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

Minye Shao; Chaohui Yu; Yixuan Wu; Fan Wang; Ling Shao; Yang Long

cs.CVSep 30, 2026

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

Authors: Minye Shao, Chaohui Yu, Yixuan Wu, Fan Wang, Ling Shao, Yang Long

Organizations: Durham University · DAMO Academy, Alibaba Group · Hupan Laboratory · University of the Chinese Academy of Sciences

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 Claude Opus 4.8. Code will be available at https://github.com/VinyehShaw/CASE.

Figures & tables

Figure 1 : The CASE longitudinal evidence-seeking benchmark. Fixed clinical targets are paired with heterogeneous patient environments. (1) Cohort, task, and raw-recording composition. (2) Clinical domains and answer formats. (3) Five organ-specific MRI protocols. (4) Paired-visit age, task density, and agent visual load. (5) Multi-organ coverage. Raw recordings precede training selection; W/H/P/B/L denote whole body, heart, pancreas, brain, and liver. Zoom in for details.

Figure 2 : CASE training and execution. (1) Two-stage training: RSFT distills screened teacher trajectories into Qwen3-VL-8B-Instruct, then agentic RL refines the policy on its own trajectories via outcome reward, OPSD, and RLAIF from an external non-ancestor judge. (2) One shared policy instantiates three evidence-seeking workflows through role prompts and tool registries. (3) A common contract alternates evidence decisions and tool observations until the P1–P5 synthesis. Privileged information scores completed training trajectories only and never determines the tool sequence.

Base model
Domain	Oncology	Endocrine/ metabolic	Cardiovascular			Respiratory/ allergy		Multisystem		Medical history	Overall (%) $\uparrow$
Method / Question	Cancer (B)	Diabetes (B)	Condition (S)	Condition set (M)	Hyper- tension (B)	Condition (S)	Condition set (M)	ICD-10 (S)	Disease systems (M)	History (B)	Micro	Macro
Qwen3-VL-8B	67.7	46.1	72.9	70.0	38.4	52.5	49.3	26.9	5.8	41.1	46.0	47.1
Medical models
Lingshu-32B	85.0	70.4	55.0	54.6	48.8	58.9	51.6	35.2	2.9	40.5	49.9	50.3
Lingshu-7B	65.4	83.1	48.6	40.3	65.8	50.7	38.6	28.5	3.3	42.8	45.5	46.7

Table 1 : Clinical accuracy (%) averaged over the three roles across the benchmark’s domains and question types, comparing medical specialist models, frontier general-purpose multimodal models, and CASE; “base” is the untrained Qwen3-VL-8B. Columns are question types (B, S, M = binary, single-select, multi-select) under six domains, with Micro and Macro at right.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Clinical area	Target	Form	Questions
Oncology	Doctor-diagnosed cancer	B	4,710
Endocrine/metabolic	Doctor-diagnosed diabetes	B	4,718
Cardiovascular	Condition identification	S	4,685
Cardiovascular	Condition set	M	4,739
Cardiovascular	Essential hypertension (I10)	B	4,037
Respiratory/allergy	Condition identification	S	5,013

Appendix

Table 3: Clinical organization of the ten task–format strata. B, S, and M denote binary, single-select, and multi-select; multi-select cardinality varies by question. Hypertension is grouped with cardiovascular medicine, while its source label remains a distinct target family.

Figure 4 : Benchmark difficulty and interaction depth. (a) First-pass tool-call depth per role over 46,613 / 48,587 / 49,827 auditable Anatomist / Radiologist / Consultant trajectories; the median is 3 calls for Anatomist (almost all $\leq 4$ ), 5 for Radiologist (90% $\leq 6$ ), and 7 for Consultant (90% $\leq 9$ ), a hierarchy of evidence-acquisition complexity. (b) Answer cardinality of the 3,637 ICD multi-select items; the number of correct options $k$ spans 1–9 and concentrates at $k\leq 4$ (965 / 899 / 702 / 511 for $k{=}1$ – $4$ , ${\approx}85\%$ ), so exact-set scoring must recover a few correct options out of nine. (c) Positive-class proportion of the four binary families, Cancer 32.1%, Diabetes 43.4%, History 62.0%, and Hypertension 21.3% over 4,710 / 4,718 / 4,037 / 4,037 items; none is near 0 or 1, so always answering the majority class cannot inflate accuracy.

Paradigm	Benchmark input	Model behavior	Evaluation target
Medical QA	Question and task-provided text or images; external knowledge access where permitted.	Infer an answer from available context, optionally retrieving external knowledge.	Primarily answer correctness.
Interactive medical tasks	Task, patient state or records, and tools or simulated encounters.	Acquire information and perform task-dependent decisions or actions.	Task-specific success; some benchmarks also assess reasoning or information acquisition.
CASE	Clinical question and a participant-specific longitudinal evidence environment.	Select tools, accumulate text and images, synthesize findings, and answer.	Exact-set accuracy; logged retrieval and completion metrics. Independent grounding assessment is an additional validation protocol.

Appendix

Table 4: Illustrative input–interaction–evaluation distinctions. The categories are not mutually exclusive, and existing interactive benchmarks need not prescribe a tool sequence.

System	Evidence use and interaction	Reported adaptation mechanism
MDAgents ( Kim et al., 2024 )	Medical problem solving with complexity-dependent individual or collaborative consultation.	Inference-time selection of roles and collaboration structure.
MedRAX ( Fallahpour et al., 2025 )	Iterative chest X-ray analysis using specialist tools.	Tool orchestration without additional agent training.
MedAgentSim ( Almansoori et al., 2025 )	Doctor–patient dialogue and requested tests or imaging results.	Retrieval of successful encounters and reflections from corrected failures, combined with multi-agent reasoning.
MedChain-Agent ( Liu et al., 2025 )	Information gathering and sequential referral, examination, diagnosis, and treatment.	Iterative agent feedback and MedCase-RAG with an expanding case database.
ClinicalAgent ( Yan et al., 2025 )	Department routing and staged consultation with laboratory and imaging reports.	Performance-informed clinician assignment and collaborative synthesis.
MedAgent-Pro ( Wang et al., 2026b )	Guideline-derived plans, patient-specific tool execution, and quantitative clinical indicators.	Hierarchical reasoning with stepwise evidence checks.

Appendix

Table 5: Workflow and context-based adaptation in medical agents. These mechanisms can change actions within or across cases without an agent-policy gradient update. The classification concerns the described agent adaptation, not the training history of every component.

System	Parameter-learning mechanism	Relation to the present study
MMedAgent ( Li et al., 2024 )	Instruction tuning of tool calls and answers across medical tasks, including composed tool use.	Establishes medical tool-policy learning through supervised examples.
MedVR ( Jiang et al., 2026b )	RL with uncertainty-guided visual exploration and rollout-consensus supervision.	Learns image inspection; correctness gates its auxiliary tool reward.
Ophiuchus ( Jiang et al., 2026a )	Tool-oriented SFT, reflection fine-tuning, and RL over interleaved visual interactions.	Establishes learning beyond demonstrations for medical visual tools.
MACRO ( Fan et al., 2026 )	Supervised initialization and GRPO encouraging discovered composite-tool use.	Combines parameter learning with experience memory and tool-set expansion.
CASE	Demonstration initialization followed by learner-directed RL with an insight-focused, detached self-distillation penalty.	Studies acquisition across longitudinal patient sources and synthesis feedback combined with outcome and rubric rewards before group normalization.

Appendix

Table 6: Learned medical tool policies. The comparison identifies differences in supervision and task structure, not evidence of comparative performance on a shared benchmark.

Configuration	Tools	Returned evidence	Call budget
Anatomist	vista3d_segment ; vista3d_compare_stations ; vista3d_compare_visits	Organ/station measurements; longitudinal changes; segmentation overlays.	4
Radiologist	list_available_images ; view_image ; compare_visits	Acquisition availability and metadata; selected or paired MRI views.	6
Consultant	All six above; list_available_reports ; get_report ; recall_skills ; write_skill	Combined patient evidence, generated report claims, and reusable textual guidance.	12

Appendix

Table 7: Executable registries. Anatomist and Radiologist each have three tools; Consultant has ten. Tool availability does not prescribe the policy’s action sequence.

Figure 5 : Token composition of a generic trajectory. Gray, fixed context (system prompt and question); blue, model-generated tokens, the only spans RL trains (tool choice, arguments, reasoning, synthesis); orange, tool-returned text (evidence, input only); pink, image placeholders for returned images, re-passed through the vision encoder.

Step	Content
Question	Based on longitudinal medical imaging and clinical records, which hepatic conditions are supported by the available evidence? Select all that apply. Options: (A) hepatic steatosis (fatty liver disease); (B) benign hepatic lesion; (C) primary liver malignancy; (D) liver cirrhosis; (E) none of the above.
System	You are a clinical consultant agent, the final arbitration over heterogeneous clinical evidence rather than any single source. You command the full ten-tool Consultant registry of table 7 . To inspect imaging you may list available acquisitions ( list_available_images ), view a selected slice or montage ( view_image ), and compare matched views across visits ( compare_visits ); for quantitative structure you may segment an organ ( vista3d_segment ) and compare volumes across stations ( vista3d_compare_stations ) or across visits ( vista3d_compare_visits ); for textual claims you may list the generated MedGemma reports ( list_available_reports ) and read one via get_report(organ, report_type) with report_type in findings, impression, or followup for liver, pancreas, or wholebody, treating these draft reports as claims to verify because they may hallucinate; and for accumulated experience you may recall reusable skills ( recall_skills ), recording new ones ( write_skill ) only in the later review stage. Required workflow: (1) call recall_skills for the implicated organ(s) first; (2) review evidence from multiple sources (images, volumes, reports), using each for what it is best at; (3) actively doubt—mark key claims [ verified ], [ suspicious ], or [ contradicted ] and verify every decision-relevant suspicious claim with a direct tool call; (4) reflect—build the causal chain and explicitly consider what could mislead you here. When done calling tools, the final message (no tool calls) must contain all five headers in order, under 2000 characters: Phase 1 Skill Recall & Review Plan; Phase 2 Multi-source Review; Phase 3 Doubt & Verification; Phase 4 Reflective Insight (the causal chain from verified evidence toward the answer, plus what could have misled and why it was rejected); Phase 5 Final Answer. Rules: maximum 12 tool calls (hard-enforced), and once about 9 are used stop calling tools and report with the evidence at hand; prefer high-value calls ( recall_skills first, vista3d_compare_visits for change questions, get_report for report evidence, view_image only for targeted verification) and skip sources irrelevant to this question; never copy a report conclusion without verification when a quantitative tool can check it; the Answer: line in Phase 5 is mandatory and the answer is extracted only from it; do not call write_skill at this stage (skill recording happens in a later review stage).
User (patient info)	Patient ID: 2306330. This female participant was born in 1955 and was 64 years of age at the time of the most recent assessment (Instance 3). Longitudinal follow-up data were available from two assessment visits: baseline assessment (Instance 2) was conducted on 2017-08-16 at age 61 years, and follow-up assessment (Instance 3) on 2019-12-16 at age 64 years, a follow-up interval of 2.3 years. Regarding housing tenure, the participant transitioned from own with a mortgage (Instance 0) to own outright (Instance 3), indicating a change in housing stability, and resided in a house or bungalow. Household composition at Instance 3 comprised 2 resident(s); the household reported ownership of three vehicle(s), a change from two at Instance 2. The reported average total household income before tax at Instance 3 was less than 18,000, compared to 52,000–100,000 at Instance 2, suggesting a potential decrease in household economic resources. Regarding occupational characteristics, the participant reported commuting distance of nan units from home to workplace (changed from 13 units at Instance 2) and employment duration of nan units in their primary occupation. This comprehensive sociodemographic profile provides essential context for interpreting clinical and imaging findings, as socioeconomic factors are well-established determinants of health outcomes and disease risk in population-based cohort studies.
P1 Plan	The case contains liver abnormalities including a reported focal lesion. I begin by recalling liver-specific skills, then select the sources best suited to each sub-question. The main clinical challenge is distinguishing an incidental or benign hepatic lesion from primary liver malignancy. I will not directly convert an imaging abnormality into a cancer diagnosis. Instead, I will characterize lesion morphology, evaluate longitudinal stability, quantify liver structural changes, and verify whether clinical evidence supports malignancy.
Tool call 1	recall_skills(organ=liver; keywords=hepatic lesion characterization, hepatocellular carcinoma, hepatic steatosis, malignancy verification) $\to$ Liver lesion assessment requires correlation between morphology, interval change, and clinical confirmation. Stable lesions without invasive characteristics are less suggestive of malignancy. Imaging suspicion alone should not be interpreted as confirmed cancer. call_6b57153c6e84494a94af5fa8
Tool call 2	list_available_images() $\to$ Available imaging (Visit 1/Visit 2 pairs): liver—IDEAL multiecho and IDEAL temporal; whole body—Dixon BH 17s W (stations 4/8/12/20/24); pancreas—gradient-echo multiecho. call_26553c9a231a453ea5b4c90

Appendix

Table 8: A complete Consultant Agent trajectory on a multi-select hepatic-diagnosis question for participant 2306330, rendered in the recorded trajectory format. The agent recalls a liver skill, cross-checks fat-sensitive and high-resolution imaging, longitudinal volumetry, and draft reports, verifies or rejects each candidate condition in P3, and commits to steatosis plus a benign lesion rather than malignancy.

Setting	RSFT	Agentic RL defaults
Training data and epochs	Mixed three-role; one epoch	Mixed three-role; one epoch
Learning rate	$3\times 10^{-6}$	$2\times 10^{-6}$
Batch	$8$ devices $\times 2$ samples $\times 2$ accumulation	$16$ prompts $\times 8$ rollouts
Sequence budget	8,192 tokens	4,096 prompt; 6,144 continuation; 12,288 total
Assistant-turn budget	Recorded conversation	10
Image handling	Resize parameter 512	Backend-specific cached images

Appendix

Table 9: RSFT and Agentic RL launcher defaults. Run-specific overrides and initialization checkpoints are recorded separately.

Figure 6 : Weight lifecycle within one RL training step. The weights

\bm{\theta}_{t}

produced by the previous step are the step’s only active policy. vLLM samples rollouts from

\bm{\theta}_{t}

; the two OPSD scoring forwards (ordinary and answer-privileged) run on the same

\bm{\theta}_{t}

; the optimizer computes

\bm{\theta}_{t+1}

from the collected batch and syncs it back to vLLM for the next step. The retired weights

\bm{\theta}_{t-1}

leave their stored sampling log-probabilities ( old_log_probs ) as the denominator of the clipped ratio, while the reference policy

\pi_{\mathrm{ref}}=\bm{\theta}_{0}

stays frozen and serves only the

\beta\,\mathcal{K}_{\mathrm{ref}}

penalty. Hyperparameters in table 9 .

Variant

Phase-4 synthesis for participant 1588571

GRPO + RLAIF (no OPSD)

The pancreatic imaging findings offer only indirect evidence relevant to diabetes, including fat deposition, atrophy, or secondary complications. However, these findings are not diagnostic of diabetes, these findings are not diagnostic of diabetes, these findings are not diagnostic of diabetes…

CASE-8B (full)

The current MRI findings provide no direct basis for diagnosing diabetes, although pancreatic fat, atrophy, or associated complications may offer supporting context. In this case, the reports describe preserved anatomy and improving pancreatic fat content, without diabetes-related complications or a documented diagnosis in the clinical history. Therefore, the available evidence does not support selecting the diabetes option.

Appendix

Table 10 : Repetition behavior on participant 1588571 with and without OPSD. Without the privileged brake the synthesis degenerates into verbatim restatement and never commits; the full model states the supporting context, weighs the draft reports, and derives the answer strictly from its insight.

Method	Teacher / privileged information	How feedback enters learning
GKD ( Agarwal et al., 2024 )	External teacher on student-generated sequences	Distribution matching on learner-visited prefixes; compatible with joint RL fine-tuning.
RLAD ( Zhang et al., 2026c )	External teacher and old-student mixture	Mixture-anchored policy ratios make imitation selective according to rollout advantage.
TGPO ( Liu et al., 2026 )	External teacher at student prefixes	Cross entropy on teacher-preferred next tokens supplements trajectory-level verifiable-reward optimization.
OPDVR ( Lin et al., 2026 )	External teacher and outcome verifier	Outcome correctness gates the direction of sampled-token distillation rewards within a policy-gradient objective.
TRACE ( Wang et al., 2026a )	Synchronized self-teacher with coarse diagnostic labels	Differentiable student distillation is routed to annotated reasoning spans; other positions retain GRPO. The default emphasizes forward KL on key spans of correct rollouts.
SSOPD ( Tan and Hong, 2026 )	Self-generated correct completion	A forward-KL auxiliary loss supervises early prefixes of an unsuccessful completion; GRPO remains the outcome objective.

Appendix

Table 11 : Mechanism comparison, not an empirical ranking. The formulations correspond to the versions cited. Direct distillation differentiates through student probabilities; privileged replay scores can instead be held fixed while optimizing the rollout policy.

Criterion	Allocation	Assessment
Evidence identification	$[0,0.3]$	Specific findings rather than generic abnormality claims.
Logical reasoning	$[0,0.3]$	A coherent connection between findings, interpretation, and answer.
Answer support	$[0,0.2]$	Evidence that supports the selected conclusion.
Completeness	$[0,0.2]$	The intermediate steps needed to justify the conclusion.

Appendix

Table 12: Fixed rubric for batch insight scoring. Group-relative selection changes the linkage adjustment, not these scoring criteria.

Panel of Figure 8 — Figure 8 : Judge-source robustness of the training-time rubric signal under three judge APIs (Kimi K2.6, blue solid; Qwen3.7-Plus, red dashed; DeepSeek V4-Flash, teal dashed). (a) Causal-KL decreases steadily for every judge; (b) the judged insight score rises for all three despite different absolute strictness, showing a consistent upward judged-score trend across the three tested sources.

Domain	Oncology	Endocrine/ metabolic	Cardiovascular			Respiratory/ allergy		Multisystem		Medical history	Overall (%) $\uparrow$
Method / Question	Cancer (B)	Diabetes (B)	Condition (S)	Condition set (M)	Hyper- tension (B)	Condition (S)	Condition set (M)	ICD-10 (S)	Disease systems (M)	History (B)	Micro	Macro
(a) Anatomist Agent
Qwen3-VL-8B-Instruct	67.6	27.6	78.7	69.9	36.5	43.4	38.5	8.8	5.9	45.3	39.5	42.2
Lingshu-32B	89.4	82.9	98.5	97.8	72.4	65.5	68.7	17.7	0.6	28.4	59.2	62.2
Lingshu-7B	88.0	95.2	54.3	35.2	78.0	54.5	35.4	13.6	2.8	44.6	47.1	50.2
MedGemma-27B	41.2	25.5	19.2	46.9	23.7	13.0	14.6	7.9	2.5	27.9	21.0	22.2

Appendix

Table 13 : Per-role clinical accuracy: the full breakdown of the role-averaged table 1 . Within each panel all methods share prompts, patient evidence, execution limits, and parsing rules; models follow the base, medical, and general grouping of table 1 , followed by CASE. Columns are the ten task–format strata (B, S, M) under six clinical domains, with Micro and Macro as in eq. 2 . Missing or failed episodes count as incorrect.

(a) Anatomist Agent
Method	Micro (%) $\pm$ 95% CI	Full $-$ method (pp) $\pm$ 95% CI	Relative gain (%) $\pm$ 95% CI
CASE (full)	77.5 $\pm$ 4.3	—	—
CASE (RSFT only)	64.9 $\pm$ 3.9	12.6 $\pm$ 1.8	19.4 $\pm$ 3.8
Qwen3-VL-8B-Instruct	39.5 $\pm$ 3.4	38.0 $\pm$ 2.3	96.2 $\pm$ 13.5
Qwen3.8-Max	65.8 $\pm$ 4.0	11.7 $\pm$ 2.1	17.8 $\pm$ 3.8
GPT-5.4-0305	63.6 $\pm$ 4.0	13.9 $\pm$ 2.0	21.9 $\pm$ 3.7

Appendix

Table 14 : Role-specific paired gains of CASE (full) over each comparator: Micro accuracy, the difference in percentage points (pp), and the relative gain, each computed from the corresponding role manifest. Every cell is the point estimate

\pm

the half-width of the 95% paired participant-cluster bootstrap percentile interval (differences and relative gains recomputed within replicates from the same panel manifest); “—” marks the reference policy, and the pp and relative-gain columns are CASE-full minus the comparator, so positive values favor CASE.

Training variant	Micro (%) $\uparrow$	Insight quality $\uparrow$	Judge coverage (%)
(a) Anatomist Agent Shared judge coverage: 84.1%
RSFT only	64.9	0.45	94.2
GRPO (correctness + format)	73.0	0.48	90.6
GRPO + RLAIF	74.3	0.65	85.3
GRPO + OPSD	77.1	0.59	93.1
CASE-8B	77.5	0.70	95.0

Appendix

Table 15 : Role-specific ablations underlying fig. 3 . Micro uses the complete role-specific manifest. Insight quality is the fixed judge’s score in

[0,1]

of the logical chain from the Phase-4 key insight to the Phase-5 answer, averaged over the questions judged for all variants (the shared subset). Judge coverage is each variant’s judge-scored fraction and shared coverage is their intersection. The last Consultant row adds inference-time skill-memory retrieval (Appendix B ), a Consultant-only augmentation rather than a training variant.

Evidence condition	RSFT micro	Full micro	Full macro	Full drop (pp)	Full calls
Question only	35.8	38.6	36.9	37.5	0
Initial clinical context	48.4	52.7	50.9	23.4	0
Fixed evidence packet	62.2	66.0	64.4	10.1	0
Adaptive acquisition, all sources	65.7	76.1	75.1	—	5.3
Adaptive, without MRI views	60.9	65.9	65.1	10.2	4.7
Adaptive, earlier visit only	59.8	64.6	62.9	11.5	3.9

Appendix

Table 16 : Evidence sensitivity under a fixed test manifest. Accuracies are percentages; drop is full-policy micro accuracy with all sources minus that in the stated condition, in percentage points. Zero calls is an imposed no-tool condition, not a measured result. Packet construction is accounted for separately from online calls.

Policy	Calls	Image blocks	Visual positions	Assistant tokens	Latency (s) median / p95	Failure (%)
CASE-8B (full)	5.1	8.4	3612	1452	34.2 / 68.5	2.6
CASE-8B (RSFT only)	4.6	7.2	3087	1248	31.5 / 61.2	6.8
Qwen3-VL-8B-Instruct	2.9	3.6	1514	782	22.4 / 49.1	21.4
Qwen3.8-Max	5.3	8.6	3704	1603	43.6 / 86.2	4.9
GPT-5.4-0305	4.3	7.0	3012	1324	46.1 / 92.4	5.7
Claude Opus 4.8	4.1	6.8	2896	1281	44.3 / 89.0	5.2

Appendix

Table 17 : First-pass interaction cost per scheduled question on the same test bank as table 1 . Failure is

100(1-N^{-1}\sum_{i}V_{i})

(unparseable or tool-error terminations), not the clinical error rate; resource means include work from failed runs and distinguish unavailable instrumentation from zero use. Read with table 1 : the full policy reaches the best accuracy at the lowest failure rate and lower latency than the frontier APIs, so autonomous acquisition is paid for in tool calls and images rather than brute-force length.

Figure 9 : The top- $128$ logits carry ${\approx}98\%$ of next-token mass in Qwen3-VL-8B-Instruct. (a) Mean sorted next-token distribution over $768$ rollout positions ( $12$ prompts $\times$ $64$ tokens), log–log: ranks $\leq 128$ (blue) hold $97.9\%$ , ranks $129$ – $2000$ add $2.1\%$ , and ranks $>2000$ add $<0.4\%$ . (b) Cumulative coverage $C(k)$ saturates by $k{=}128$ ( $C(128){=}97.9\%$ ; $C(64){=}96.9\%$ , $C(1024){=}99.4\%$ ), so $K{=}128$ buckets suffice for the causal-KL term.

Forward	Context	Parameters	Role in learning
Ordinary scoring	Student history	Current snapshot, fixed	Produces $q$ ; no scoring gradient.
Privileged scoring	Same history plus $z_{j}$	Same snapshot, fixed	Produces $p$ ; no scoring gradient.
Actor	Ordinary history	Trainable policy	Optimizes sampled assistant actions.
Reference	Ordinary history	Frozen SFT policy	Supplies the separate reference penalty.

Appendix

Table 18: Distinct forward roles in the specified algorithm. The external text judge is neither the OPSD teacher nor the reference policy.

Explore similar work

May 19, 2026cs.CL

ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning

Large language models (LLMs) and agentic systems have shown promise for clinical decision support, but existing works largely assume that evidence has already been curated and handed to the model. Real-world clinical workflows instead require agents to actively seek, iteratively plan, and synthesize multimodal evidence from heterogeneous sources. In this paper, we introduce ClinSeekAgent, an automated agentic framework for dynamic multimodal evidence seeking that shifts the paradigm from passive evidence consumption to active evidence acquisition. Given only a clinical query and access to raw data sources, ClinSeekAgent gathers evidence by querying medical knowledge bases, navigating raw EHRs, and invoking medical imaging tools; refines its hypotheses as new information emerges; and integrates the collected evidence into grounded clinical decisions. ClinSeekAgent serves both as an inference-time agent for frontier LLMs and as a training-time pipeline for distilling high-quality agent trajectories into compact open-source models. To validate its inference-time effectiveness, we construct ClinSeek-Bench, which pairs Curated Input reasoning from fixed pre-selected evidence with Automated Evidence-Seeking over raw clinical data. On text-only EHR tasks, ClinSeekAgent improves Claude Opus 4.6 from 60.0 to 63.2 overall F1 and MiniMax M2.5 from 43.1 to 47.3, with positive risk-prediction gains in 7 out of 9 evaluated host models. On multimodal tasks, ClinSeekAgent improves Claude Opus 4.6 from 47.5 to 62.6 (+15.1); all evaluated models improve across the three CXR-related task groups. We further validate ClinSeekAgent as a training pipeline by distilling agentic evidence-seeking trajectories into ClinSeek-35B-A3B, which achieves 34.0 average F1 on existing AgentEHR-Bench, improving over its Qwen3.5-35B-A3B baseline by +11.9 points and approaching Claude Opus 4.6.

Juncheng Wu, Letian Zhang, Yuhan Wang +5

UC Santa Cruz

Real-World Clinical Workflows

Ask PDF

Jul 10, 2026cs.AI

LongMedBench: Benchmarking Medical Agents for Long-Horizon Clinical Decision-Making

In this work, we introduce LongMedBench, a real-world EHR-based benchmark for long-horizon clinical decision-making. Prior evaluations of LLM-based medical agents have largely emphasized short-context knowledge QA and tool use. However, real-world medical care is inherently longitudinal, and clinicians must aggregate evidence across repeated visits, tests, and evolving treatments. Therefore, long-horizon interaction is essential for realistic assessment. LongMedBench is constructed via a reproducible pipeline that integrates MIMIC-IV admission records and clinical notes into time-series event streams and long-context memory datasets, enabling long-horizon, multi-session interactions between agents and a clinical environment. It comprises 335 patients, with 19.72 inpatient visits per patient on average and 44.91 medical events per visit. Guided by the long-horizon decision process, we propose an evaluation taxonomy with three suites: fact-based QA, temporal reasoning, and long-horizon decision-making. This taxonomy measures how agents understand and leverage historical patient information over extended horizons. Our experiments show that while recent LLMs can make good use of explicit timestamps, they have challenges in implicit time inference; The RAG and agent memory system can improve the performance of information retrieval tasks, but the performance of decision-making tasks is highly dependent on the model's immediate context.

Zihan Xu, Yanzhen Chen, Xiaocheng Zhang +4

Zhejiang University, Hangzhou, China · Alibaba Group, Hangzhou, China · Transvascular Implantation Devices Research Institute, Hangzhou, China

Clinical Reasoning Training Longitudinal Patient Records Long-Horizon Agents Long-Horizon Task Planning Mimic-Iv Datasets

Ask PDF

May 14, 2026cs.CL

COTCAgent: Preventive Consultation via Probabilistic Chain-of-Thought Completion

As large language models empower healthcare, intelligent clinical decision support has developed rapidly. Longitudinal electronic health records (EHR) provide essential temporal evidence for accurate clinical diagnosis and analysis. However, current large language models have critical flaws in longitudinal EHR reasoning. First, lacking fine-grained statistical reasoning, they often hallucinate clinical trends and metrics when quantitative evidence is textually implied, biasing diagnostic inference. Second, non-uniform time series and scarce labels in longitudinal EHR hinder models from capturing long-range temporal dependencies, limiting reliable clinical reasoning. To address the above limitations, this work presents the Probabilistic Chain-of-Thought Completion Agent (COTCAgent), a hierarchical reasoning framework for longitudinal electronic health records. It consists of three core modules. The Temporal-Statistics Adapter (TSA) converts analytical plans into executable code for standardized trend output. The Chain-of-Thought Completion (COTC) layer leverages a symptom-trend-disease knowledge base with weighted scoring to evaluate disease risk, while the bounded completion module acquires structured evidence through standardized inquiries and iterative scoring constraints to ensure rigorous reasoning. By decoupling statistical computation, feature matching, and language generation, the framework eliminates reliance on complex multi-modal inputs and enables efficient longitudinal record analysis with lower computational overhead. Experimental results show that COTCAgent powered by Baichuan-M2 achieves 90.47% Top-1 accuracy on the self-built dataset and 70.41% on HealthBench, outperforming existing medical agents and mainstream large language models. The code is available at https://github.com/FrankDengAI/COTCAgent/.

Zihan Deng, Xiaozhen Zhong, Chuanzhi Xu

School of Computing and Data Science, The University of Hong Kong · 2Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China · School of Computer Science, The University of Sydney

Electronic Health Records Clinical Decision Support Chain-of-Thought Reasoning

Ask PDF