AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

REVA-PO: Stabilizing Reinforcement Learning for Chest X-ray Report Generation

Li Guo, Anas Tahir, Z. Jane Wang

ECCV 2026 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add classifier-guided SFT (Stage 2) after warm-up (ablation)REVA-PO LVLM · MIMIC-CXR testCheXpert F1: 0.407 -> 0.474Improved
add RL stage (Stage 3) after Stages 1 and 2 (ablation)REVA-PO LVLM · MIMIC-CXR testBLEU-4: 0.118 -> 0.143Improved
add warm-up (Stage 1) before Stages 2 and 3 (ablation)REVA-PO LVLM · MIMIC-CXR testCheXpert F1: 0.481 -> 0.506Improved
add Response-Weighted Regularization instead of constant KL (ablation)REVA-PO · MIMIC-CXR testCheXpert F1: 0.484 -> 0.506Improved
add KL regularization (ablation)REVA-PO · MIMIC-CXR testCheXpert F1: 0.463 -> 0.506Improved
add Validation-Anchored Policy Reset (ablation)REVA-PO · MIMIC-CXR testCheXpert F1: 0.481 -> 0.506Improved
use BLEU-4 rather than RadGraph F1 as RL reward (ablation)REVA-PO · MIMIC-CXR testCheXpert F1: 0.480 -> 0.506Improved
use CheXpert F1 as RL reward (ablation)REVA-PO · MIMIC-CXR testtraining stabilityHurt
increase group size G from 4 to 6 (ablation)RER · MIMIC-CXRfinal metricsNo clear effect
scale base LVLM from 3B to 7B (ablation)REVA-PO · MIMIC-CXR testBLEU-4: 0.135 -> 0.143Improved
scale base LVLM from 3B to 7B (ablation)REVA-PO · IU-Xray testBLEU-4: 0.195 -> 0.199No clear effect
replace constant beta with response-weighted beta (compute cost)REVA-PO · MIMIC-CXR teststep time (s): 11.80 -> 12.02Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy