acceptodds
Withdrawn from ICLR 2027

Evolving Cross-Modal Medical Agent Through an Interactive Evidence-Credit Environment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

16 views

Abstract

Medical agents built on vision–language models can integrate reasoning with active evidence acquisition to address complex clinical tasks. However, current medical agents have limited evidence–feedback for long-horizon evidence acquisition. In this paper, we introduce **RAD-Gym**, a unified evidence-credit interactive environment designed to support human-like evidence seeking and medical agent evolution through evidence–credit interaction. We define a unified action space enabling agents to locate sparsely distributed findings and ground their reasoning in evidence acquired from highly redundant medical images. Based on this action space, we further train agents on teacher trajectories augmented with synthesized pre-action evidence acquisition thinking and use another teacher model to filter trajectories to obtain reasonable credit for evidential diagnosis. We then propose Evidence-Credit Interaction Reinforcement Learning (ECI-RL) to improve long-horizon evidence seeking through evidence–credit interaction. The environment uses LLM-based expert judges to assess accumulated evidence and assign credit to newly verified diagnostic progress, providing credit feedback for policy optimization. Experiments across four heterogeneous multimodal benchmarks show that our training recipes improve radiology reporting performance over the base model, supporting successful medical agent evolution from evidence-credit interaction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Radiology Report Generation, Agent

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LQcQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces RAD-Gym, an interactive environment for volumetric medical image reporting with a shared four-action interface: locate_organ, view_slice, write, and submit. It proposes ECI-RL, a GRPO-based RL method that combines terminal report rewards with process credit assigned to newly verified diagnostic progress, tracked through a coverage ledger and LLM-based expert judges. The paper also introduces trajectory-level metrics such as EHR, EER, NPHR, SE, and F-N, and evaluates on CT-RATE, AMOS-MM, Gastric-X, and MR-RATE. The claimed contributions are a modality-agnostic interactive environment, evidence-credit RL for long-horizon evidence seeking, and new state-of-the-art results on multiple medical report benchmarks.

Critical strengths

The paper targets interactive evidence acquisition in volumetric medical imaging, where fixed-input report generation cannot capture how evidence is selected. RAD-Gym’s shared four-action interface is a reasonable abstraction for CT and MRI, and the trajectory-level metrics are a useful direction for auditable medical agents.

Process-level evaluation is a valuable motivation. Extending report evaluation to evidence coverage and exploration efficiency is clinically meaningful. If validated properly, such metrics could help distinguish agents that produce plausible reports from agents that actually acquire supporting evidence.

Critical weaknesses

My main concern is that the empirical results do not support the paper’s main claim. ECI-RL is framed as a consistent improvement, but it is worse than SFT on several reported metrics: CT-RATE CIDEr drops from 27.81 to 26.82; AMOS-MM CIDEr drops from 22.58 to 22.20 and GREEN drops from 23.81 to 21.85; Gastric-X CIDEr drops from 32.65 to 30.59 and BioBERT drops from 70.54 to 70.28. F-N is also not best on AMOS-MM or MR-RATE. The paper reports point estimates only, with no seeds, variance, or significance tests. Given these mixed results, the “consistent” and SOTA claims are too strong.

The other major issue is the evidence-credit mechanism. The process reward relies on LLM judges for G, I, and X; only T is checked programmatically. There is no human validation, no agreement study, and no judge/prompt ablation. Since the method’s novelty is to reward “verified” diagnostic progress, this is a serious gap. Reproducibility is also limited: the pipeline uses several proprietary models, and the paper provides no code, compute budget, or sufficient training details.

Questions for authors

How were the LLM judges validated against radiologists? Please report agreement, failure cases, and sensitivity to the judge model and prompt.

Why does ECI-RL improve some clinical metrics but degrade CIDEr, GREEN, and BioBERT on several datasets? Which tradeoff should readers trust?

1 comment

Response to Reviewer LQcQ

We have submitted a revised PDF version with the blue-highlighted revisions for reference.

[Weakness 1 (Concern 1) & Question 2]: Finding-level and process evaluation expose clinically important omissions that high report-metric scores can obscure, and our experiments support their complementary value. The SFT comparison shows both directions explicitly (Tables 1–2):

DatasetFinding F1: SFT → ECI-RL ↑Report metric: SFT → ECI-RL ↑
CT-RATE25.47 → 26.64CIDEr: 27.81 → 26.82
AMOS-MM13.80 → 15.05GREEN: 23.81 → 21.85
Gastric-X4.65 → 7.89CIDEr: 32.65 → 30.59
MR-RATE26.88 → 27.51F-N: 24.27 → 24.47

Recent work likewise extracts and matches structured findings (RadMatch [1]), aligns atomic finding sentences (RadSEM [2]), and scores abnormal findings with attribute-level errors (CRIMSON [3]). These support evaluating clinical units and their details. Our fixed-acquisition test additionally supports observation-aware scoring (Table 4):

MetricRanking agreement ↑Tie rate ↓
Finding F10.5210.130
NPHR × finding F10.6230.073

Here is a representative example on Gastric-X (Gastric-X 1240) that has high overlap scores despite missing the primary gastric lesion. CIDEr and RadGraph-XL reach the 97th and 95th cohort percentiles, yet only 1/6 reference findings are matched, and EER is 0.

Reference (excerpts): “Ulcerative infiltrative lesion in the gastric body, consideration of gastric carcinoma not invading the serosal layer [...]” “The gastric wall at the lesion shows abnormal enhancement in the arterial phase.”

Model (excerpts): “No acute abnormality identified in the abdomen or pelvis on this multiphase CT.” “Moderately distended stomach with fluid/ingested material [...]”

[Weakness 2 (Concern 2)]: ECI-RL rewards intermediate evidence acquisition with supporting ablations, and the revised appendices specify training and evaluation details. The ablations show (Table 3):

RewardFinding F1 ↑EER ↑F-N ↑
Coarse report only17.397.7015.88
Joint terminal only17.148.1415.28
Full ECI-RL19.7810.5017.79

The revised appendices provide training settings, prompt templates and evaluator contracts (Appendices B.4, C and D).

[Question 1]: The reviewed radiologist evaluation achieves 73.5% agreement with the judge, and our failure case highlights the need to reserve budget for targeted verification. Several radiologists assessed whether the acquired images support each finding across 20 independent cases. After review, 147/200 valid judgments agree with the judge, with one skipped item (EER: 73.6%, NPHR: 73.1%).

Gastric-X 1240: broad exploration uses 30/32 images, leaving only two slices for the final arterial review. The report then omits the reference gastric lesion.

References

[1] Corbière et al. RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching. arXiv:2609.01470, 2026.

[2] Yang et al. RadSEM: A Finding-by-Finding Metric for Clinical Consistency in Radiology Reports. arXiv:2606.17062, 2026.

[3] Baharoon et al. CRIMSON: A Clinically-Grounded LLM-Based Metric for Generative Radiology Report Evaluation. MICCAI, 2026.

Reviewer_PZiuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The training pipeline first applies teacher-guided SFT on tool-use trajectories with synthesized pre-action reasoning, and then applies Evidence-Credit Interaction Reinforcement Learning (ECI-RL), which uses evidence-gated terminal rewards and temporally weighted process credit. The paper evaluates report quality and trajectory grounding across CT-RATE, AMOS-MM, Gastric-X, and MR-RATE, claiming that ECI-RL improves finding-level reporting and acquired-evidence metrics over the base model, SFT, and report-level GRPO rewards.

Critical strengths
  1. The paper addresses an important limitation of fixed-input medical VLM evaluation: final report quality alone cannot verify whether an agent inspected the relevant part of a large 3D scan before making a claim.

  2. The explicit separation among image acquisition, report drafting, final submission, and post-hoc trajectory evaluation is a useful design choice, especially because the environment records slice provenance and attempts to link report claims to actually viewed images.

3.The paper provides unusually detailed reward definitions, evaluator contracts, token masking rules, and trajectory metrics, which makes the intended evidence-credit mechanism more inspectable than many agent-RL papers.

Critical weaknesses
  1. Could the authors clarify the central claim that ECI-RL assigns credit “strictly” to verifiable diagnostic progress, given that the terminal quality score qjq_j and error penalty UU are generated by report-level judges that access frozen claim slots derived from the reference report? Although the policy does not observe the reference report, the reward still depends on reference-conditioned claim decomposition. Please explain why this should be interpreted as evidence-based process credit rather than privileged reference-based supervision.

  2. How are the frozen claim slots Ω\Omega and the core/detail tier weights (ωcore⁡,ωdetail⁡)^=(0.8,0.2)\widehat{(\omega_{\operatorname{core}},\omega_{\operatorname{detail}})}=(0.8,0.2) constructed for each case? Please report whether slots are manually annotated, extracted by an LLM, or derived from reports automatically; how their correctness is validated; and how often the slot set omits clinically important findings or includes ambiguous findings.

  3. How reliable are the LLM-based expert judges for acquisition relevance GG, imaging support II, differential justification XX, and report quality qjq_j? The paper provides evaluator contracts but does not report physician agreement, inter-judge agreement, repeated-run variance, judge-model ablations, or error rates for the certificates that directly determine RL rewards. Please provide human validation stratified by modality, dataset, positive/negative findings, and difficult cases involving subtle lesions.

Questions for authors
  1. How is the comparison between adaptive agents and fixed-input baselines made fair under the image budget? The fixed-input baseline receives uniformly sampled slices, whereas RAD-Gym agents can use up to 20 rounds, six slices per view, and 32 images per trajectory, plus localization metadata and iterative report revision. Please report results for fixed-input baselines with matched image counts, matched token budgets, matched image resolutions, matched localization information, and an oracle uniform/stratified sampling baseline.

  2. Could the authors provide stronger interactive baselines trained under the same environment and supervision budget? The main comparisons include zero-shot medical VLMs, general VLMs, and GRPO variants, but do not clearly isolate the contribution of the proposed credit design from the benefit of training Qwen3.5-4B on thousands of GPT-5.5 trajectories. At minimum, please compare against behavior cloning without synthesized reasoning, SFT with randomly permuted action rationales, RL with dense action-cost rewards, and a learned process-reward model trained on the same expert judgments.

  3. The teacher-trajectory pipeline appears vulnerable to teacher-model and synthesizer-model leakage. GPT-5.5 produces action trajectories, Qwen3.8-Max reconstructs pre-action reasoning, and the final submission audit is given the ground-truth report. How do the authors ensure that filtering decisions, trajectory plausibility labels, or synthesized reasoning do not encode future diagnostic information indirectly into SFT data? Please provide blinded audits of pre-action reasoning against the information visible before each action.

1 comment

Response to Reviewer PZiu

We have submitted a revised PDF version with the blue-highlighted revisions for reference.

[Weakness 1]: References define the supervised targets, while acquired evidence determines their process credit. Correct report text alone cannot earn image-grounded credit, and repeated mentions do not accumulate new credit (Section 2.4, Appendix D). Retaining reference-conditioned terminal scoring but removing process credit lowers performance (Table 3):

RewardFinding F1 ↑EER ↑
Joint terminal only17.148.14
Full ECI-RL19.7810.50

[Weakness 2]: Reference claims are extracted automatically and frozen before rollout, and 0.8/0.2 improves recall and F1 over 0.5/0.5 in our parameter experiment. Core/detail masses are normalized across nonempty tiers (Appendix D). The step-20 models give the following results on Mix-100, comprising 50 Gastric-X cases and 50 CT-RATE cases:

Core/detail weightsNPrecision ↑Recall ↑F1 ↑Accuracy ↑
0.8/0.21000.11850.07110.08390.890
0.5/0.51000.12750.06200.08240.890

[Weakness 3]: The reviewed radiologist annotations agree with the judge on 73.6% of EER judgments and 73.1% of NPHR judgments, while reward ablations support process feedback. Several radiologists assessed image support across 20 independent CT/MRI cases, yielding agreement with the judge of 109/148 for EER and 38/52 for NPHR after review, with one item skipped. Table 3 isolates reward contributions, and Table 4 provides complementary counterfactual evidence. CABS [4] independently validates abnormality decomposition with clinicians and reports counterfactual concordance above 0.90 versus below 0.80 for RadGraph and RaTEScore. Together with our Table 4, this motivates unit-level and observation-aware assessment when conventional scores miss clinical errors [1–3]. We accordingly retain conventional report metrics as auxiliary measures rather than treating them as definitive measures of clinical correctness.

[Question 1]: ECI-RL retains an aggregate advantage under the same 32-image cap, although token budgets and localization access remain model-specific. We evaluate the same 100 cases per dataset, with ECI-RL using 19–32 images under the same cap (E3):

EvaluationModelImagesRadGraph-XL F1 ↑RaTEScore ↑Finding F1 ↑
Four datasetsHulu-Med-7B320.06970.46630.1316
Four datasetsECI-RL19–320.11060.53260.1978
CT-RATEMediX-R1-8B320.10210.51600.1641
CT-RATEECI-RL190.11350.55140.3135

This matches the image cap. Models retain their native processors and interfaces, so token budgets and localization access are not fully matched.

[Question 2]: Synthesized reasoning improves all five aggregate metrics in our matched LoRA comparison. Both arms use Qwen3.5-4B, the same 800 trajectories, one epoch and matched optimization. All 400 test cases enter the averages, with failed or missing cases scored zero (E4):

LoRA SFTRadGraph-XL F1 ↑RaTEScore ↑Finding F1 ↑LocAcc ↑AttrAcc ↑
With reasoning0.09110.48380.05740.06750.0238
Without reasoning0.08470.41280.05210.05630.0200

[Question 3]: Pre-action reasoning excludes future observations and reference reports, and reference access is confined to the terminal audit. The synthesizer receives the revealed history and current action before its tool observation is disclosed. The mask in the SFT training loss also ensures no leakage. The reference enters only the terminal audit, outside policy observations (Section 2.3, Appendix C.2).

References

[1] Corbière et al. RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching. arXiv:2609.01470, 2026.

[2] Yang et al. RadSEM: A Finding-by-Finding Metric for Clinical Consistency in Radiology Reports. arXiv:2606.17062, 2026.

[3] Baharoon et al. CRIMSON: A Clinically-Grounded LLM-Based Metric for Generative Radiology Report Evaluation. MICCAI, 2026.

[4] Lin et al. Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis. arXiv:2605.20277, 2026.

Reviewer_5XvhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RAD-Gym, an interactive environment for evidence-seeking in volumetric medical image reporting, and ECI-RL, which assigns process-level credit based on verified diagnostic progress. Experiments on four CT/MRI benchmarks show improvements over SFT and several GRPO baselines on selected report and evidence metrics.

Critical strengths
  1. The paper addresses an important problem: enabling medical agents to actively acquire relevant visual evidence during long-horizon report generation.

  2. RAD-Gym offers a unified interaction interface for CT and MRI volumes, including localization, slice viewing, report drafting, and submission.

  3. The proposed evidence-credit mechanism is conceptually reasonable and attempts to connect intermediate evidence acquisition with final report correctness.

  4. The experiments cover four heterogeneous datasets and include several reward ablations, report metrics, and trajectory-level metrics.

Critical weaknesses
  1. The main reward and evaluation depend heavily on LLM-based judges and automatically defined claim slots. There is no radiologist validation or agreement analysis, so the clinical reliability of EER/NPHR remains unclear.

  2. The effect of ECI-RL is not cleanly isolated. SFT and ECI-RL use different checkpoints and training stages, while some ablations are evaluated at different update numbers.

  3. Results are not consistently better across metrics. ECI-RL improves some F1/evidence scores but decreases CIDEr or GREEN on several datasets, and its MR-RATE F-N is below the RadGraph baseline.

  4. Reproducibility is limited because teacher trajectories and reasoning synthesis rely on proprietary models. Main results also lack multiple-seed variance or statistical significance tests.

Questions for authors
  1. What is the agreement between the automated evaluators and expert radiologists?

  2. Can the authors compare SFT and ECI-RL under the same training budget and checkpoint selection?

  3. How sensitive are results to the LLM judge, claim weights, and reward coefficients?

  4. Are teacher-generation data and test cases completely disjoint?

1 comment

Response to Reviewer 5Xvh

We have submitted a revised PDF version with the blue-highlighted revisions for reference.

[Weakness 1 & Question 1]: Our 20-case evaluation with radiologists yields 73.5% agreement with the judge after review, and counterfactual experiments support observation-aware scoring. We asked several radiologists to assess whether the acquired images support each finding across 5 independent cases per dataset (20 cases total). Of 201 support judgments, one was skipped and 147/200 agreed with the judge after review:

JudgmentAgreement
EER109/148 (73.6%)
NPHR38/52 (73.1%)

Table 4 additionally shows higher ranking agreement and fewer ties when observation support is included:

MetricRanking agreement ↑Tie rate ↓
Finding F10.5210.130
NPHR × finding F10.6230.073

CABS [4] also reports clinician validation of its abnormality decomposition and stronger counterfactual ranking consistency than conventional metrics.

[Weakness 2 & Question 2]: ECI-RL starts from the reported SFT checkpoint, and reward ablations separately test the contribution of process credit. In our revised version, Appendix B.4 now clearly states that our ECI-RL starting from the selected SFT checkpoint, ECI-RL performs 20 updates using low-rank adaptation. This identifies the RL increment after SFT. The reward ablations separately test process credit (Table 3):

RewardFinding F1 ↑EER ↑
Coarse report only17.397.70
Joint terminal only17.148.14
Full ECI-RL19.7810.50

[Weakness 3 & Question 3]: ECI-RL improves finding correctness and evidence coverage, and the 0.8/0.2 setting achieves higher recall and F1 than equal weighting on the mixed 100-case evaluation. MR-RATE illustrates this directly (Tables 1–2):

ModelFinding F1 ↑EER ↑F-N ↑
SFT26.8810.0924.27
GRPO–RadGraph27.209.8224.86
ECI-RL27.5113.2124.47

The step-20 models of different hyperparameter (0.8/0.2 & 0.5/0.5) are evaluated on a Mix-100 dataset, comprising 50 Gastric-X cases and 50 CT-RATE cases:

Core/detail weightsNPrecision ↑Recall ↑F1 ↑Accuracy ↑
0.8/0.21000.11850.07110.08390.890
0.5/0.51000.12750.06200.08240.890

High report similarity can overlook clinically important omissions, motivating finding-level evaluation in RadMatch, RadSEM and CRIMSON [1–3]. Our counterfactual study similarly gives NPHR × finding F1 higher ranking agreement than CIDEr (0.623 vs. 0.392, Table 4). We therefore regard conventional report metrics as imperfect proxies for clinical correctness and use them as auxiliary measures alongside unit-matching and observation-support metrics.

[Weakness 4 & Question 4]: Teacher generation and evaluation use separate splits, and the revision documents generation, training and checkpoint selection. The revision specifies teacher-trajectory generation, reasoning synthesis, filtering, optimization and checkpoint selection (Appendices B.4 and C). Evaluation uses temperature 0, with one reported training seed.

References

[1] Corbière et al. RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching. arXiv:2609.01470, 2026.

[2] Yang et al. RadSEM: A Finding-by-Finding Metric for Clinical Consistency in Radiology Reports. arXiv:2606.17062, 2026.

[3] Baharoon et al. CRIMSON: A Clinically-Grounded LLM-Based Metric for Generative Radiology Report Evaluation. MICCAI, 2026.

[4] Lin et al. Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis. arXiv:2605.20277, 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.