li](yingshu.li) , tong.chen, [luping.zhou}@sydney.edu.au](luping.zhou}@sydney.edu.au)[ ](luping.zhou}@sydney.edu.au)[lingqiao.liu@adelaide.edu.au](lingqiao.liu@adelaide.edu.au) [leiw@uow.edu.au](leiw@uow.edu.au) Abstract Radiology report generators often produce fluent text yet miss crucial details, leading to local semantic conflicts or flipped findings that require stronger penalties. Crossentropy (CE) merely increases the probability of the ground-truth token y ∗ without directly suppressing the model’s current wrong choice yˆ, and treats all positions uniformly, so corrections are not prioritized. We introduce a self-adaptive optimization framework that dynamically adjusts token-level gradients based on semantic discrepancy cues derived from a frozen LLM referee. The LLM itself is not the contribution—it merely provides weak supervision to trigger the adaptive learning process. Within this framework,(i) semantic conflicts between the predicted and reference reports are automatically localized and tagged with . . . (used only during training), and (ii) adaptive, stronger penalties are applied within these sparse but critical spans. Updates follow a push–pull scheme: error spans are pushed down, while non-error tokens are reinforced. The update strength is governed by two complementary signals—normalized entropy (for uncertainty calibration) and focal-style confidence (for handling overand under-confident predictions). On MIMIC-CXR and IUXray, our framework consistently improves both language metrics (BLEU-4, ROUGE-L, METEOR) and clinical metrics (RadGraph F1, CheXbert), and remains robust to noisy or imperfect error tags. 1. Introduction Automating radiology report generation (RRG) promises to reduce radiologists’workload and improve diagnostic consistency. Recent encoder–decoder models and large language models (LLMs) [2–4, 7, 34, 35] can produce fluent narratives from images, yet they frequently miss clinically salient details: flipped findings, stray negations, or *Equal contribution.†Corresponding author. local contradictions. A central reason is that standard training with cross-entropy (CE) treats all positions uniformly and only increases the likelihood of the ground-truth token, without directly pushing down the model’s current wrong choice. Consequently, models lack a built-in mechanism for semantic self-consistency, making them unable to identify or prioritize corrections where they matter most[13] . Prior work explores token-level feedback in general NLG via reinforcement learning [43], contrastive objectives [17], or post-hoc correction. However, these approaches often depend on hand-crafted rewards, extra annotations, or auxiliary networks, which limits scalability in medical NLP where expert supervision is costly. Such reliance on coarse, sequence-level feedback leaves models insensitive to token-level semantic conflicts—a key challenge our method addresses. We propose SAT-RRG, a self-adaptive training framework that dynamically modulates token-level gradients using semantic discrepancy cues derived from a frozen LLM referee. The referee only provides weak supervision during training, highlighting phrase-level conflicts via . . . tags to yield sparse, conflict-aware token masks. Guided by these masks, SAT-RRG converts learning into a closed loop that concentrates updates on suspect spans to correct what is likely wrong while consolidating stable evidence el

FAQs

Why does cross-entropy training underperform for radiology report generation?

Cross-entropy boosts the ground-truth token probability but does not explicitly suppress the model’s current wrong choice, and it applies penalties uniformly across positions.

How does SAT-RRG obtain and use semantic discrepancy cues?

A frozen LLM referee provides weak supervision during training, highlighting phrase-level conflicts using ... tags to create sparse, conflict-aware token masks.

What is the push–pull optimization scheme in SAT-RRG?

Inside error spans the loss drives down the probability of the chosen token, while outside errors it reinforces correct tokens; update strength is governed by normalized entropy uncertainty and focal-style confidence.

Sep 14, 2026

Uploaded by Olivia Brown