acceptodds
Under review as a conference paper at ICLR 2027

KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy

Abstract

Most clinical benchmarks evaluate language models (LMs) on diagnosis using complete case descriptions. In clinical practice, however, patients present information in different ways, and clinicians must obtain relevant history and determine which examinations are needed before reaching a diagnosis. Diagnostic accuracy alone therefore cannot establish whether an agent gathered essential information or conducted an appropriate clinical assessment. Furthermore, existing benchmarks lacks professional clinicians' verification. To address this gap, we introduce KlinikeBench, a benchmark of 333 clinician-authored tasks, each providing an isolated sandbox environment with a virtual patient, clinical tools, and task-specific success criteria. More than 35 clinicians contributed to case authoring and benchmark evaluation. In an empirical study, clinicians gave simulated dialogues higher mean quality ratings than reference conversations, which is adopted from real conversation. In each task, an LM has a fixed budget of turns to communicate with the patient, ask about relevant history, request examinations, follow action constraints, and record a final diagnosis. We score these steps separately as well as together. Across 31 models and seven model families, the best-performing models (e.g., GPT-6-astra and Claude Opus 5) succeed on of tasks, even though their diagnosis accuracy reaches 90.7%. Some models benefit from talking with the patient; others diagnose well from a complete chart but perform much worse in conversation. Overall, KlinikeBench provides a testbed for evaluating the full clinical encounter and reveals a substantial gap between diagnostic accuracy and performance in interactive clinical assessment.

Primary area:
datasets and benchmarks
Keywords:
agents evaluation, clinical agent evaluation, large language models, benchmarking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wt65OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KlinikeBench presents 333 clinician-authored tasks for evaluating clinical agents beyond diagnostic accuracy, together with a platform for constructing and revising cases. I find the paper interesting. The clinical agent environment is somewhat incremental relative to existing benchmarks, but the clinician involvement and case-authoring platform are valuable contributions.

Critical strengths

I appreciate the emphasis on evaluating history-taking and clinical actions alongside the final diagnosis (i.e. process-level focus). The clinician authoring and revision process is a strength, and the dialogue-quality assessment (30 clinicians providing 4,500 ratings) represents substantial effort. The platform for developing additional cases could also make the contribution useful beyond the released benchmark.

Critical weaknesses
  1. Unnecessary or harmful actions are insufficiently addressed (L191–203). The evaluation emphasizes completing required tools and asking required questions. However, clinical decision-making also involves knowing which tests or interventions should not be performed. How does the evaluation penalize unnecessary or potentially harmful actions? Do the argument-level action constraints capture this? A measure of action appropriateness or precision would complement the current emphasis on coverage.

  2. Clinical validity and interpretation of task difficulty need clarification (L163–170; Table 2, L275–277). Beyond the authoring and revision process, are cases independently assessed by other clinicians for clarity, solvability, and the appropriateness of their requirements? Performance from two or three independent physicians, reporting diagnostic accuracy and strict pass@1, would help contextualize the low model pass rates. Please also clarify whether “accepted alternative diagnoses” means synonymous diagnostic labels or genuinely different, clinically plausible diagnoses. Some presentations may reasonably support several differential diagnoses even after investigation; how are these handled?

  3. The dialogue-quality study needs clearer justification and reporting (L307–314, L370–376; Table 3, L397–403). Why is MTS-Dialog an appropriate comparison, and what exact instructions were given to the human raters? Please report whether the observed differences are statistically significant. I also remain uncertain how strongly progressive history disclosure, emotional appropriateness, and human-likeness establish the clinical quality of the underlying tasks. These dimensions assess aspects of conversational realism, but do not directly establish that a case is clinically coherent or supports its intended diagnosis.

  4. The interaction-budget results need a clearer separation of trends and variability (L330–340). Although the point estimates are not monotonic, they appear compatible with an overall improvement as the budget increases. I appreciate the confidence intervals in Appendix Table 8; repeated rollouts and uncertainty shown in the main figure would further help distinguish a reproducible trend from noise. It would also be useful to contextualize this result against AgentClinic’s interaction-budget analysis, which found declining diagnostic accuracy beyond a certain budget.

Questions for authors

Smaller comments and suggestions.

  • Related work and novelty (L067–073; Table 1, L076–087): Please include PhysicianBench in Table 1 and discuss its process checkpoints in the main comparison. It is already discussed in Appendix E, but is particularly relevant to the motivation here. Similarly, patient-simulator sensitivity analysis (L364–368) has precedent in AgentClinic, so its novelty should be framed accordingly.
  • Table 1, L083: AgentClinic was initially released in 2024.
  • L125–127: The discussion of clinician-authored LLM benchmarks could also acknowledge prior work like SMMILE (Neurips '25 I believe).
  • Figure 2, L116–122: Show the actual ratings instead of only “Higher mean quality ratings.”
  • L241–242: As an optional extension, pass@3 or pass@5 would help characterize the benefit of additional attempts.
  • L307–308 and Figure 5, L393–394: Consistently distinguish human clinicians from LLM clinician agents. In Figure 5, explicitly identify the agent models represented by the colored curves and the judge/patient models on the horizontal axes. Showing absolute scores alongside ranks would also make the figure more informative; these scores are currently in Table 9.
  • Figure 6, p. 8; caption L429–431: The “Diagnosis” y-axis label is clipped.
  • Section 4.2, particularly L432–438: The small SFT/RL study feels peripheral to the main contribution, and the RL findings remain inconclusive. Consider moving it to the appendix to keep the main presentation focused.
7 comments

3. Additional metrics on harmful actions and precision (Added two columns in the table 7)

How does the evaluation penalize unnecessary or potentially harmful actions? Do the argument-level action constraints capture ... A measure of action appropriateness or precision would complement the current emphasis on coverage.

We added two metrics to emphasis on the coverage and harmful behaviour into the main result (table 7) in the updated manuscript, and part of the results is shown bellow. We found that the GPT-5.5/5.6/6 family is the most careful (1–14%), with GPT-6-astra almost always checking allergies first (1.0%). The Claude models differ widely within the family. Opus 5.5 is at 12.3%, while Opus 4.8 (52.6%) and Sonnet 4.6 (79.5%) often prescribe without a check. For precision, broad tool users score lower: Grok, Gemini and DeepSeek are at 64–72%. The strongest models is able to use model wisely while achiving higher tool coverage, e.g. GPT-5.6-sol at 78.5% precision with 36.6% tools, and Opus 5.5 at 76.2% with 49.8%.

Unchecked prescribing: among encounters in which the agent prescribed (40–135 per model), the share where the first prescription came before any allergy or medication check, whether by tool or by asking the patient about allergies.

Action precision: for each encounter, the share of distinct clinical tools the agent called that were required for the case, averaged over encounters. Documentation tools such as record_diagnosis are excluded.

ModelStrict pass@1DiagnosisToolsMust-askAction precisionUnchecked prescribing ↓
Claude Opus 529.4%90.7%44.1%89.9%75.2%27.5%
Claude Opus 5.528.8%90.7%49.8%84.4%76.2%12.3%
Claude Opus 4.816.8%89.2%27.9%85.5%75.8%52.6%
Claude Sonnet 513.8%87.4%24.0%85.1%73.5%38.6%
GPT-6-astra24.3%87.4%39.3%88.0%77.0%1.0%
GPT-5.6-sol23.7%86.2%36.6%90.8%78.5%9.2%
GPT-5.423.4%80.2%53.2%73.5%74.9%22.5%
GPT-4o3.9%66.4%12.9%79.5%76.7%42.5%
Gemini-3.8-flash18.9%80.8%53.8%76.5%71.6%16.5%
Gemini-3.5-flash15.6%75.7%43.8%76.6%68.7%21.1%
Grok-4.722.2%83.2%35.1%87.5%67.6%16.5%
Grok-4.621.9%81.4%36.9%89.0%66.8%10.6%
Grok-4.321.3%77.8%45.6%83.3%71.6%15.2%
DeepSeek-v4-pro8.1%29.4%13.8%87.8%67.9%19.0%
DeepSeek-v4-flash3.9%18.0%9.3%81.3%63.6%21.5%
Kimi-K2.623.4%77.2%48.7%84.7%75.5%27.3%
GLM-5.14.8%24.9%10.8%30.8%77.9%25.0%

Our annotation platform lets clinician authors mark actions as forbidden for a case, and 204 of the 333 tasks include such entries, most often admit_to_hospital, prescribe_medication, refer_to_specialist and order_imaging. When the cases were reviewed, however, clinicians did not consistently agree on which actions should be prohibited. Whether a patient with panic symptoms or superficial thrombophlebitis should never be admitted, for example, depends on local practice and on details outside the encounter. We therefore did not score forbidden actions as a gate. In the submitted version they are instead withheld from the agent's tool set (Appendix B.1), so they restrict the action space but are not measured. The task format also supports argument-level constraints, such as requiring or prohibiting a specific drug in prescribe_medication(drug_name). This design follows the structure of clinical guidelines. They state what should be done for a presentation (the recommended work-up, safety checks and disposition) far more consistently than what must never be done; prohibitions are usually patient-specific, such as contraindications or allergies, rather than properties of the disease.

AuthorsOpenReview ↗

Expanded Interaction-Budget Sensitivity and Accepted Alternative Diagnoses

4. Extended interaction-budget results against AgentClinic (Appendices D.4–D.5)

Contextualize this result against AgentClinic’s interaction-budget analysis, which found declining diagnostic accuracy beyond a certain budget.

Following your suggestion, we evaluated five models on 50 matched tasks with step caps from 4 to 60, plus an uncapped condition. The default is 32 steps, and the budget includes both dialogue and tool calls.

Model4 steps8 steps16 steps32 (default)45 steps60 stepsNo cap
GPT-6-astra00728343236
Claude Opus 5111034364038
Grok-4.710022503830
Gemini-3.8-flash01420221420
Claude Sonnet 500320262226

Strict pass@1 falls from 20–34% at 32 steps to 0–10% at 16 steps and 0–1% at 4–8 steps, with all paired-bootstrap 95% confidence intervals excluding zero for these reductions. For GPT-6-astra, halving the budget leaves diagnosis accuracy at 86% but reduces required-tool completion from 46% to 15%, showing that correct diagnosis can coexist with incomplete assessment. Larger budgets do not consistently help: Grok-4.7 rises from 22% at 32 steps to 50% at 45 steps, then falls to 30% without a cap

5. whether “accepted alternative diagnoses” means synonymous diagnostic labels or genuinely different, clinically plausible diagnoses.

In our cases, accepted alternative diagnoses includes both synonymous diagnostic labels and genuinely different. During the data annotation process, we use knowledge graph of sympotoms (L156) to add negative symptoms to disabiguiate the targeted diseases, in the cases there is no way to disguiwish the diseases from symptoms, the clinically plausible diagnoses are added as accepted alterantive diagnoses. We also provide the detailed pseudocode for how the diagnose is compared to the ground truth label in the reply to the review (XdHN) on the second question he/she proposed.

6. Clarification on the Diagnose-quality study.

Purpose of the study As we use simulated patients, it is very critical to ensure the behaviour of these simualted patients follows the real patients pattern. So we design this study to check if the simulated patient is distinguiable from the real patients. To study this, we sample real patient-clinician conversatoin from MTS-Dialog dataset. As performing such study we need to ensure the quality of the data and also the proper license for the data. the data provided by MTS-Dialog is hereby used.

Instructions Given to the clinicians in the study:

Survey format and clinician instructions

The questionnaire collects clinicians’ professional background information and presents patient–physician dialogues for evaluation. Below are the background questions, a sample dialogue, and the rating schema.

1. Clinician background questions

  1. Department and Current Professional Title
    (e.g., Attending Physician, Department of Cardiology)
  2. Which hospital or institution are you currently affiliated with?
  3. How many years of work experience do you have?

2. Instructions for dialogue evaluation

“Please read the dialogue in the figure and rate the patient's behavior. Use a 5-point scale: 1 = strongly disagree, 5 = strongly agree.”

3. Sample dialogue

Patient: I've had this pain in my belly for about a day and a half, and it's getting really bad, especially when I cough or move around. It started near my belly button but now it's mostly on the lower right side.

Physician: That pattern is concerning for appendicitis. Do you also have fever, nausea/vomiting, loss of appetite, or diarrhea/constipation, and when you press on the lower right side and let go does it worsen?

Patient: Yeah, I feel kind of feverish and sweaty, and I've been really nauseous—I threw up once and I basically haven't wanted to eat at all. I've also had some diarrhea and constipation, and yes, pressing on the lower right side and letting go makes it hurt a lot more.

Physician: This sounds very concerning for appendicitis, and because the pain is worsening with movement/coughing and you have rebound tenderness, you need urgent in-person evaluation in the ER now. Are you able to get to the emergency department right away, and do you have any severe dizziness, fainting, or trouble breathing?

Patient: I can get there right away, and I don't have any trouble breathing. I'm a little shaky and worried, but I haven't fainted or anything like that.

4. Rating schema

For each dialogue, clinicians select one score for each of the following three statements. The response options are 1, 2, 3, 4, and 5, anchored at 1 = strongly disagree and 5 = strongly agree.

Evaluation dimensionStatement presented to cliniciansResponse options
Progressive information disclosureThe patient only gradually supply the medical history when asked by the physician.1 / 2 / 3 / 4 / 5
Emotional CongruenceThe patient's mood and worries match how serious the condition is.1 / 2 / 3 / 4 / 5
Overall impressionThe patient is more like a real human patient rather than a robot.1 / 2 / 3 / 4 / 5

Each evaluated dialogue therefore receives three separate ratings, one per dimension.

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

KlinikeBench is an expert-authored benchmark for evaluating LLMs with virtual patients with incomplete information. It additionally evaluates tool use, musk-ask coverage beyond diagnosis correctness.

Critical strengths
  1. The motivation is clear: diagnostic accuracy alone can overlook incomplete clinical assessment.
  2. Clinician involvement makes the benchmark potentially valuable.
  3. They explored the KlinikeBench’s potential as a training environment for SFT and RL.
Critical weaknesses

My main concern is why each required tool and must-ask topic is considered necessary. Diagnosis accuracy is high, while strict success is low. Without more evidence, it is hard to tell whether this gap reflects clinically important omissions. Clinicians wrote these lists, but do other clinicians agree that every item is necessary? I would like the authors to address the following questions:

  1. Explain how clinicians decide which tools and topics are mandatory.
  2. Clarify whether clinically equivalent alternatives are accepted.
  3. Give the model all must-ask information and required-tool results directly. Then test whether this improves diagnosis accuracy.
Questions for authors

How were the 35 conversations chosen in section 3.5 Clinician assessment? How is diagnosis correctness determined?

5 comments

Thank you for reading through the comments. We provide the further clarification below.

Q1. whether it is a widely accepted clinical requirement

A direct response to this is we provide both pass@1 and individual gate in our benchmark, and we also provide the must-ask score column, which provide the equally weighted scores over conversations, with additional scores on diagnosis and tool use.

Q2.1 The new experiment only has 50 cases and three models.

We admitted this experiments is not on the full set, but as each tasks solves cost around 3 us dollars, e.g. this particular ablation cost around 200 dollars. We promise we will extend to larger scale for the full set in the format manuscript.

Q2.2 A more reasonable baseline would be the model in a single-turn setting with tool access, rather than intake alone.

We follow the original suggestion to setup the experiment with must-ask information and required-tool results, we will try to add this additional experiment result with your new request. But at the same time, we think there could be inconsistent in the original review.

Give the model all must-ask information and required-tool results directly

HealthBench uses weighted scores over thousands of conversations, so the comparison does not settle this question.

We cited HealthBench as an example of how the must-ask is generated in the same way as the rubrics of health bench, where clinicians write rubrics such as "advises seeking emergency care", "asks how long the symptoms have lasted", "mentions that dose depends on kidney function".

Thank you for the response. I think my concern has been misunderstood.

Again, my concern is the validity and necessity of the must-ask topics and required tools. If these items are not validated or accepted by most physicians, the failures the benchmark reports may not be real failures. Reporting the gates separately does not answer this, because it changes how the scores are aggregated, not if the items are valid.

I mentioned HealthBench’s size because scale reduces individual bias, and its weighted scores because the weights calibrate the importance of each item. KlinikeBench does neither: every item counts the same, and missing means a failure. So I don’t think it is fair to say the approach is consistent with HealthBench. I am looking for evidence that KlinikeBench reflect accepted clinical standards and not one author’s judgment: was the necessity of each item agreed upon by other clinicians? Appendix D.9 mentions that cases are reviewed by two clinicians, but it does not say whether this review covers the necessity of each required item. In Table 2, for example, tool-use coverage is the largest gap, so a natural question is whether those tools are really necessary. Additionally, in the human baseline (Appendix D.10), even the human clinicians complete the required tools on only half of the cases, although I recognize that this is an exploratory study of eight cases.

I do not think my two comments are inconsistent. My original suggestion was meant as one way to show that the items models miss are important and thereby support the validity. It did not specify the setup, which I left to the authors. For example, the information could also be supplied at the end of a multi-turn encounter. The single-turn comparison against intake alone can show that extra information helps in general, but not that the items models miss matter, because the baseline does not reflect what the model would obtain on its own. That is why I brought up a baseline of the model with tool. I do not intend to burden the authors with more experiments. What matters is that the concern is addressed, and the authors are free to use any evidence that shows the missed items are clinically necessary.

Reviewer_h4ZwOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Goal: Establish a beyond-diagnostic benchmark for evaluating LLM

Details:
The paper presents KlinikBench, a benchmark consisting of patient cases, test reports and related tools primed with a LLM-powered patient simulator. Each sample has a hidden scoring key consisting of the gold diagnosis and tools required (hallmark/must ask questions, ordering tests, etc). The LLM being tested interacts with the patient simulator in a multi turn session to identify the diagnosis. KlinikBench evaluates the accuracy of this predicted diagnosis, but also assesses the interaction taken for reaching this diagnosis (tool calls made). The paper proposes a pass@1 metric which measures whether the agent has reached the correct diagnosis, covered all must-ask questions and made correct tool calls along the way. Results indicate that many state-of-the-art models achieve good diagnostic accuracy but fail when their interactions are evaluated.

The dataset is created by doctors as follows. A doctor first chooses a target disease for which associated findings are provided from a knowledge graph. Then, the doctor authors the patient profile, persona, tool usage and must ask topics. The doctor also lists positive and negative findings for the patient. Finally, doctors can refine the created profile by observing a simulated doctor patient interaction. The paper compares the simulated dialogue to publicly available MTD Dialogue dataset. However, I could not find a discussion on the quality of the created profiles themselves.

Critical strengths
  • I heartily welcome any new resource in the medical domain. I especially appreciate KlinikBench cases that contain explicitly marked negative symptoms which are very essential for accurate diagnostic information. Recording alternative diagnosis is very important and meaningful for future studies.
  • I also like that extensive suite of models being tested and the SFT/RL baselines as well.
Critical weaknesses
  1. The main contribution of the paper is a benchmark dataset curated with the help of a team of doctors. Hence, authors should provide some qualitative measurements on the samples being created. Have authors verified that the instructions given to the doctors are being followed across the team to a good degree? Could there be any subjective interpretations causing two doctors to judge a case differently? I understand that standard agreement measures such a Cohen Kappa may not be applicable due to the complex nature of the data. But, requesting a doctor to verify a subset of cases manually could be viable and helpful.
  2. The evaluation criteria is unnecessarily strict in my opinion. Results in table 2 indicate that frontier models such as GPT 6 Astra and Opus 5.5 operate at a reasonably higher diagnostic accuracy. The overall drop in the performance is mostly due to models failing to make all the tool calls listed for the task. What motivates the authors to use these strict measures? As the authors have themselves pointed out in the introduction, there can be alternate but acceptable routes to reach a certain diagnosis. Such a route may be different from what is recommended in the task but can still be viable.
  3. What could be a human baseline in this case? I don’t think even human doctors could possibly reach a high pass@1 score due to its strict nature. If possible, authors could ask a fresh pair of doctors in the simulation and provide a human baseline over 50 samples. Their response may motivate authors to reconsider such strict measures. I also recommend considering standard precision/recall/f1 based measures. An agent that exhaustively inquires everything would naturally have a high tool and must ask coverage, but is not really useful.
  4. Authors check whether or not increasing the number of turns in session help. But, equally important is to verify whether the models break at lower turn budget, especially from an efficiency perspective.
  5. Authors motivate KlinikBench in the introduction (lines 68-72) stating that an agent may reach correct diagnosis early on by guessing or may end up with incorrect diagnosis due to interaction failures. I don’t think the results from the paper substantiate these motivations. Can authors provide example interactions from their evaluations discussing such cases. As discussed in point 2, analysing results at lower turns (say 4/8) can help understand whether the agent has reached a conclusion without enough exploration or whether it fails early on.
  6. In single turn v/s multi-turn setting, GPT-5.4 reports decrease in the accuracy when full snapshot is provided. This is counter intuitive. Can authors provide an explanation?
  7. KlinikBench uses GPT-5.4-mini for simulating patients. It is not clear what motivated this choice of the model or whether the simulator was evaluated for hallucinations or unfaithful responses. The main paper does not state or reference this information, but shows that switching patient LLM maintains model ranking to some degree.
  8. Can authors motivate comparing KlinikBench with MTS Dialogue? KlinikBench’s main contribution is patient profiles augmented with tools, must-asks and diagnosis. MTS dialogue contains doctor-patient dialogues, many of which are incomplete or short. I don’t believe this evaluation is meaningful unless authors want to showcase the quality of the dialogue (not their main contribution).
Questions for authors

Questions and Writing Issues:

I request elaborations on my points from the weakness section. Further, here I have queries related to the writing of the paper.

  1. KlinikBench has 333 tasks. It is unclear from the introduction whether it means 333 test samples/cases or each sample has a subset of 333 tasks. The paper should provide clarification early on.
  2. The introduction and section 2 does not clearly indicate what the constraints in the tasks are.
  3. What is the purpose of role-play during the creation of the tasks?
  4. Where is the conclusion? Shouldn’t we expect every paper to have one?
  5. Since the main contribution of this work is a benchmark dataset, I believe it is crucial to have a related works section in the main paper but not at the end of the appending.
  6. Many figures in the results section do not have a caption or label. Although their positioning is sufficient, it is better to have a label at least for referencing.
2 comments

14. Correct diagnoses can mask incomplete encounters

We examined 8,897 encounters across 333 tasks and 27 models from the main leaderboard runs, using GPT-5.4-mini as the patient simulator and must-ask judge.

StatisticResult
Encounters with a correct diagnosis7,447/8,897 (84%)
Among correct diagnoses: incomplete must-ask coverage47%
Among correct diagnoses: incomplete required actions67%
Among correct diagnoses: either gap80%
Diagnosis accuracy with full vs incomplete must-ask coverage86% vs 81%
Encounters without a recorded diagnosis60 (0.7%), including 31 from GPT-5-mini

Thus, 80% of encounters with a correct diagnosis still missed at least one required history item or clinical action. A diagnosis-only metric would count these encounters as successes.

Example: acute angle-closure glaucoma, Gemini-3.8-flash.

The patient opens with: “I suddenly got really bad pain in one eye, and now lights have rainbow rings and my vision looks doubled.”

StepAgent action
1Retrieves the patient’s chart.
2Asks about onset, triggers, and family history of glaucoma.
3Checks vital signs.
4States, “Based on your symptoms and examination, you have acute angle-closure glaucoma,” despite no recorded eye examination or eye-pressure measurement.
5Admits the patient and records the diagnosis with high confidence.

Outcome: correct diagnosis; 0/4 must-ask items credited; required-action gate failed; strict pass failed.

The agent identifies the reference diagnosis after one question, but the encounter leaves required history and safety checks incomplete, including the affected eye, severity of visual changes, associated symptoms, and allergies. It also claims examination support that is absent from the recorded trajectory. This example shows why we evaluate the assessment process alongside the final diagnosis: a correct label alone does not demonstrate that the required clinical work was completed.

Really appreciate for your detailed feedback, as the discussion period is reaching the end, we hope that our comments address your initial concerns.

Comments on the paper

1 comment
Area_Chair_Zg7COpenReview ↗

Author-Reviewer Discussion

Dear Reviewers:

Thank you for your thoughtful reviews. The authors have responded to your reviews. Please check their response and reach out to them if you have any further questions. Please act as soon as possible, as the discussion period with the authors closes on Nov 18.

Best regards,

  • AC.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.