Researchers propose X-Ragent, a post-training framework for chest X-ray (CXR) reasoning that uses reinforcement learning to help an AI system seek supporting evidence when needed and accept only valid quantitative measurements. The work was published on Oct 12 and submitted to the Hugging Face Papers page by Zhenlong Yuan on Oct 14; the listed affiliation is Carnegie Mellon University. The authors say the framework addresses two limits of tool-augmented medical agents: learning when auxiliary evidence is needed and keeping inapplicable measurements out of diagnostic reasoning.

The team built Rad-CoT, a 7.7K-sample corpus combining tool-use sequences ordered by cause and effect with diagnostic reasoning written in report style. They used agentic supervised fine-tuning (training on examples of multi-step tool use) to initialize Qwen3-VL-8B, then optimized it with Group Relative Policy Optimization (a method that compares groups of model responses during training). Its accuracy-gated reward applies process-level terms for measurement consistency and tool cost only to correct trajectories, so plausible tool-use steps cannot offset an incorrect diagnosis.

Across three chest X-ray benchmarks, X-Ragent achieved the strongest results, including an 8.6-point improvement on the fully held-out ChestAgentBench. The authors say selectively seeking evidence and admitting only valid measurements supports more faithful, evidence-grounded reasoning. The source does not specify the benchmark score scale or give results for the other two benchmarks.

Researchers propose X-Ragent, a post-training framework for chest X-ray (CXR) reasoning that uses reinforcement learning to help an AI system seek supporting evidence when needed and accept only valid quantitative measurements. The work was published on Oct 12 and submitted to the Hugging Face Papers page by Zhenlong Yuan on Oct 14; the listed affiliation is Carnegie Mellon University. The authors say the framework addresses two limits of tool-augmented medical agents: learning when auxiliary evidence is needed and keeping inapplicable measurements out of diagnostic reasoning.

The team built Rad-CoT, a 7.7K-sample corpus combining tool-use sequences ordered by cause and effect with diagnostic reasoning written in report style. They used agentic supervised fine-tuning (training on examples of multi-step tool use) to initialize Qwen3-VL-8B, then optimized it with Group Relative Policy Optimization (a method that compares groups of model responses during training). Its accuracy-gated reward applies process-level terms for measurement consistency and tool cost only to correct trajectories, so plausible tool-use steps cannot offset an incorrect diagnosis.

Across three chest X-ray benchmarks, X-Ragent achieved the strongest results, including an 8.6-point improvement on the fully held-out ChestAgentBench. The authors say selectively seeking evidence and admitting only valid measurements supports more faithful, evidence-grounded reasoning. The source does not specify the benchmark score scale or give results for the other two benchmarks.