A chest X-ray agent stops calling tools by default: it first asks whether evidence is needed, then blocks unusable measurements
Synopsis
The chest X-ray agent X-Ragent splits "should I seek evidence" from "is this measurement usable": the policy invokes region zoom, anatomical parsing and knowledge lookup only when needed, while quantitative measurements such as the cardiothoracic ratio enter reasoning only through an independent validity gate. Training runs in two stages, agentic supervised fine-tuning on 7,715 structured diagnostic trajectories followed by GRPO with an accuracy-gated reward that applies process terms only to correct trajectories. On the fully held-out ChestAgentBench, accuracy rises from 63.1% for the strongest baseline to 71.7%, with gains across all seven interpretation capabilities.
Interpretation
Chest X-ray agents now separate whether evidence should be sought from whether a retrieved measurement may be used, instead of feeding every tool output into reasoning. Earlier medical agents could invoke external modules but lacked an explicit mechanism for learning when auxiliary evidence is needed, and could not keep inapplicable quantitative evidence out of diagnostic reasoning. On ChestAgentBench, entirely excluded from training, X-Ragent reaches 71.7%, 8.6 points above the strongest baseline MedRAX at 63.1%, and exceeds that baseline on all seven interpretation capabilities.
The policy first performs a global radiographic assessment, then decides whether to invoke optional visual-refinement or knowledge-lookup tools, while quantitative measurement follows a separate channel gated by acquisition conditions. Tool calls are generated by the policy, whereas cardiothoracic ratio, lung-field asymmetry and thoracic shift are computed by a deterministic backend and exposed only when view_check confirms applicability. The Rad-CoT corpus holds 7,715 samples, 3,715 routed two-turn trajectories and 4,000 report-style diagnostic CoT examples; GRPO optimizes on 4,715 training examples for 300 steps with at most 5 policy-routed calls per trajectory.
The reward multiplies process terms by answer correctness, so measurement consistency and tool cost only refine preferences among correct trajectories. Without that gate, an incorrect trajectory could still earn reward from superficially desirable process behavior, letting process quality compensate for diagnostic error. Ablations show the tool-cost term and the evidence-consistency term each beat the baseline alone, yet combining them without the gate underperforms either alone, while gating both gives the best results on all three benchmarks.
The validity gate yields a net gain on non-standard radiographs while leaving standard frontal films unchanged. With the same measurement backend, gating raises accuracy on 613 non-frontal or atypical CXRs from 69.17% to 70.15%, while 1,887 frontal CXRs stay at 72.18%. Paired prediction changes show 9 corrections and 3 regressions on non-frontal or atypical inputs versus 8 corrections and 8 regressions on frontal inputs, so all six net corrections come from the validity-sensitive subset.
Perspective
The report covers chest X-ray interpretation: accuracy on three benchmarks, ablations of the training pipeline and reward design, and a controlled comparison of the validity gate on standard and non-standard views. The intended reader is a researcher working on medical multimodal agents and tool-augmented reasoning. X-Ragent is explicitly described as a research prototype, not a medical device, and should not be used for clinical decision-making.
The authors position X-Ragent as a research prototype and state it should not be used for clinical decision-making. The measurement channel covers only three deterministic quantities, cardiothoracic ratio, lung-field asymmetry and thoracic shift, and the validity gate withholds a substantial share of inputs: on ChestAgentBench, view_check prevents measurement execution for 24.5% of inputs, with 31% ultimately withheld after the mask-quality check. On the 115-question Rad-ReStruct subset, X-Ragent matches its base model at 70.4%.