acceptodds
Under review as a conference paper at ICLR 2027

Counterfactual Self-Evolving Agents for Evidence-Grounded Reasoning

Abstract

Self-play proposer–solver methods improve reasoning by generating tasks and learning from verified solutions. However, for evidence-identifiable tasks, where case-specific evidence and domain knowledge determine a checkable answer, self-play requires generating plausible cases whose answers can be independently verified. We introduce counterfactual self-evolution, which generates counterfactual context for reconsidering the original case. A trainable Proposer constructs targeted evidence edits and describes potential outcome changes with causal explanations. We handcraft an expert-verified counterfactual instruction-tuning dataset to teach the Proposer to generate high-quality counterfactuals across a broad range of action–outcome scenarios. Each counterfactual instruction-tuning example specifies an edit within a defined category and explains its hypothesized causal effect on the decision, teaching the Proposer to reason systematically about what changes and why. We instruction-tune the Proposer on these examples, then formulate a fine-tuning reward that integrates feedback from the Solver and Verifier. Across diverse counterfactual scenarios, this reward favors high-quality counterfactuals and warranted revisions, while penalizing changes that overturn correct decisions. The counterfactual context aims to correct errors and strengthen confidence in correct decisions. Accepted counterfactuals accumulate in memory that supplies in-context evidence to the frozen Solver; the Solver adapts through evolving context rather than weight updates. We apply the framework to clinical reasoning, fact verification, and business reasoning. Our evaluation tracks performance over successive rounds as counterfactual memory grows, including transfer to harder cases. Our method achieves superior results across diverse frontier models.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Counterfactual reasoning, Self-evolving agents

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

This paper introduces CF-Evolve, a Proposer–Verifier–Solver framework that improves a frozen Solver through learned counterfactual guidance and accumulated memory. The Proposer is trained to correct erroneous decisions while preserving correct ones. Experiments cover clinical prediction, diagnosis, business reasoning, and fact verification.

Critical strengths
  1. The distinction between beneficial corrections and harmful revisions is well motivated and more informative than reporting accuracy alone.
  2. Learned guidance produces substantial improvements in some settings, particularly DiagnosisArena.
  3. The appendix transparently reports implementation details, negative transfer results, and limitations of the theoretical guarantees.
Critical weaknesses

1.The clinical curriculum progressively enables patient-specific counterfactual guidance, rather than demonstrating transfer from earlier cases to new ones. The separate held-out memory experiment yields approximately one percentage point of improvement, with confidence intervals including zero. *Thus, evidence for sustained self-evolution is limited. *

  1. Appendix D states that the clinical RL experiment omits the correctness veto in Equation (5) and assigns zero weight to the Verifier reward. Consequently, the results do not directly validate the complete proposed optimization mechanism. I concern about the reported implementation differs from the proposed objective.

  2. A matched, supervised correction or evidence-review baseline is needed to distinguish counterfactual benefits from additional supervision, model calls, and acceptance policies. Currently, the contribution of counterfactual reasoning is insufficiently isolated.

Questions for authors
  1. Which experiments implement Equation (5), including the correctness veto and Verifier feedback?
  2. How does CF-Evolve compare with an ordinary correction Proposer under matched supervision and inference budgets?
  3. Can the authors demonstrate sequential memory transfer on independent new cases while keeping current-case counterfactual processing fixed?
  4. Can the main clinical results be reproduced on a test cohort untouched during development?
3 comments

Thank you for your constructive feedback. To further clarify the contributions and mechanisms of CF-Evolve, we have conducted additional experiments examining the Verifier's role, comparisons with ordinary correction, and the impact of accumulated counterfactual supervision on challenging tasks. We also provide additional clinical evaluations to assess generalization.

1. Reward Ablation and Verifier Contribution

The clinical RL experiment reported in the appendix was a preliminary demonstration. Due to runtime constraints, retrieval-based verification (RAG) had not yet been integrated into that run. We have now supplemented it with evaluations incorporating the correctness veto and the Verifier, using the same 10,471 IHM records and full-clinical inputs. Adding the Verifier on top of the veto improves AUROC and AUPRC by 0.108 and 0.077 percentage points, respectively. The three configurations are summarized below.

Table R3-1. IHM reward ablation. All metrics are percentages.

Reward configurationAccuracyAUROCAUPRCF1
Additive reward, no veto or Verifier84.165875.531132.371637.1970
Veto only, no Verifier82.427775.563031.450639.2338
Veto + Verifier81.692375.671331.527238.6167

2. Comparison with Direct Correction

Structured counterfactual guidance enables more effective and selective correction than ordinary diagnostic feedback by identifying evidence boundaries and guiding targeted revisions while preserving correct predictions.

Under matched supervision and inference budgets, CF-Evolve outperformed answer-blind Direct Correction across all six model–dataset comparisons on DiagnosisArena and MMMU Business with Qwen, Llama, and MedGemma. The gains over Direct Correction were particularly substantial on DiagnosisArena (+28.3, +32.6, and +22.5 percentage points, respectively), compared with +0.7, +4.8, and +11.0 points on MMMU Business. The correction–harm analysis further illustrates this advantage: on DiagnosisArena, CF-Evolve achieved 43 corrections with zero harms for Llama, versus zero corrections and two harms for Direct Correction. On MMMU Business, harmful revisions decreased from 24 to 13 for Llama and from 31 to 1 for MedGemma.

3. CF Supervision and Harder Tasks

Our self-evolution experiments aim to improve the Proposer by learning from accumulated counterfactual experience, rather than simply reusing historical cases, which can introduce inapplicable reasoning. We trained the IHM Proposer on progressively accumulated cases ordered from easy to hard and evaluated LOS Proposers trained with increasing amounts of CF supervision. On IHM, the balanced-accuracy gain over no CF increases from 5.95 percentage points on the 24-hour task to 8.76 and 9.33 points on the 72- and 168-hour tasks, respectively. On LOS, 2-class improves accuracy by 2.05 and 6-class improves 4.83 points. These results suggest that additional CF supervision can provide greater benefits in more challenging settings, although the relationship is not uniformly monotonic.

4. Data Usage and Additional Evaluation

The 3,000 CF training examples were constructed from training records, not from test cases. The “test set” comes from the early oracle experiment, where we analyzed reasoning failure modes and evaluated the potential benefit of counterfactual guidance. In subsequent pipeline development and evaluation, those inspected cases were treated as validation data and excluded from the final test set.

To avoid potential data leak problem, we also assess whether the benefits extend beyond the oracle-analysis cases, we additionally evaluated the original Proposer examples on 20,000 additional MIMIC-IV stays and 325 eICU cases. The Proposer and Solver weights remained frozen, with no additional training on these evaluation records.

Table R3-2. Additional MIMIC-IV and external eICU evaluation. All metrics are percentages.

Evaluation cohortMethodnAccuracyAUROCAUPRCF1
Additional MIMIC-IV staysFrozen Solver20,00058.8469.5921.1329.41
Additional MIMIC-IV stays+ CF20,00062.7973.8624.3532.21
eICUFrozen Solver32588.6277.8618.4527.45
eICU+ CF32590.7785.3330.6425.00

CF improves all four metrics on the additional MIMIC-IV cohort, including gains of 4.27 percentage points in AUROC and 3.21 points in AUPRC. On the external eICU cohort, AUROC and AUPRC improve by 7.47 and 12.18 points, respectively, while F1 decreases. These results support benefits beyond the original oracle-analysis cases and demonstrate transfer of the MIMIC-trained Proposer to an external clinical dataset.

The core architectural idea a learnable feedback model, a frozen execution model, and reward signals tied to whether revisions yield correct outputs has been explored in prior work like RL4F (ACL 2023), Retroformer (ICLR 2024), and CTRL (ICML 2025). Other lines of work such as Reflexion, ReasoningBank, Agent Hospital and GSEM pursue memory-based self-evolution without weight updates. In clinical diagnosis, counterfactual case modification was studied by You et al. (2026).The main incremental advance here is limiting feedback to structured counterfactuals and drawing a clear distinction between correction and retention. The theoretical analysis appears to be a tailored instantiation of known results. I recommend discussing RL4F, Retroformer and CTRL and making the key differences explicit.

AuthorsOpenReview ↗

Thank you for your suggestion!

Dear Reviewer VmcY,

We appreciate your suggestion to clarify the differences from additional baselines. We carefully reviewed the reference you suggested and added a new paragraph to the Related Work section discussing all of the relevant methods you mentioned. Here we provide further details to clarify the distinctions. Specifically, our novelty lies in three aspects:

  1. Feedback design: RL4F and CTRL directly critique a candidate answer, while Retroformer reflects on a failed trajectory. In contrast, our feedback is built on causally grounded counterfactuals: it edits the evidence in the case itself.

  2. Supervision: preserving a correct answer is an explicit training objective in our method. We assess whether a counterfactual is useful or harmful, functioning as a safety guard. For example, our instruction-tuning data include preservation examples, and the clinical reward penalizes harmful revisions more heavily than it rewards corrections.

  3. Task formulation: CTRL is a code-generation method trained against unit tests, while Retroformer is trained on environment returns from repeated attempts at the same task. Because we use counterfactuals to support decision-making, our method evaluates each counterfactual by its effect on the Solver’s decision for the original labeled case, together with feedback from the Verifier.

We have added new self-evolution experiments in Section 4.3 of the revised draft. Beyond memory evolution, i.e. growing the CF context while keeping the Proposer fixed, we now study evolution of the Proposer itself. Specifically, we progressively tune the Proposer using batches of increasingly high-quality counterfactual instruction-tuning examples. This produces a clearer trend: performance improves across all tasks as training grows, with larger gains in more challenging settings. These results show that memory evolution and Proposer evolution are complementary forms of self-evolution, each providing distinct insights. Our approach therefore extends beyond memory evolution alone.

Circle back to the additional related works, Retroformer is methodologically closest to our approach. RL4F and CTRL update only the feedback model’s weights, while Reflexion, ReasoningBank, Agent Hospital, and GSEM update only memory. In contrast, both our method and Retroformer update model behavior as well as memory. We will prioritize an implementation-based comparison with Retroformer, along with other methods you mentioned, and further clarify the distinctions between these methods in the revision.

Thank you again for your valuable suggestions for helping us further improve our work. Please do not hesitate to let us know if you have additional questions.

Sincerely,

Authors

Reviewer_R5iXOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a three-agent system in which a LoRA-tuned Proposer generates evidence-grounded counterfactuals for a frozen Solver, a retrieval-based Verifier checks them against external sources, and a reward explicitly distinguishes corrections of wrong answers from preservations of correct ones. The Proposer is first supervised on hand-crafted, expert-verified counterfactual records and then trained for one epoch with a GRPO-style objective. Evaluation covers MIMIC-IV in-hospital mortality on 10,471 test cases, length of stay, two multiple-choice benchmarks, a fact-verification set, and a cross-case memory transfer experiment. Three propositions are given relating correction and harm probabilities, an inject-or-abstain gate, and a finite-sample bound on accumulated revisions.

Critical strengths

The paper evaluates its system with a paired correction-versus-harm decomposition rather than accuracy alone, which is the correct instrument for a method whose stated goal is to avoid flipping correct answers, and the counts are reported per model and per dataset rather than aggregated away. The authors also report results that do not support them, including a null memory-transfer experiment and a verifier ablation that undercuts the verifier, and the accompanying audits explain why rather than leaving the null unexplained. That candor is worth noting explicitly.

Critical weaknesses

One of the three named agents does nothing in the reported configuration. The clinical reward sets the verifier weight to zero, and Table 22 shows the best adoption policy is the one that explicitly ignores the retrieval signal and reaches 66.7 percent, while requiring positive verifier support drops to 48.6. A three-agent framework whose middle agent is ablated at negative cost is empirically a two-agent framework, and the headline DiagnosisArena number comes from the policy that discards it. The paper's own words about the four propositions go further: Proposition 1 is credited as a specialization of an existing result, Proposition 2's guarantee is stated not to apply to the implemented gate because the probabilities are not estimated and the confidence is not calibrated, and Proposition 3 is stated not to be enforced by the admission rule and not to certify the reported AUROC gains. Second, both headline claims fail against the paper's own evidence. The one-fourteenth-the-tuning-data framing is contradicted by Appendix F, where plain answer-only Solver SFT on the same training question IDs beats CF-Evolve on DiagnosisArena for all three backbones by 7.9, 17.4, and 7.3 points, and by the IHM row where task SFT reaches higher accuracy on 41,886 examples. The self-evolution claim underlying the title rests on a cross-case transfer experiment in which all four confidence intervals span zero with McNemar p between 0.12 and 0.17, and the authors' content audit attributes this to 90.7 percent of stored memories not changing the answer.

Questions for authors

MIMIC-IV in-hospital mortality prevalence is roughly ten percent, so a constant-survival predictor would score about 88 to 90 percent accuracy, above the 84.17 percent reported for CF-Evolve. Please add a majority-class and a threshold-swept baseline and reconsider whether accuracy should be the primary metric for this task, since the SFT-to-GRPO jump moves accuracy by 10.5 points while AUROC moves by 0.1, which looks like calibration rather than discrimination. The paper concedes that test reasoning was examined during development and Figure 3c evaluates hand-authored counterfactuals on 10,470 test cases; please clarify whether any authoring used test labels, and whether that condition can be reproduced. How often does the adoption gate fire for reasons attributable to counterfactual plausibility as opposed to agreement with the Proposer's proposed answer, given that gate for two of three backbones requires that agreement? The development audit noted in Section 2.2 used test reasoning to identify error types; please describe the boundary in more detail. Finally, the reported objective is not the correctness-veto objective the paper motivates. Will the authors reconcile Equation 5 with Equation 16, and either rescue the verifier or remove it from the architecture?

1 comment

Thank you for your feedback. We added class-imbalance baselines, reward ablations, gate analyses, and clinical evaluations.

1. IHM Class Imbalance and Threshold Analysis

With 12.3% mortality prevalence, always predicting survival achieves 87.73% accuracy, exceeding our reported 84.17%. Threshold sweeping and five-fold cross-fitted threshold selection (selecting on four folds and evaluating on the fifth) yield approximately 87.7–88.1%. We therefore prioritize AUROC, AUPRC, and MCC over accuracy.

Table R2-1. IHM comparison (n = 10,471). Metrics are percentages; MCC is ×100. Accuracy and MCC use threshold 0.5; constant-predictor MCC is zero by convention.

MethodAccuracyAUROCAUPRCMCC (×100)
Majority class (always survival)87.7350.0012.270.00
Zero-shot Solver58.3970.0022.5220.1
+ SFT Proposer73.7275.4329.8527.8
+ SFT + GRPO Proposer84.1775.5332.3728.2

Most discrimination gains come from CF instruction tuning: SFT improves AUROC by 5.4 points, AUPRC by 7.3, and MCC (×100) by 7.7 over zero-shot. GRPO adds smaller ranking improvements; its larger accuracy gain is threshold-sensitive.

2. Data Usage and Additional Evaluation

The 3,000 CF training examples came exclusively from training records. The term “test set” referred to the early oracle experiment used to analyze reasoning failures. Those inspected cases were subsequently treated as validation data and excluded from the final test set.

To assess generalization beyond these cases, we evaluated the original Proposer, trained on 3,000 CF examples, on 20,000 additional MIMIC-IV stays and 325 eICU cases. Both Proposer and Solver remained frozen, without training on these evaluation records.

Table R2-2. Additional clinical evaluation. All metrics are percentages.

Evaluation cohortMethodnAccuracyAUROCAUPRCF1
Additional MIMIC-IVFrozen Solver20,00058.8469.5921.1329.41
Additional MIMIC-IV+ CF20,00062.7973.8624.3532.21
eICUFrozen Solver32588.6277.8618.4527.45
eICU+ CF32590.7785.3330.6425.00

CF improves all four metrics on additional MIMIC-IV stays. On eICU, AUROC and AUPRC increase by 7.47 and 12.18 points, respectively, although F1 declines. These results support transfer beyond the oracle-analysis cases, including to an external clinical dataset.

3. Adoption-gate Attribution

The gate checks whether the frozen Solver independently reaches the Proposer’s anticipated answer, without seeing that answer, and filters excessive confidence drops. Across Llama and MedGemma’s frozen DiagnosisArena logs, answer disagreement contributed to 12 of 16 post-Solver rejections and confidence checks to 10, with six failing both. The headline policies did not activate RAG plausibility filtering (0/16); these rejections therefore cannot establish retrieval-based verification’s contribution.

In a separate fixed-candidate Qwen ablation, incorporating the RAG Verifier’s rejection signal reduced harmful revisions from three to two while preserving accuracy at 92/138 (66.7%). This distinguishes answer/confidence-based adoption from external evidence verification.

4. Reward Ablation and Verifier Contribution

The appendix’s clinical RL run was a preliminary demonstration without RAG verification due to runtime constraints. We added a three-way comparison on the same 10,471 IHM records with full-clinical inputs. Adding the Verifier to the veto increases AUROC and AUPRC by 0.108 and 0.077 points, respectively, while accuracy and F1 decline.

Table R2-3. IHM reward ablation. All metrics are percentages.

Reward configurationAccuracyAUROCAUPRCF1
Additive reward, no veto or Verifier84.165875.531132.371637.1970
Veto only, no Verifier82.427775.563031.450639.2338
Veto + Verifier81.692375.671331.527238.6167
Reviewer_dzSvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CF-Evolve, which trains a Proposer to generate counterfactual guidance for a frozen Solver. Instruction tuning teaches evidence edits, alternative explanations, and interventions; subsequent reinforcement learning rewards useful answer revisions. Mechanical checks, retrieval-based verification, and task-specific acceptance policies regulate the guidance. Experiments cover clinical prediction, DiagnosisArena, business questions, and fact verification with three small language-model backbones. The reported systems improve several outcomes relative to their zero-shot counterparts. The paper also studies patient-specific context accumulation and transfer from a separate counterfactual memory bank. Its appendices provide important qualifications: the clinical reward differs from the main formulation, some headline gains include preprocessing, and the held-out memory-transfer gains are small and statistically uncertain.

Critical strengths
  1. Useful improvements with frozen Solvers. The approach provides a practical way to apply supervised corrective guidance without updating the decision model. Table 19 reports substantial DiagnosisArena gains across all three backbones, with corrected answers exceeding harmed answers in each setting. The clinical experiments also report ranking metrics alongside threshold-dependent accuracy, allowing a more informative assessment of the improvements.
  2. Detailed reporting of corrections and limitations. The appendices distinguish full-system gains from subsequent counterfactual revisions, include direct Solver fine-tuning, and examine alternative acceptance policies. Table 22 shows that a REJECT-veto policy can match Qwen's 66.7% DiagnosisArena accuracy while reducing harmful revisions from three to two. The held-out memory study reports both corrections and harms with uncertainty, making its limited effect assessable.
Critical weaknesses
  1. The central reward and verification descriptions do not consistently identify the evaluated method. Equation 5 uses a correctness veto and includes Verifier feedback. Appendix D instead states that the reported clinical run uses an additive reward with weights (2, 1, 0), without the veto. Consequently, the main-text guarantee that unresolved errors always receive negative total reward does not describe that run. Separately, Qwen's headline DiagnosisArena result uses the validation-selected policy that ignores RAG verdicts (Appendix E.8). Mechanical checks remain active, and an alternative verifier policy matches its accuracy, but the headline result alone cannot establish the contribution of retrieval-based verification.
  2. Evidence for improvement through reusable memory is substantially weaker than evidence for case-specific guidance. Appendix G.1 progressively enables each patient's own counterfactual context on a fixed cohort; this does not isolate transfer from earlier cases. Appendix G.2 does test disjoint memory sources and evaluation cases. Its selected bank improves accuracy by 1.13 percentage points, with 17 corrections, nine harms, and McNemar p = 0.17; all reported gain intervals include zero. These results support a tentative memory-transfer benefit, while the stronger claim of progressive improvement on harder cases remains insufficiently established.
  3. The performance advantage needs tighter attribution and qualification. The existing data-matched Solver-SFT baseline exceeds CF-Evolve on DiagnosisArena for every backbone: 74.6/79.0/70.3% versus 66.7/61.6/63.0% (Table 23). On business questions, MedGemma's counterfactual revision adds only 1.4 points after output recovery, compared with the 11.7-point full-system gain over the separately reported zero-shot result (Appendix E.6). These findings still support a useful frozen-Solver alternative, but not general superiority attributable to counterfactual evolution.
  4. Development was not fully independent of clinical test inputs. Appendix C.4 explicitly says that development audits inspected test reasoning to identify error types. The stated exclusion of test facts and outcomes from training messages is valuable, but does not remove possible adaptation to test-set error patterns. A fresh evaluation cohort is needed to establish how well the developed procedure generalizes beyond that audit.
Questions for authors
  1. Which experiments use Equation 5, which use the additive reward, and how much do the correctness veto and Verifier reward contribute under otherwise matched settings?
  2. Can you separate current-case counterfactual guidance from cross-case memory on the same frozen test set, and measure transfer to harder cases with adequate uncertainty?
  3. How do the systems compare with Solver-SFT and a matched reconsideration baseline when preprocessing, supervision, candidate count, and inference cost are aligned?
  4. Which design decisions were informed by the clinical test-reasoning audit, and can the final frozen pipeline be evaluated on a previously unexamined cohort?
  5. Can the abstract and main results explicitly reflect the small-model evaluation, the stronger DiagnosisArena Solver-SFT baseline, and the uncertain memory-transfer effect?
1 comment

Thank you for your constructive feedback. We have conducted additional experiments to address your concerns, including reward ablations, stronger baseline comparisons, CF supervision scaling, and external validation on MIMIC-IV and eICU.

1. Reward Ablation and Verifier Contribution

The clinical RL experiment reported in the appendix was a preliminary demonstration. Due to runtime constraints, retrieval-based verification (RAG) had not yet been integrated into that run. We have now supplemented it with evaluations incorporating the correctness veto and the Verifier, using the same 10,471 IHM records and full-clinical inputs. Adding the Verifier on top of the veto improves AUROC and AUPRC by 0.108 and 0.077 percentage points, respectively. The three configurations are summarized below.

Table R1-1. IHM reward ablation. All metrics are percentages.

Reward configurationAccuracyAUROCAUPRCF1
Additive reward, no veto or Verifier84.165875.531132.371637.1970
Veto only, no Verifier82.427775.563031.450639.2338
Veto + Verifier81.692375.671331.527238.6167

2. CF Supervision and Harder Tasks

Our self-evolution experiments aim to improve the Proposer by learning from accumulated counterfactual experience, rather than simply reusing historical cases, which can introduce inapplicable reasoning. We trained the IHM Proposer on progressively accumulated cases ordered from easy to hard and evaluated LOS Proposers trained with increasing amounts of CF supervision. On IHM, the balanced-accuracy gain over no CF increases from 5.95 percentage points on the 24-hour task to 8.76 and 9.33 points on the 72- and 168-hour tasks, respectively. On LOS, 2-class improves accuracy by 2.05 and 6-class improves 4.83 points. These results suggest that additional CF supervision can provide greater benefits in more challenging settings, although the relationship is not uniformly monotonic.

3. Comparison with Direct Correction

Structured counterfactual guidance enables more effective and selective correction than ordinary diagnostic feedback by identifying evidence boundaries, guiding targeted revisions, and helping preserve correct predictions.

Under matched supervision and inference budgets, CF-Evolve outperformed answer-blind Direct Correction across all six model–dataset comparisons on DiagnosisArena and MMMU Business with Qwen, Llama, and MedGemma. The gains over Direct Correction were largest on DiagnosisArena (+28.3, +32.6, and +22.5 percentage points, respectively), compared with +0.7, +4.8, and +11.0 points on MMMU Business. The correction–harm analysis further supports this advantage: on DiagnosisArena, CF-Evolve achieved 43 corrections with zero harms for Llama, compared with zero corrections and two harms for Direct Correction. On MMMU Business, harmful revisions decreased from 24 to 13 for Llama and from 31 to 1 for MedGemma. Unlike direct Solver-SFT, CF-Evolve achieves its improvements while keeping the Solver parameters frozen.

4. Data Usage and Additional Evaluation

The 3,000 CF training examples were constructed from training records, not from test cases. The “test set” comes from the early oracle experiment, where we analyzed reasoning failure modes and evaluated the potential benefit of counterfactual guidance. In subsequent pipeline development and evaluation, those inspected cases were treated as validation data and excluded from the final test set.

To avoid potential data leak problem, we also assess whether the benefits extend beyond the oracle-analysis cases, we additionally evaluated the original Proposer examples on 20,000 additional MIMIC-IV stays and 325 eICU cases. The Proposer and Solver weights remained frozen, with no additional training on these evaluation records.

Table R1-2. Additional MIMIC-IV and external eICU evaluation. All metrics are percentages.

Evaluation cohortMethodnAccuracyAUROCAUPRCF1
Additional MIMIC-IV staysFrozen Solver20,00058.8469.5921.1329.41
Additional MIMIC-IV stays+ CF20,00062.7973.8624.3532.21
eICUFrozen Solver32588.6277.8618.4527.45
eICU+ CF32590.7785.3330.6425.00

CF improves all four metrics on the additional MIMIC-IV cohort, including gains of 4.27 percentage points in AUROC and 3.21 points in AUPRC. On the external eICU cohort, AUROC and AUPRC improve by 7.47 and 12.18 points, respectively, while F1 decreases. These results support benefits beyond the original oracle-analysis cases and demonstrate transfer of the MIMIC-trained Proposer to an external clinical dataset.

5. Manuscript Updates

We have revised the manuscript accordingly in response to your suggestions, including updates to the abstract and main results to better reflect our additional experimental findings and clarify the scope of our contributions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewers,

Thank you for your constructive feedback and for recognizing the value of counterfactual guidance for frozen Solvers and our correction–harm analysis. In response to the reviews, we conducted additional experiments, including stronger baseline comparisons, component ablations, and external clinical evaluations. We took additional time to complete and validate these experiments to ensure that our responses are supported by concrete evidence. We have also revised the manuscript to incorporate the new findings, clarify the methodology, and better reflect the scope of our contributions. Please feel free to let us know if you have additional questions.

Sincerely,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.