acceptodds
Under review as a conference paper at ICLR 2027

CEF-OPD: Learning to Correct Consequential Errors in Radiology Report Generation

Abstract

Radiology report generators are typically trained on reference prefixes but deployed on their own predictions. An early clinical contradiction may influence subsequent claims, yet its local prediction error does not directly quantify the downstream clinical consequences. We introduce a clinical error frontier on-policy distillation (CEF-OPD) framework that allocates corrective supervision according to the estimated downstream consequences of model-generated errors. CEF-OPD identifies the earliest contradiction with extracted reference facts and constructs a reference-supported minimal correction. Paired continuations from the erroneous and corrected prefixes estimate the change in downstream clinical cost, excluding the directly edited claim. This estimate determines the weight assigned to local correction and distillation, emphasizing errors with greater estimated downstream impact. Reference clinical facts provide privileged guidance for self-distillation from the pre-error state, targeting the decision that introduces the contradiction. We characterize the paired cost difference as a model-internal cost-to-go advantage and bound local clinical risk differences through propagation-weighted distillation under stated assumptions. Experiments on MIMIC-CXR and IU X-ray datasets demonstrate the effectiveness of our proposed method.

Primary area:
generative models
Keywords:
Radiology report generation, On-policy distillation, Clinical error propagation, Exposure bias

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_aX7KOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces CEF-OPD for corrective training in radiology report generation. A frozen RadGraph-XL extractor and deterministic rules identify the earliest generated claim that explicitly contradicts reference facts and construct a minimal correction preserving other extracted facts. Paired continuations from the erroneous and corrected claims estimate the downstream reference-relative cost difference.

Critical strengths

Sections 3.1–3.3 place supervision before the contradictory claim, targeting the decision that introduces the error. Excluding the edited claim from scoring and sharing the remaining reference targets prevent a positive propagation benefit from arising merely because the original error was edited away. Retaining omission penalties also discourages empty continuations when reference targets remain.

On MIMIC-CXR, the full method improves GREEN from 0.606 to 0.651, RG-ER from 0.397 to 0.326, and CE-F1 from 0.461 to 0.485 over Continued-SFT. These results support the practical potential of the training procedure.

Critical weaknesses

Equation (9) normalizes by batch size rather than the sum of weights. Replacing w_i=[ΔC_i]+ with uniform weights therefore changes both the allocation of supervision and its overall strength relative to SFT. Appendix B usefully matches episodes and optimization budgets and includes a lambda sensitivity analysis, but does not establish that the uniform baseline receives independent lambda tuning or matched mean weight. Consequently, the gains do not yet isolate the information supplied by ΔC. A constant-weight baseline matched to the mean weight, and a shuffled-weight baseline preserving the weight distribution, zero-weight fraction, and overall scale, would directly test this claim. Theorem 1 holds for arbitrary fixed nonnegative weights and bounds deviation from the mixed target's local risk; it does not establish that ΔC weights outperform other weights or that an update reduces free-running generation risk.

Section 3.1 relies on custom normalization, anatomical compatibility, claim segmentation, and editing rules, excluding several ambiguous cases. Validation of the underlying RadGraph-XL extractor does not by itself validate these additional decisions. Equation (2) checks edits using the same extractor, so shared extraction blind spots can survive this check.

Questions for authors
  1. What exact weight is used in the uniform baseline? Was its lambda selected independently, or was its mean auxiliary weight matched to CEF-OPD?

  2. How many examples remain at each stage: generated reports, detected contradictions, accepted edits, and positive-weight episodes? Which error types do they cover?

  3. Can you quantify teacher–student support mismatch when conditioning the teacher on reference facts, and test anchored/partial-privilege targets to mitigate hindsight?

Reviewer_A6y5OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Clinical Error Frontier On-Policy Distillation (CEF-OPD) to improve the clinical accuracy of radiology report generation. The method identifies the earliest clinical contradiction in a model-generated report, constructs a reference-supported correction, and uses paired continuations from the erroneous and corrected prefixes to estimate the error's downstream impact. This estimated impact is then used to weight local correction and privileged on-policy self-distillation, prioritizing errors with greater downstream consequences.

Experiments on MIMIC-CXR and IU-Xray demonstrate improvements in clinical accuracy and report quality over existing methods. Ablation studies support the contributions of frontier correction, privileged distillation, and propagation-aware weighting, while paired-continuation analyses provide evidence that correcting different errors can have substantially different downstream effects.

Critical strengths
  1. The paper introduces a sensible strategy for prioritizing errors based on their estimated downstream consequences, rather than treating all prediction errors equally. The use of paired continuations to estimate error propagation provides a meaningful signal for allocating corrective supervision.
  2. The paper demonstrates consistent improvements on two radiology report generation benchmarks. Importantly, the ablation studies systematically evaluate the contributions of frontier correction, privileged on-policy distillation, and propagation-aware weighting. The paired-continuation analysis further supports the underlying hypothesis that different errors have different downstream impacts.
Critical weaknesses
  1. Although the proposed weighting improves performance over uniform weighting, the gains are relatively modest.
  2. The choice of weight, defined as the positive part of the downstream cost difference, is largely heuristic. The paper does not investigate alternative mappings, such as threshold-based, nonlinear, or baseline-adjusted weighting. Consequently, it remains unclear whether the specific weighting formulation is well justified。
Questions for authors

What motivates the specific choice (w_i=\max(0,\Delta C_i))? In particular, why should the supervision strength scale linearly with the estimated downstream cost reduction, and why should errors with non-positive (\Delta C_i) receive zero additional supervision? Could the authors compare alternative weighting functions, such as binary, nonlinear, or baseline-adjusted weighting.

Reviewer_W2uHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training framework that identifies the earliest reference-supported contradiction in a generated report and constructs a minimal correction. Paired continuations from erroneous and corrected prefixes estimate downstream clinical benefit, excluding the edited claim. This benefit weights local correction and self-distillation from a reference-informed teacher sharing the student’s parameters. All auxiliary components are restricted to training.

Critical strengths
  1. Well-motivated supervision strategy. Targeting the pre-error state is a sensible way to supervise decisions made under model-generated context. Excluding the corrected claim from downstream scoring, while retaining omission penalties, helps distinguish propagation effects from the direct benefit of replacing an incorrect statement.

  2. Useful component comparisons. The paper includes continued supervised fine-tuning, correction-only and distillation-only variants, uniform weighting, and alternative corrective-training strategies. On MIMIC-CXR, the full method improves GREEN from 0.602 to 0.651 and clinical F1 from 0.458 to 0.485 relative to SFT

Critical weaknesses
  1. The benefit of propagation-aware weighting is not cleanly isolated. Equation 9 normalizes the auxiliary loss by batch size rather than total weight. Consequently, replacing uniform weights with $w_i=[\Delta C_i]_+$ generally changes both the relative allocation of supervision and its overall strength relative to SFT. Unless the weights are explicitly rescaled, improvements could partly result from changing the effective auxiliary-loss coefficient rather than identifying consequential errors. Appendix B matches episodes and optimization budgets and examines sensitivity to λ\lambda, but does not establish that the uniform baseline has matched mean weight or independently tuned loss strength. A mean-matched constant-weight baseline and a shuffled-weight baseline would better isolate the information contributed by ΔC\Delta C. This is particularly important because propagation-aware weighting is the principal distinction from frontier-uniform training.

  2. Frontier detection, correction validation, and downstream cost estimation depend on the same extractor and normalization rules. Re-extracting an edited claim establishes consistency under that representation, but does not independently establish that the contradiction was correctly identified or that the edit preserves clinical meaning. The implementation specifies a confidence threshold of 0.9 without adequately defining its calculation or reporting frontier-detection accuracy, correction accuracy, and acceptance coverage. An expert audit of accepted and rejected examples would strengthen confidence in the supervision. The paper appropriately describes its cost as reference-relative fidelity; the remaining question is how reliably this proxy identifies clinically meaningful training targets.

  3. The mixed-target identity and local risk bound appear consistent with their stated assumptions. However, the bound applies to arbitrary nonnegative weights and follows from a mixture-distribution identity, Pinsker’s inequality, and Cauchy–Schwarz. It therefore does not demonstrate that propagation-derived weights are preferable to uniform or alternative weights.

  4. The principal tables provide point estimates without confidence intervals. In particular, the incremental MIMIC-CXR clinical F1 improvement over uniform weighting is 0.478 to 0.485. It does not establish the uncertainty of the main paired comparisons. The matched ablations are valuable, but comparisons against factuality-oriented post-training methods already cited in the manuscript, such as RRG-DPO or SAT-RRG, are absent from the reported tables. Such comparisons under the same backbone and training budget would better establish the contribution beyond the selected OPD variants.

Questions for authors
  1. Why gate both corrective objectives by downstream benefit? When ΔCi≤0\Delta C_i\leq0, both local correction and distillation receive zero weight, even if the frontier claim is demonstrably incorrect. SFT remains active, but does not directly supervise that model-generated state. Would retaining an unweighted correction loss while weighting only distillation better address isolated or terminal errors?

  2. Please specify accepted episode counts, refresh frequency, trainable parameters, post-training duration, and rollout cost. These details are not sufficiently specified in the supplied manuscript’s description of periodic episode collection and optimization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.