acceptodds
Withdrawn from ICLR 2027

From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

15 views

Abstract

Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal imaging. We propose CASE: a series of role-specific Clinical Agents for Seeking Evidence, together with a tool-use harness and an agentic post-training framework for compact vision-language policy models. We further introduce a longitudinal multimodal benchmark built on UK Biobank, comprising 50,401 clinical questions derived from real-world ICD-10-coded diagnoses of 4,739 participants. Each question links to a patient-specific environment containing clinical context and multi-sequence MRI from baseline and follow-up visits, where agents autonomously select which visits, organs, modalities, slices, and specialist tools to inspect and compare. Supervised fine-tuning transfers evidence-seeking workflows from 14,734 frontier-model interaction trajectories, followed by agentic reinforcement learning on the learner's own environment interactions. Privileged on-policy self-distillation and rubric-based LLM feedback refine evidence-to-conclusion reasoning without prescribing tool sequences. Experiments show that CASE moves beyond question-answer imitation toward transferable investigation policies, strengthening evidence-grounded longitudinal reasoning. Under matched evaluation conditions, our Qwen3-VL-8B based agent achieves over 16% and 10% relative improvements in answer accuracy over GPT-5.4 and Opus 4.8.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Agentic Learning, Reinforcement Learning, Multimodal Reasoning, Longitudinal Medical Reasoning, Clinical Agents

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jFAqOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CASE, an agentic framework for longitudinal medical reasoning that enables vision-language models to actively acquire and synthesize clinical evidence. The authors construct a novel benchmark based on UK Biobank, comprising 50,401 questions from 4,739 participants, with access to longitudinal MRI, anatomical measurements, and generated medical reports. The proposed training pipeline combines rejection-sampling fine-tuning (RSFT) with agentic reinforcement learning, incorporating rubric-based AI feedback (RLAIF) and on-policy privileged self-distillation (OPSD). Extensive experiments demonstrate substantial improvements over the base VLMs and several frontier multimodal models.

Critical strengths

1. Meaningful problem formulation and benchmark design: The paper addresses an important limitation of conventional medical QA by explicitly modeling evidence acquisition as part of the reasoning process. The longitudinal multimodal environment, heterogeneous evidence sources, and role-conditioned tool interactions provide a valuable setting for studying autonomous evidence-seeking policies beyond fixed-context question answering.

2. Strong empirical results with extensive evaluations: CASE achieves substantial improvements over both its RSFT initialization and standard GRPO training. The experiments include comparisons against multiple medical and general-purpose models, role-specific evaluations, and component-level ablations. In particular, the sensitivity experiments in Table 15 demonstrate the importance of adaptive information acquisition, strengthening the empirical motivation for the proposed framework.

Critical weaknesses

1. Insufficient Validation of Evidence Grounding for Clinical Conclusions

Although the paper distinguishes label agreement from evidence sufficiency (Table 4), its evaluation still relies mainly on diagnosis-label accuracy and text-only LLM judgments, which cannot verify whether acquired multimodal evidence genuinely supports the predicted diagnoses. Evidence sensitivity experiments show that retrieved information matters, but not that evidence-to-answer reasoning is clinically valid, especially for conditions not directly identifiable from imaging. The gains may therefore reflect correlated clinical cues or imperfect generated reports. An independent grounding evaluation on a representative subset would better support the paper’s central claim.

2. Limited OPSD Validation

Appendix D clarifies how OPSD differs from existing methods, and Table 2 shows its effectiveness, but the observed gains remain insufficiently attributed to answer-privileged conditioning. Appendix F proposes relevant controls (e.g., neutral hints and shuffled answers), but their quantitative results are not reported. Furthermore, the experiments do not clearly distinguish improvements in evidence acquisition from those in synthesis stability and repetition suppression. These analyses would help substantiate the proposed OPSD mechanism.

Questions for authors

Q1. Evidence-Grounding Assessment

How do the authors validate whether the acquired evidence genuinely supports the predicted diagnoses, beyond label agreement? An independent grounding assessment, as discussed in Table 4, would help substantiate the evidence-grounded reasoning claims. Quantitative results across different clinical task categories would help validate the evidence-grounding claims.

Q2. OPSD Mechanism

To better understand the effectiveness of OPSD, could the authors report the neutral-hint, shuffled-answer, and evidence-prefix swapping analyses proposed in Appendix F? These results would help clarify the role of answer-privileged conditioning and distinguish improvements in evidence acquisition from those in synthesis quality.

Q3. Training and Evaluation Consistency

The manuscript reports different trajectory counts and notes differences in Consultant tool configurations across training stages. Could the authors clarify how the final training samples were selected and how Consultant's tool access was handled across SFT and RL? Additionally, was skill memory enabled in the main results (Tables 1 and 12), or only in the separate variant reported in Table 14?

Reviewer_mJVWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CASE, a framework that trains a compact VLM (Qwen3-VL-8B-Instruct) as a tool-using clinical agent that gathers evidence from longitudinal patient environments. It contributes (i) a UK Biobank–derived benchmark of 50,401 multiple-choice questions built from diagnosis records (doctor-diagnosed fields and ICD-10 codes) of 4,739 participants, each linked to baseline/follow-up multi-organ MRI, Vista3D measurements and MedGemma-generated reports exposed through tools; (ii) three prompt-defined roles (Radiologist, Anatomist, Consultant) sharing one policy; and (iii) a two-stage recipe: judge-filtered SFT on Qwen3.8-Max trajectories, then GRPO combining an outcome reward, a Kimi K2.6 rubric reward on the Phase-4 “insight,” and a detached answer-privileged self-distillation penalty (OPSD). On a participant-disjoint test split, the 8B agent reports 76.1% micro accuracy versus 65.3% for GPT-5.4 and 68.6% for Claude Opus 4.8, supported by component ablations, evidence-removal tests and training curves.

Critical strengths
  1. A relevant setting with a potentially useful environment. Shifting from supplied to acquired evidence over paired-visit, multi-organ MRI goes beyond fixed-context medical VQA, and the environment design (participant-disjoint hash split, request-keyed replay cache so RL observations match the evaluation tools, logged acquisition paths) could become a useful testbed if released. This is the main source of value, but it is conditional on the label-validity concerns below.
  2. A concrete diagnosis of rubric-reward collapse. The finding that rubric-only GRPO degenerates into repetitive restatement late in training, and that a detached answer-privileged divergence penalty prevents it (Figs. 3, 7; Table 14), is a transferable lesson for RLAIF on agent synthesis, and Apps. D and I are candid about what the penalty does not guarantee. Its weight is limited because each variant is a single run without seeds.
Critical weaknesses

W1. The headline claims rest on an unvalidated benchmark and a mismatched comparison.

  1. Labels are diagnosis fields (e.g., any-site cancer, diabetes, family history) that the provided MRI largely cannot determine, and App. A concedes evidence sufficiency “requires separate assessment,” which is never done. Yet accuracies are implausibly high: in the Anatomist panel, with segmentation/volume tools only, CASE and Lingshu-7B both reach 95.2% on diabetes (Table 12a), although §6.2 states specialist models “cannot exploit the harness.” This suggests shortcuts via the narrative, options, or MedGemma reports (Table 8, call 8 returns a “clinical record review”); no text-only or option-only baseline rules them out.
  2. CASE is trained in-distribution (same generator, harness and labels) but compared only with zero-shot models. No method trained on the same data (correctness-filtered SFT, tool-free QA SFT, MedVR/Ophiuchus-style RL) and no external task is evaluated, so “transferable investigation policies” and the generality of the prompt-defined role workflows are unsupported.
  3. Evidence-seeking clinical agents are established (MediQ, NeurIPS 2024; AgentClinic; VivaBench; MTBBench), and App. A.2 itself calls the design “a testable design difference, not a demonstrated advantage.” Four contributions are claimed; none is validated on its own terms.

W2. Key claims contradict the paper’s own results, and the record is internally inconsistent.

  1. §5 says RSFT “only imitates demonstrations” and genuine gains come from RL, yet RSFT supplies ~20 of the 30-point gain over the base (46.0→65.7→76.1), and its Radiologist accuracy (67.9) beats the teacher’s (40.2). Rejection-sampling fine-tuning is known to rival GRPO (Xiong et al., 2025).
  2. Kimi K2.6, the SFT filter and RL judge, scores 42.8% on this task (30.3% as Consultant), below the untrained base, and judges ≤800 characters of text against the reference without images, so “insight” scores cannot reflect grounding.
  3. Inconsistencies: ~140k vs 48,464 SFT samples; LR 3e-6 (Table 9) vs 1e-5 (App. C); an SFT serializer lacking the ten-tool Consultant branch; an evaluation judge “held out” (§4) yet possibly reused (App. F); references to “blank tables” and baselines that “remain to be supplied”; and the teacher Qwen3.8-Max, available only from August 2026, cited as a 2025 report.
Questions for authors
  1. Which environment evidence determines the diabetes, cancer and history labels? Please report per-stratum accuracy of a text-only model on the narrative and of an option-only prior, explain Lingshu-7B’s 95.2% diabetes accuracy given §6.2, and state whether MedGemma report generation saw any clinical-record input. If imaging is not decisive for most strata, the benchmark and evidence-seeking claims do not hold.
  2. Is Table 8 a verbatim recorded rollout? Its system prompt cites “Table 7,” view_image returns interpretive prose rather than images, and the multi-select hepatic question matches none of the ten strata in Table 3. Likewise, is Eq. (11) measured at a single position, given that consecutive tokens of the repeated phrase appear as alternatives?
  3. How many full RL runs underlie Tables 2/14 and Fig. 8 (its causal-KL curves differ by judge, implying separate runs), and what wall-clock time did they take on the stated single 8×H100 node? When were the 147k teacher trajectories recorded relative to Qwen3.8-Max’s August 2026 availability (seeds 20260812/20260820), and are all reported tables final?
  4. Why use Kimi K2.6 as filter and reward given its 42.8% task accuracy, and which model is the held-out evaluation judge?
  5. Same-data baselines on identical splits: correctness-filtered SFT/RAFT, tool-free SFT on question–answer pairs, and one prior medical agentic-RL recipe. A margin that persists would materially raise my assessment.
  6. Evaluation on an external longitudinal dataset or task to test whether the investigation policy transfers beyond this generator and harness.
Reviewer_8PfaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces CASE, a framework and benchmark for evidence-seeking clinical agents using longitudinal patient records and imaging from UK Biobank. A shared Qwen3-VL-8B policy operates under three role configurations and is trained through filtered teacher-trajectory SFT followed by agentic RL. The RL reward combines answer correctness, rubric-based feedback, and a detached penalty derived from distributional differences between ordinary and answer-privileged contexts. The paper reports improvements over its SFT initialization and several larger models.

Critical strengths
  1. The paper introduces a large-scale benchmark comprising 50,401 questions from 4,739 participants, covering diverse disease categories, question formats, and multimodal evidence sources. It provides a useful testbed for future research on longitudinal medical agents.
  2. The paper provides extensive implementation details, including tool configurations, training procedures, and evaluation protocols, which help readers follow the implementation and facilitate reproduction.
Critical weaknesses
  1. The motivation for the three-role design is insufficiently explained. It is unclear whether the roles primarily represent different evaluation settings or are intended to improve policy learning, and what this design contributes beyond different prompts and tool configurations.

  2. The claim that OPSD reduces repetitive reasoning lacks sufficient direct evidence. Reward curves and a single-case comparison provide limited support, while I could not locate quantitative results for the Rep-8 metric defined in Appendix F.

  3. The experiments do not sufficiently distinguish whether RL improves evidence acquisition or the interpretation and synthesis of acquired evidence. Appendix F describes an evidence-prefix swapping analysis that could address this question, but I could not find its results.

  4. The writing needs substantial improvement. The main text mixes methodological ideas with implementation details and frequently redirects readers to the appendices, making the paper difficult to follow.

Questions for authors

Please see the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.