acceptodds
Under review as a conference paper at ICLR 2027

CLINCH: Linking Evidence, Decisions, and Execution in Long-Horizon Clinical Tasks

Abstract

Clinical agents can make locally appropriate decisions yet still fail to complete long-horizon tasks in electronic health record (EHR) systems. Completing such tasks requires gathering patient evidence dispersed across records, revising decisions as that evidence changes, and distinguishing intended actions from operations actually completed, so that the final documentation reflects what was done. We introduce CLINCH, a clinical agent harness that maintains a revisable workspace linking task requirements, patient evidence, clinical decisions, and execution status. CLINCH preserves and recovers decision-relevant context, provides patient-specific medical knowledge at decision points, and checks task completion and documentation against runtime-recorded actions and results. Identified gaps return to the agent for further retrieval, execution, or revision. We evaluate CLINCH with DeepSeek-V4-Pro, GPT-5.6-Luna, and GLM-5.2 on PhysicianBench and MedAgentBench. Relative to the corresponding baseline agents, CLINCH improves task success by 13.0, 4.7, and 21.0 percentage points on PhysicianBench and by 11.1, 12.2, and 19.4 points on MedAgentBench, respectively. On PhysicianBench, it also reduces failed checkpoints in retrieval, clinical reasoning, action execution, and documentation for all three models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large language model agents, clinical agents, electronic health records, agent harnesses, long-horizon tasks, tool use, state management, execution verification

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_81uUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses disconnects among patient evidence, clinical decisions, executed actions, and final documentation in long-horizon EHR tasks. The authors propose CLINCH, a clinical agent harness that requires no fine-tuning of the main model. A Clinical Workspace maintains task, patient, and decision state. Clinically-Aware Context Management, Decision-Anchored Knowledge Integration, and Evidence-Grounded Task Verification support evidence recovery, patient-specific knowledge application, and closed-loop checks against runtime-recorded actions and results.

The experiments cover three base models, the full PhysicianBench task set, and a fixed 60-task subset of MedAgentBench. Mean single-run task success increases by 4.7–21.0 and 11.1–19.4 percentage points, respectively. The authors provide three runs per main-experiment configuration, task-level confidence intervals, a token-budget control, grouped ablations, and detailed case studies.

Critical strengths

S1. The research question is closely aligned with practical EHR workflows, and the system design is intuitive.

Distinguishing unretrieved information from negative findings and intended actions from completed operations, while checking documentation against execution, addresses important clinical-agent failure modes. The Clinical Workspace and runtime log incorporate these distinctions into ongoing execution and have a clear practical motivation.

S2. The full system improves performance across multiple models and two environments, supported by substantial analysis.

The main experiments use three runs. The authors report task-level confidence intervals, checkpoint failure categories, token usage, and a budget control. Case studies illustrate how evidence recovery, decisions, and execution connect. The selection of the MedAgentBench subset and the mechanisms enabled in that environment are also described explicitly.

Critical weaknesses

W1. The incremental value of CLINCH's specific design remains unclear. Comparisons focus on the original agents, extended execution budgets, and mechanism removal. Direct comparisons with related agents that support memory, history recovery, or review are missing, making it difficult to distinguish the benefits of standard workflow support from those of CLINCH's specific design.

W2. Statistical and mechanistic evidence for the components remains limited. Each ablation is run once with DeepSeek, while the full system uses a three-run mean; FDR and closed-loop checks are removed together. Appendix B reports ablation token usage, but finer attribution and cross-model ablations are missing. Appendix D provides cases and text excerpts, but does not show a complete input–output comparison for each mechanism.

Questions for authors

Q1. How much does CLINCH improve over other memory or review augmented agents? Please add at least one related agent baseline under the same base model, tool access, information sources, and comparable budgets, and compare task success, checkpoint outcomes, and token usage.

Q2. Are the component contributions stable, and do they vary across base models? Please repeat each ablation three times and report means, variability, and confidence intervals for paired gains. Corresponding ablations on the other models could use the same predefined subset covering the main task types. The full framework has already been tested on three models; this experiment would assess the components' cross-model stability.

Q3. Do the performance gains of each mechanism justify its additional token usage? Please present the token usage already reported in Appendix B alongside Table 4, separate FDR-call costs from subsequent revision costs, and refine the grouped ablations. Repeated runs and budget controls would help establish the marginal benefits and resource efficiency of each mechanism group.

Q4. Could the existing cases in Appendix D show more directly how each mechanism changes the information available to the model and its actions? Please provide comparisons of original tool outputs and their compressed versions, patient facts and decision records, and FDR inputs, feedback, and before-and-after text and action changes.

2 comments

We thank the reviewer for the precise and actionable questions.

Q1: Related agent baselines. We compared CLINCH with two generic agents that use the same base model, tools, and a similar token budget (details in General Response 1). Generic continuation reaches 22.0% Pass@1 and 66.9% CP, and a generic critic/reflection loop reaches 29.0% and 70.7%, compared with 34.3% and 80.6% for CLINCH. CLINCH's checkpoint pass rate exceeds the loop's by 9.9 points (95% CI [4.8, 15.2]).

Q2: Component contributions. Each ablation group targets one failure mode, and removing it raises the corresponding failure category in Table 4. Without the outline and clinical compression, retrieval failures rise from 7.7 to 14.0 per 100 tasks. Without decision-anchored retrieval, reasoning failures rise from 64.0 to 73.0. Without FDR, documentation failures rise from 37.3 to 42.0. The new no-knowledge control adds a further comparison at the component level: with knowledge cards, decision-time retrieval, and medical review all disabled, CLINCH still reaches 33.0% Pass@1 and 79.1% CP (details in General Response 2).

Q3: Token usage per mechanism. We will add token usage to Table 4 for each configuration: 2,299K for the full system, 1,679K without the outline and compression, 2,207K without decision-anchored retrieval, and 1,142K without FDR. FDR calls themselves account for 1.3% of tokens. The difference between the full system and the FDR ablation comes from the retrieval, execution, and revision work that FDR's feedback triggers.

Q4: Mechanism-level examples. We will add a section to Appendix D that traces each mechanism with verbatim excerpts from the two cases. For context management, it shows a 212K-character laboratory result stored in full, with a 20K-character visible window and an explicit marker that more content remains. For decision-anchored knowledge, it shows the query (patient context, pending decision, candidate plan), the retrieved evidence and its stated limitations, the resulting decision record with interpretation and change conditions, and the status change from planned to executed after the medication hold was confirmed by read-back. For FDR, it shows the draft's claim that the patient was counseled, FDR's finding that the operation ledger contained no matching communication, the communication the agent then created, the revised note citing that record, and the second review accepting it.

We have added the related-agent comparison requested in Q1. Our RetroSum reproduction reaches 30.0% Pass@1 and 78.4% CP with DeepSeek-V4-Pro, and 26.0% and 78.4% with GLM-5.2. CLINCH reaches 34.3% and 80.6%, and 45.7% and 84.6%, respectively. Token usage and the one-run versus three-run reporting convention are included in the General Response update.

Reviewer_QnubOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CLINCH, an new agent harness for long-horizon EHR tasks that links task requirements, patient evidence, clinical decisions, and execution status. Experiments are conducted on three LLMs (DeepSeek-V4-Pro, GPT-5.6-Luna, and GLM-5.2), and CLINCH improves both Pass@1 and CP on PhysicianBench and MedAgentBench subset compared to original baseline methods. Additional ablation studies further demonstrate the contribution of each proposed component.

Critical strengths
  • The paper studies an important and practical problem: AI agents completing clinical long-horizon tasks on real EHRs.
  • The paper proposes a new agent framework to improve the accuracy by linking the evidence retrieval, decision-making and action execution together.
Critical weaknesses
  • Additional comparison against existing clinical agent harness methods should be conducted. The related work in Section 2, such as retrospective summarization and reflection-based methods, should be directly compared in addition to the baseline agent from the original benchmark. Otherwise it is unclear how much of the gain is specific to CLINCH rather than a generic effect of adding summarization, retrieval, and a verifier loop.

  • The knowledge cards are built from clinical topics covered by the benchmark’s task instructions (Section 3.4). This design is likely to overfit a single benchmark and is hard to generalize to new tasks or out-of-domain settings. More general clinical knowledge or guidelines could be used instead, and results without the cards should be reported, since the baseline has no external knowledge access at all and part of the gain may come from this rather than from the harness design.

  • Only two benchmarks are included, and both are FHIR API-based. Only PhysicianBench matches the studied long-horizon setting while the MedAgentBench tasks are short and the documentation and knowledge-card mechanisms are disabled, so it exercises only part of the harness. It is therefore questionable how the proposed harness performs on other long-horizon benchmarks and whether the designs are optimized for these specific environments.

  • In experiments, only three LLMs are adopted for comparison. The most recent models should be also included to make the experiments more comprehensive. And it would be interesting to see and compare the improvement gains on different tiers of models.

Questions for authors

Besides comments above,

  • Why only a subset of 60 tasks (out of 300) of MedAgentBench is used for testing? And four templates are excluded?
  • What retrieval sources and search tool does the medical review use?
2 comments

We thank the reviewer for the thoughtful review.

W1: Comparison with existing methods. We compared CLINCH with a generic critic/reflection loop under a similar token budget. The loop reaches 29.0% Pass@1 and 70.7% CP, close to the baseline's 72.0% CP, while CLINCH reaches 34.3% and 80.6%; CLINCH's checkpoint pass rate exceeds the loop's by 9.9 points (95% CI [4.8, 15.2]; details in General Response 1). A reflection step that reads the agent's own trajectory tends to repeat the trajectory's claims. CLINCH keeps its working state outside the interaction history and verifies completion against what the EHR actually received.

W2: Knowledge cards. We now report results without the cards. With knowledge cards, decision-time retrieval, and medical review all disabled, CLINCH reaches 33.0% Pass@1 and 79.1% CP, compared with 21.3% and 72.0% for the baseline (details in General Response 2). Knowledge cards are also disabled in all MedAgentBench runs, where CLINCH improves over the baseline for every model. Regarding more general clinical knowledge, decision-time retrieval already draws on it: it searches public web pages for professional society guidelines and high-quality reviews, independently of the benchmark's task topics.

W3: Benchmark scope. We have extended the MedAgentBench evaluation to all 300 tasks, on which CLINCH improves Overall SR from 79.7% to 86.3%. PhysicianBench remains our primary benchmark because it scores complete long-horizon tasks across retrieval, reasoning, action, and documentation. MedAgentBench complements it by testing CLINCH in a separate EHR environment. FHIR is also the interoperability standard used by major production EHR systems, and CLINCH works with generic tool calls and their results.

W4: Model selection and model tiers. Hospital deployments usually require models that can run on premises, so that patient data stays within the institution and the model version stays fixed for auditing. For this reason, we use two strong open-weight models, DeepSeek-V4-Pro and GLM-5.2, as the primary models, and GPT-5.6-Luna as a closed-model reference. The three models differ in capability, and the gain grows with their clinical-reasoning ability: GLM improves the most (+21.0) and Luna the least (+4.7). Stronger models are therefore able to turn more of the harness's support into completed tasks.

Q1: MedAgentBench subset. The 60-task subset was fixed before any runs. It takes ten instances from each of the six templates that involve writes or conditional actions. The four excluded templates are single-step lookups: patient identification, age calculation, numerical aggregation, and latest-value retrieval. We now also report results on all 300 tasks.

Q2: Retrieval sources. Decision-time retrieval searches public web pages through the Exa search API. The queries are directed toward professional society guidelines (e.g., KDIGO, ADA, ACR) and high-quality reviews, and retrieved passages are returned with their sources. The card library is not searchable during execution. At the start of a task, the same base model selects at most two cards from the catalog based on their titles, topics, and aliases. We will describe both in Appendix C.

Following your suggestion, we directly compared CLINCH with RetroSum on all 100 PhysicianBench tasks using both DeepSeek-V4-Pro and GLM-5.2. RetroSum achieves 30.0% / 26.0% Pass@1 and 78.4% / 78.4% CP, while CLINCH achieves 34.3% / 45.7% Pass@1 and 80.6% / 84.6% CP. This provides a direct comparison with retrospective summarization beyond the generic reflection control. Details are in the General Response update.

Reviewer_13FcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces CLINCH, a clinical agent harness for long-horizon EHR tasks. It maintains a structured workspace linking task requirements, patient evidence, clinical decisions, and execution status. Three mechanisms support context preservation, decision-specific knowledge integration, and verification of actions against EHR records.

Evaluated on PhysicianBench and a subset of MedAgentBench with three base models, CLINCH improves task success by 4.7–21.0 percentage points on PhysicianBench and 11.1–19.4 on MedAgentBench.

Critical strengths

CLINCH presents a coherent and practically motivated clinical agent framework with encouraging improvements across two benchmarks. Critical strenghts are:

  • Well-motivated problem and coherent design. Explicitly linking evidence, decisions, execution, and documentation addresses important failure modes in long-horizon clinical tasks. Distinguishing missing evidence from negative findings and planned actions from completed operations is particularly valuable.

  • Meaningful empirical improvements. The evaluation spans three models and two EHR benchmarks, with repeated runs, confidence intervals, checkpoint-level analyses, and a token-budget control.

  • Useful implementation and failure analyses. The structured workspace and runtime-grounded verification are sensible design choices. The detailed case studies and component ablations help explain the observed improvements.

Critical weaknesses
  1. Incremental novelty and limited comparisons with stronger baselines (Sections 2–4, Table 2, Appendix B).

CLINCH integrates established agent mechanisms, including structured planning, context compression, retrieval augmentation, and execution verification. Its clinical specialization and integration are useful, but the incremental contribution over existing agent frameworks is not sufficiently isolated.

The experiments compare primarily against the original benchmark baselines. More informative comparisons would include a simpler planning/verification agent or a relevant clinical framework such as AgentEHR/RetroSum or ReflecTool, ideally under comparable inference budgets.

This matters because CLINCH uses 4.4–10.7× more inference tokens than the baseline on PhysicianBench (2.4–4.2× on MedAgentBench). The DeepSeek continuation control is useful but limited: it is a single run, and the generic continuation prompt reached only 22.0% Pass@1 at 2.13M tokens/task, with CP dropping from 72.0 to 66.9. This suggests the control spends extra computation ineffectively rather than establishing that CLINCH beats stronger ways of using a similar budget, such as best-of-N sampling or a generic critic/reflection loop at matched tokens.

  1. Task-informed knowledge preparation requires clearer controls (Section 3.4, Appendix C.3).

The offline knowledge cards are constructed from clinical topics covered by evaluation-task instructions. Although patient records and evaluation scores are excluded, this introduces task-informed preparation that the baseline does not receive.

Please clarify how topics are selected, which sources are available during decision-time retrieval, and whether the gains persist without the precomputed cards or when cards are prepared independently of evaluation topics.

Such a comparison would help distinguish improvements from the proposed workspace architecture from those attributable to additional task-specific medical knowledge.

  1. Component ablations do not sufficiently establish causal contributions (Section 4.4, Table 4, Appendices B and C.2).

Each ablated configuration is evaluated in a single run on DeepSeek only, while the full system is averaged over three runs. Mechanisms are also removed in groups (e.g., the Dynamic Clinical Outline together with clinical compression), making their individual contributions difficult to identify.

For example, removing Final Draft Review reduces Pass@1 from 34.3% to 32.0%, a difference within one run-level standard deviation of the full system (±2.5), while nearly halving inference-token usage (2,299K → 1,142K per task). Its failed-checkpoint profile also does not clearly match the failure mode it is meant to address: action-execution failures change little (21.0 → 23.0), and reasoning failures decrease (64.0 → 61.0). This raises a question about FDR’s incremental benefit relative to its cost.

Multi-run, more targeted ablations, ideally on more than one base model and reporting resource usage, would strengthen the analysis. The authors should also clarify the use of different context-compression thresholds across DeepSeek runs (80k/20k for the first PhysicianBench run vs. 192k/40k for the others; Appendix C.2), since this mixed configuration serves as both the headline result and the reference for all ablations, and should state which thresholds the ablated runs used.

  1. Some conclusions are stronger than the supporting evidence (Sections 4–5).

The paper states that CLINCH does not change clinical reasoning, yet decision-specific retrieval and review can correct clinical judgments, as illustrated by the reduced reasoning-checkpoint failures and Appendix D examples. The claim would be more precise if framed as not changing the base model's underlying reasoning capability.

Similarly, the proposed three failure modes are illustrated qualitatively but not measured directly. A quantitative breakdown of their prevalence, together with cases where CLINCH introduces regressions, would better support the proposed explanation of performance gains.

Finally, GPT-5.6-Luna's PhysicianBench improvement has a 95% confidence interval spanning zero, and MedAgentBench uses a selected 60-task subset with some CLINCH mechanisms disabled. These qualifications should be reflected more clearly in the conclusions.

Questions for authors

See the weakness section + additional questions below:

  • How much of CLINCH's gain remains when task-informed knowledge cards are removed or constructed independently of evaluation topics?

  • How does CLINCH compare with a simpler planning or verification baseline at a comparable inference budget?

  • Can the authors clarify the contribution and computational cost of individual mechanisms, particularly Final Draft Review, using repeated ablation runs?

  • Can the authors quantify the proposed failure modes and distinguish improvements in clinical judgments from improvements in execution consistency?

2 comments

We thank the reviewer for the detailed and constructive review.

Novelty and comparison with stronger baselines. CLINCH uses familiar building blocks such as planning, summarization, retrieval, and verification. Its contribution lies in how these are connected for long-horizon clinical tasks: patient evidence keeps its source and status, each decision is linked to the actions it requires, and completion is checked against the EHR's operation log. Adding generic counterparts of these components does not reach CLINCH's results. With a similar token budget, generic continuation reaches 22.0% Pass@1 and 66.9% CP, and a generic critic/reflection loop reaches 29.0% and 70.7%, compared with 34.3% and 80.6% for CLINCH. CLINCH's checkpoint pass rate exceeds the loop's by 9.9 points (95% CI [4.8, 15.2]), and CLINCH solves 32 tasks that the loop fails, against 4 in the other direction (details in General Response 1).

Knowledge cards. CLINCH keeps most of its gain when knowledge cards, decision-time retrieval, and medical review are all disabled: it reaches 33.0% Pass@1 and 79.1% CP, compared with 21.3% and 72.0% for the baseline (details in General Response 2). When the cards were written, the card-writing agents received only topic names and a common specification. They had no access to patient records, agent trajectories, checkpoints, or scores. During execution, decision-time retrieval searches public web pages through the Exa search API, with queries directed toward professional society guidelines and high-quality reviews; the card library itself is not searchable at that stage.

Ablations and the cost of FDR. FDR calls themselves account for 1.3% of tokens. Most of the additional usage comes from the follow-up retrieval, execution, and revision that FDR's feedback triggers. Removing FDR increases documentation failures (37.3 → 42.0 per 100 tasks) and retrieval failures (7.7 → 11.0), the two categories FDR is designed to catch. We will list token usage next to each configuration in Table 4.

Compression thresholds. The first DeepSeek CLINCH run used an earlier 80k/20k setting. The other two runs and all ablations used 192k/40k. The two 192k/40k runs average 34.5% Pass@1, in line with the three-run mean of 34.3%. We will state the setting of every run in Appendix C.2.

Clinical reasoning. We agree with the reviewer and will revise the wording. CLINCH does not change the base model's underlying reasoning capability, but decision-time retrieval and medical review can help the model correct individual judgments, as in Appendix D.1.

Clinical judgment and execution. The checkpoint categories distinguish the two. For Luna, CLINCH reduces action-execution failures from 42.3 to 22.0 and retrieval failures from 25.7 to 11.7 per 100 tasks, while clinical-reasoning failures decrease from 105.0 to 86.0. Most of the improvement for this model therefore comes from evidence gathering and execution, and the remaining errors reflect the base model's clinical judgment. A per-task comparison, which we will add to the appendix, shows the same picture: the number of tasks with more vs. fewer successful runs under CLINCH is 38 vs. 15 for DeepSeek, 30 vs. 20 for Luna, and 46 vs. 14 for GLM.

Luna and MedAgentBench scope. We will report Luna's Pass@1 confidence interval alongside its checkpoint-level gain, which is significant (+9.0 points, 95% CI [3.7, 14.4]). We also now report results on all 300 MedAgentBench tasks (General Response 4).

We have added the suggested RetroSum comparison on both DeepSeek-V4-Pro and GLM-5.2. RetroSum reaches 30.0% Pass@1 and 78.4% CP with DeepSeek, and 26.0% and 78.4% with GLM-5.2, compared with 34.3% and 80.6%, and 45.7% and 84.6%, respectively, for CLINCH. The comparison uses one reproduction run and the reported three-run CLINCH means. Full results and token usage are in the General Response update.

Comments on the paper

2 comments

We thank all reviewers for their careful reading and constructive comments. The reviewers raised four shared questions: whether the gains could be obtained with extra compute or a generic reflection loop, whether they depend on the knowledge cards, how much the individual mechanisms cost, and how CLINCH performs on the full MedAgentBench. We address each below and will incorporate these results into the revised paper.

  1. Extra compute and generic reflection do not reproduce the gains. We compared CLINCH with two generic ways of spending a similar token budget on top of the baseline agent. The first keeps the agent working with a generic continuation prompt until it reaches the budget. The second adds a critic/reflection loop, in which the model reviews its own trajectory and draft and revises them. All configurations use DeepSeek-V4-Pro on the same 100 PhysicianBench tasks.
ConfigurationRunsPass@1 (%)CP (%)Tokens/task
Baseline321.372.0524K
Baseline + generic continuation122.066.92,132K
Baseline + generic critic/reflection loop129.070.71,666K
CLINCH334.380.62,299K

Additional budget alone leaves task success at the baseline level. The critic/reflection loop also leaves the checkpoint pass rate close to the baseline, whereas CLINCH raises it by 9.9 points over the loop (paired task-bootstrap 95% CI [4.8, 15.2]). At the task level, CLINCH solves 32 tasks that the loop fails, while the loop solves 4 tasks that CLINCH fails. Reviewing the agent's own trajectory tends to repeat what the trajectory claims; CLINCH instead checks the requirements and the note against the operations the EHR actually recorded.

  1. The gains do not depend on the knowledge cards. To separate the harness from the added medical knowledge, we ran CLINCH with knowledge cards, decision-time retrieval, and medical review all disabled. In this setting, CLINCH has no knowledge source that the baseline lacks. It still reaches 33.0% Pass@1 and 79.1% CP, compared with 21.3% and 72.0% for the baseline and 34.3% and 80.6% for the full system. Knowledge cards are also disabled in all MedAgentBench runs. The improvement therefore comes mainly from the workspace, context management, and verification.

  2. Where the cost of verification comes from. Calls to Final Draft Review (FDR) account for 1.3% of CLINCH's recorded tokens. The remaining increase comes from the retrieval, execution, and revision steps that the main agent performs after FDR identifies a gap. Removing FDR raises documentation failures from 37.3 to 42.0 and retrieval failures from 7.7 to 11.0 per 100 tasks, the categories FDR is designed to check. We will add token usage for every ablated configuration to Table 4.

  3. Full MedAgentBench. We evaluated all 300 MedAgentBench tasks across the ten templates with DeepSeek-V4-Pro. Against the baseline with the same runtime and all CLINCH mechanisms disabled, CLINCH improves Overall SR from 79.7% to 86.3%.

AuthorsOpenReview ↗

pdate: comparison with RetroSum

Following the reviewers' suggestion, we reproduced RetroSum on all 100 PhysicianBench tasks with DeepSeek-V4-Pro and GLM-5.2. We used the open-source RetroSum implementation and adapted it to the PhysicianBench environment.

Base modelAgentPass@1 (%)CP (%)Tokens/task
DeepSeek-V4-ProBaseline21.372.0524K
RetroSum reproduction (1 run)30.078.41,695K
CLINCH (3-run mean)34.380.62,299K
GLM-5.2Baseline24.776.5377K
RetroSum reproduction (1 run)26.078.41,638K
CLINCH (3-run mean)45.784.62,066K

RetroSum improves over the DeepSeek baseline, showing that retrospective summarization is useful in this setting. CLINCH achieves higher task success and checkpoint completion than this reproduction with both base models. The difference is particularly clear with GLM-5.2: CLINCH exceeds RetroSum by 19.7 points in Pass@1 and 6.2 points in CP. These results suggest that maintaining patient-specific evidence and decision states and checking completion against EHR operations complements retrospective summarization. We will add the comparison, reproduction details, and token usage to the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.