acceptodds
Under review as a conference paper at ICLR 2027

Conformity Supervised Evidential Deep Learning

Abstract

While deep neural networks achieve remarkable performance, their tendency to produce overconfident predictions remains a critical issue. Evidential deep learning (EDL) solves this problem by modeling class probabilities with a Dirichlet distribution and quantifying the uncertainty in the model’s predictions. However, conventional EDL lacks an explicit supervision for uncertainty quantification, which leads to a risk of assigning excessive evidence to unreliable samples. In this paper, we propose Conformity Supervised Evidential Deep Learning (CSEDL), a EDL framework that learns class prediction and uncertainty through complementary supervision. Specifically, we introduce an extra uncertainty head and the rank of nonconformity score as a sample difficulty metric to supervise it without additional annotations. Together with the direction head supervised by ground-truth labels, CSEDL forms a dual-head architecture, and the two heads are then combined to reconstruct Dirichlet parameters for final prediction. Extensive experiments on out-of-distribution (OOD) detection, tumor referral prediction, autonomous-driving trajectory risk forecasting, trusted multi-view classification and calibrating large language models (LLMs) indicates that CSEDL improves uncertainty quantification ability across multiple tasks. **Code is available in the supplementary materials.**

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Evidential deep learning, uncertainty estimation, conformal-rank

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_gPC9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that conventional evidential deep learning (EDL) provides no explicit training target for uncertainty, so evidence can grow on all inputs as training proceeds (Eq. 3, Fig. 1). The proposed CSEDL splits the output layer into a direction head π, trained with cross-entropy, and a scalar uncertainty head mapped to u ∈ (0.05, 0.95). The two are recombined into Dirichlet parameters α_k = 1 + (K/u − K)π_k, so the predictive distribution is p_k = u/K + (1 − u)π_k. The uncertainty head is regressed onto a "conformal rank": the rank of each training sample's nonconformity score, 1 − p_y, among the scores of recent training batches held in a first-in-first-out cache, smoothed with a sigmoid. The standard EDL KL regularizer on non-target evidence is kept. Experiments cover OOD detection with CIFAR-10 as the in-distribution set and six OOD sets, tumor referral on PatchCamelyon, trajectory-risk forecasting on Argoverse 1, multi-view classification with missing views, and calibration of a fine-tuned Qwen2.5-3B on four multiple-choice QA sets. CSEDL reports the best average OOD AUROC (91.43, against 89.46 for a five-model deep ensemble) and the best results on most other metrics. The supplement provides code for the CIFAR-10 OOD experiment.

Critical strengths

S1. A simple, low-cost mechanism with strong OOD results. CSEDL adds one scalar head and a cache of recent nonconformity scores. The supplementary script implements it compactly and consistently with Eqs. 4–16 and the stated hyperparameters. Training with the cache takes 218.1 s, against about 340 s when the scores are recomputed over the training set once per epoch (Fig. 4b). With a single model, CSEDL gives the best average OOD AUROC in Table 1 (91.43 vs. 89.46 for deep ensembles), with large margins on SVHN (97.01 vs. 93.59) and LFWPeople (99.44 vs. 93.10). This is consequential because a single-model plug-in that outperforms a five-model ensemble at OOD detection would be useful in practice, provided the gains survive the concerns below.

S2. Breadth of evaluation. The method is tested on five task families: image OOD detection, medical referral, driving risk, multi-view fusion and LLM calibration. It is compared with up to seven EDL variants (three, plus IB-EDL, for the LLM task) and, for OOD detection, with deep ensembles, PostN and NatPN, with means and standard deviations reported. CSEDL is best or close to best on most metrics. This matters because it suggests the effect is not tied to one setting, which is also why the attribution problems in W2 matter.

Critical weaknesses

W1. The uncertainty target is a uniform, in-sample rank. On CIFAR-10 this appears to leave the predictive distribution strongly underconfident, and the training signal is partly self-referential.

  • Uniform by construction. The target is the empirical CDF of the nonconformity score (Eqs. 11–13), which the paper describes as "a smooth and uniform measure of uncertainty" (l.245). Over the training data it is therefore approximately uniform on [0, 1], however accurate the model is. On CIFAR-10 the learned uncertainty follows suit. In Fig. 5 (epoch 100), CIFAR-10 samples are spread over all five uncertainty bins (about 15–28% each), with about 28% in [0.8, 1], the bin that holds about 93% of SVHN samples.
  • Consequence for calibration on CIFAR-10.
    • Because p_k = u/K + (1 − u)π_k (Eq. 8), each prediction's confidence is at most 1 − (1 − 1/K)u.
    • Fig. 1 reports u_ID = 0.3956. The four u values in Fig. 1 equal K/(K + mean evidence), so they are at most the arithmetic mean of u, and the bound below is conservative.
    • If Fig. 1 shows CIFAR-10 test samples for the reported model, average confidence is at most about 0.64, against 92.5% accuracy (Table 1). Since ECE is at least the gap between accuracy and mean confidence, CIFAR-10 ECE would be at least about 28 points. Fig. 5's histogram implies a similar or larger gap.
    • Whether the figures describe the Table 1 model needs confirming. The released script trains for 20 epochs by default, while Figs. 1 and 5 run to epoch 100.
    • The paper reports no in-distribution calibration for CIFAR-10 or PCam, although the script already collects p on the test set. Calibration appears only for the LLM and multi-view tasks, where the reported ECE is low.
    • By the same bound, the LLM results require much smaller u. With K = 4, an ECE of 2.54% at 94.07% accuracy (SciQ, Table 4) allows a mean u of at most about 0.11. The paper does not explain why the same uniform target yields small u there but not on CIFAR-10. (Fused multi-view predictions escape the bound because Dempster–Shafer fusion multiplies view-level uncertainties.)
  • Self-referential target. Eq. 10 computes the nonconformity score from p, which contains u, and the released code does the same (nonconformity in csedl.py uses prob). This gives Z_i = 1 − π_y + u(π_y − 1/K). For training samples the classifier already fits (π_y ≈ 1), Z_i ≈ (1 − 1/K)u. The target then reduces to the rank of the sample's current uncertainty among the uncertainty values the model produced for earlier batches. The stop-gradient prevents a direct collapse, but it is unclear what information the target carries once training accuracy saturates. The paper reports neither the training accuracy nor any analysis of this.
  • No conformal guarantees. The scores are computed in-sample, on training data, against a cache of the last 1,024 training scores (four batches of 256 in the released code). The cache is filled under earlier parameters, and there is no held-out calibration set and no exchangeability. None of conformal prediction's guarantees therefore apply. The claim that the rank "guarantees a better uncertainty quantification performance" (l.245–246) is supported only empirically, by the Entropy and Raw-score variants in Table 1 on one benchmark.

This is consequential because the paper's central claim is better uncertainty quantification, and these issues stem from the design of the target rather than from a particular experiment.

W2. The evaluation does not isolate what drives the gains.

  • Ablation.

    • Without the KL term, conformal-rank supervision gives OOD AUROC of 76.74 / 71.25 / 80.22 on SVHN / CIFAR-100 / GTSRB (Table 5). That is below vanilla EDL on SVHN and CIFAR-100 (79.84 / 79.20, Table 1).
    • The full model's 97.01 / 87.59 / 90.68 appears only when both terms are combined. "Complementary roles" (l.477–479) does not explain jumps of 10–20 points.
    • A likely reason, not discussed in the paper: because α's non-target entries scale with K/u − K, the KL term also trains u directly. By my calculation, consider a training sample whose π puts at least 10% of its mass on wrong classes. Unless u is already high, the KL gradient on the uncertainty logit exceeds the largest possible rank-loss gradient (0.02 per sample with λ_unc = 0.01) by roughly 10× to over 1,000×. For confidently fitted samples it is near zero. The KL term may therefore be the main uncertainty signal for hard samples, with the rank loss mainly spreading the easy ones.
    • The variant without either term sits at chance (50.00 / 49.98 / 50.01). The uncertainty head then receives no gradient (in csedl.py, L_dir uses only the direction head's logits), so this is not an informative ablation.
    • Table 5 reports no standard deviations.
  • Missing baselines. Classification is decided entirely by the direction head (argmax p = argmax π under Eq. 8), which is trained with cross-entropy rather than an EDL loss. CSEDL's accuracy (92.50%) exceeds every EDL baseline (86.23–90.40%) and the deep ensemble (92.33%). Yet its own cross-entropy-only variant reaches 87.31% (Table 5), so the source of the extra accuracy is unclear. The paper also reports no standard single-model baseline (a cross-entropy model scored by maximum softmax probability, entropy or energy) and no auxiliary-confidence baseline such as ConfidNet. Without them, it is unclear whether CSEDL improves on standard practice outside the EDL family.

  • Selection on test data. App. F chooses M_max = 1024 "because it achieves the highest AUROC among the tested cache sizes" (l.874–876). The paper describes no validation split, and the released script trains on the full CIFAR-10 training set and evaluates directly on the test sets. The AUROC at 1024 in Fig. 4a equals Table 1's SVHN result (97.01), so the cache size appears to have been chosen on test OOD data. The choice matters: at M_max = 256, the SVHN margin over deep ensembles shrinks from 3.42 to 0.11 points. The paper does not say how λ_unc = 0.01 and T = 0.05 were chosen.

  • Reproducibility. The supplement contains a single script, for CSEDL on the CIFAR-10 OOD benchmark. There is no code for the baselines, the Entropy and Raw-score variants, Figs. 1, 4 and 5, or the other four tasks. The script's argument checks also reject λ_unc = 0 and λ_kl = 0, so Table 5 cannot be run as released. Most reported results therefore cannot be reproduced from the supplement, although the paper says the code is provided "to facilitate the reproduction of our results" (l.500–501).

  • Positioning and scope. The paper claims to be the first to supervise uncertainty explicitly in EDL (l.49–50). Sec. 2.3 also states that CP-based methods typically use conformal scores post hoc rather than as a training signal (l.153–155). Both overlook closely related work:

    • auxiliary confidence heads, such as ConfidNet, which regresses the true-class probability (Corbière et al., NeurIPS 2019), and DeVries & Taylor (2018);
    • conformal training (Stutz et al., ICLR 2022; Einbinder et al., NeurIPS 2022). Einbinder et al. penalize the deviation of hold-out conformity scores from uniformity using a smoothed empirical CDF, which is close to Eq. 13;
    • analyses showing that EDL objectives do not yield faithful epistemic uncertainty, and that the resulting measures are relative rather than absolute (Bengs et al., NeurIPS 2022; Jürgens et al., ICML 2024; Shen et al., NeurIPS 2024).

    Beyond OOD detection, comparisons are limited to EDL variants (Tables 2, 3, 6), and Table 4 uses a single LLM. Several of Table 4's differences are within one standard deviation, although the text says it "proves the generalization ability" (l.462–463).

This is consequential because the evidence that the conformal-rank target drives the gains comes from within-method ablations on one benchmark. That evidence also depends on the KL term, uses a cache size apparently chosen on test data, lacks standard single-model baselines, and mostly cannot be reproduced from the supplement.

Questions for authors
  1. Additional result requested (could change my assessment; W1).
    • What are the ECE and NLL of the final prediction p (Eq. 8) on the CIFAR-10 and PCam test sets? The released script already collects p for the test set.
    • Do Fig. 1 and Fig. 5 show test samples of the model in Table 1? The script trains for 20 epochs by default, while the figures run to epoch 100.
    • The u values in Fig. 1 equal K/(K + mean evidence), and u_ID = 0.3956 caps mean confidence at about 0.64, against 92.5% accuracy. Why, by contrast, does Table 4 report ECE of 2.5–6.3%? Under Eq. 8 that requires a mean u of at most about 0.11–0.33.
  2. Clarification (W1). The released code computes Z_i from p, as Eq. 10 states. What is the training accuracy at the end of training? Once most training samples have π_y ≈ 1, so that Z_i ≈ (1 − 1/K)u, what does the rank target capture?
  3. Clarification and additional result (W2). Which data were used to choose M_max, λ_unc and T? Could you report Table 1 with these chosen on held-out in-distribution data only?
  4. Additional result requested (W2).
    • How does CSEDL compare with a single cross-entropy model scored by maximum softmax probability, entropy or energy, and with a ConfidNet-style confidence head on the same backbone?
    • Why does conformal-rank supervision without the KL term (Table 5) fall below vanilla EDL (Table 1) on SVHN and CIFAR-100? How much of u's training signal comes from the KL term, which acts on u through α?
    • What are the standard deviations in Table 5?
  5. Clarification. The supplement gives the CIFAR-10 setup: VGG16-BN with global average pooling, AdamW, cosine schedule, batch size 256, 20 epochs and seed 42 by default.
    • Did Table 1 use 20 epochs, or the 100 shown in Figs. 1 and 5?
    • How many runs are behind the ± values?
    • What setups were used for PCam, Argoverse, the multi-view task and the Qwen2.5-3B fine-tuning, and which quantity is the uncertainty score for each?
    • Will code for the baselines, ablations and other tasks be released?
Reviewer_ePzpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Conformity Supervised Evidential Deep Learning (CSEDL), an EDL method that separately learns class prediction and uncertainty estimation. The prediction is supervised by the ground-truth label, and the uncertainty is supervised by the ranks of nonconformity scores and then combined to construct a Dirichlet distribution for final predictions. CSEDL demonstrates strong UQ performance across OOD detection, tumor referral prediction, and autonomous trajectory risk forecasting, multi-view classification, and LLM calibration.

Critical strengths

(1. Motivation and novelty) The idea of separating class prediction and uncertainty quantification in EDL is interesting and timely. Moreover, to the best of my knowledge, this is the first work to incorporate conformal-rank-based supervision into the EDL framework. This offers a promising direction for addressing the existing limitations of EDL.

(2. Broad experimental evaluation) This paper provides experimental results beyond standard vision settings, including tumor referral prediction and trajectory risk forecasting, which demonstrate the broad applicability of the model.

Critical weaknesses

(1. Unclear justification for uncertainty supervision) This paper lacks a clear justification for using the ranks of nonconformity scores as ground-truth targets for uncertainty supervision. In particular, while nonconformity scores reflect predictive difficulty, it remains unclear why their ranks can be treated as valid supervision targets for uncertainty estimation, especially for epistemic uncertainty.

(2. Unclear justification for the evidential formulation) While CSEDL effectively addresses some limitations of conventional EDL by separating prediction and uncertainty estimation, it remains unclear why the method needs to be built upon EDL in the first place. To my understanding, the main contribution of the paper is conformal rank-based uncertainty supervision, which does not necessarily require an evidential formulation. Moreover, since prediction and uncertainty are largely learned separately, the necessity of the Dirichlet-based formulation is questionable.

Questions for authors

(1. Justification for uncertainty supervision) What is the theoretical or conceptual justification for using nonconformity ranks as targets for supervising uncertainty? Is it valid to understand that the nonconformity rank is a ground truth for uncertainty? If so, what notion of uncertainty (e.g., epistemic or aleatoric) does the conformal rank-based uncertainty score intend to capture?

(2. Necessity of the evidential formulation) It is not clear whether the evidential formulation in CSEDL is necessary. Conceptually, I’m confused about the necessity of the Dirichlet-based formulation if prediction and uncertainty are mostly separately supervised and learned. Empirically, I want the authors to compare CSEDL with a standard neural classifier (softmax-based) equipped with an additional uncertainty head supervised by the same conformal-rank targets and report the performance gain.

(3. OOD detection performance and experimental details) How can uncertainty supervision based on nonconformity ranks of in-distribution samples generalize to unseen OOD samples and lead to strong OOD detection performance? Moreover, could the authors provide more details on the experiments, including the backbone architecture and whether the baseline results are reproduced or taken from the respective papers?

(4. Self-referential uncertainty supervision) To my understanding, the nonconformity scores are derived from the model's own predictive probabilities, making the uncertainty supervision somewhat self-referential. It is unclear whether these model-derived targets provide additional information about uncertainty beyond the model's existing predictive confidence.

6 comments

Dear Reviewer ePzp,

Thanks again for your valuable comments. We provided a rebuttal to your review comments four days ago. Would you please kindly check whether your concerns are well addressed by those responses? Thanks.

Regards from the authors.

I thank the authors for their detailed responses and additional experiments. In particular, I appreciate the comparison with a softmax classifier equipped with an uncertainty head in Q2. The results provide useful empirical evidence supporting the benefit of the evidential formulation.

However, my major concern regarding the justification for nonconformity-rank-based uncertainty supervision remains unresolved. I would appreciate further clarification on the following points.

(1. Justification for uncertainty supervision — W1/Q1)

I understand the authors’ clarification that nonconformity ranks are not ground-truth uncertainty values but instead represent relative prediction difficulty under the model. However, this does not fully explain why these ranks constitute appropriate supervision targets for uncertainty estimation. In particular, why should learning relative prediction difficulty lead to improved uncertainty estimation, especially epistemic uncertainty? Prediction difficulty does not necessarily reflect epistemic uncertainty, as even well-supported ID samples can be difficult to classify due to inherent ambiguity. Therefore, the conceptual connection between the proposed supervision target and the uncertainty that CSEDL aims to estimate remains unclear.

(2. Role of rank supervision in epistemic uncertainty estimation — Q3)

A central goal of epistemic uncertainty estimation is to reflect the model’s lack of knowledge about inputs that are not sufficiently supported by the training data, and OOD detection is a common way to evaluate this capability.

However, the authors attribute CSEDL’s ability to assign high uncertainty to unseen OOD inputs primarily to the EDL framework rather than to nonconformity-rank supervision. This explanation seems difficult to reconcile with the central motivation of the paper, which is that conventional EDL lacks adequate uncertainty supervision and may assign excessive evidence to unreliable samples.

If the ability to assign high uncertainty to unseen OOD inputs primarily comes from the EDL framework itself, what exactly is the role of nonconformity-rank supervision in improving epistemic uncertainty estimation? More fundamentally, since the proposed supervision targets reflect relative prediction difficulty among ID training samples, it remains unclear why learning these targets should improve uncertainty estimation for unfamiliar OOD inputs. While the reported performance gains are encouraging, they do not fully explain how the proposed supervision contributes to this capability.

(3. Self-referential uncertainty supervision — Q4)

I understand the authors’ example showing that nonconformity scores distinguish confidently correct from confidently incorrect predictions, thereby providing information beyond maximum predictive confidence. However, this does not resolve my original concern about self-referential supervision.

The nonconformity scores are still derived from the model’s own predictive probabilities, and their ranks are subsequently used to supervise the same model’s uncertainty estimates. What ensures that this self-referential learning process provides meaningful uncertainty information rather than simply reinforcing the model’s existing predictions? In particular, once the model correctly predicts most training samples, what additional information does the rank-based supervision provide beyond the uncertainty or confidence already expressed by the model?

Overall, while the additional experiments strengthen the empirical evidence for CSEDL, I remain unconvinced that the central uncertainty supervision mechanism is sufficiently justified. My concern is not simply whether the proposed method improves downstream performance, but why supervising relative prediction difficulty should lead to meaningful epistemic uncertainty estimation.

We sincerely thank reviewer ePzp for continuing the discussion and for acknowledging the empirical benefit of the evidential formulation in our comparison with a softmax classifier equipped with an uncertainty head trained using the same rank targets. Building on the discussion above, we provide further clarifications about your concerns.

1. Justification for uncertainty supervision — W1/Q1

First, we would like to clarify that CSEDL learns evidential uncertainty, defined as u(x)=K/S(x)u(x)=K/S(x), and uses it to identify unreliable predictions. We do not claim that evidential uncertainty faithfully recovers epistemic uncertainty. Furthermore, previous studies have also identified limitations in interpreting the outputs of standard EDL objectives as faithful quantifications of epistemic uncertainty [1–3].

We agree that prediction difficulty can arise from inherent ambiguity in ID samples. For our objective of identifying unreliable predictions by evidential uncertainty, assigning higher uncertainty to such samples is appropriate. In summary, CSEDL estimates evidential uncertainty for identifying unreliable samples, rather than epistemic uncertainty. Our claim concerns the effectiveness of this supervision for learning evidential uncertainty and improving the identification of unreliable predictions, and it does not equate prediction difficulty with epistemic uncertainty.

2. Role of rank supervision in epistemic uncertainty estimation — Q3

The evidential formulation and rank supervision play complementary roles. EDL provides the Dirichlet parameterization linking evidence strength to uncertainty, while rank supervision provides a supervision signal for evidence allocation to improve this process. The usefulness of EDL for OOD detection is therefore consistent with the need to improve how its evidence strength is learned.

Specifically, difficult ID samples and unfamiliar OOD inputs may both exhibit incomplete or conflicting class-discriminative cues, providing insufficient support for a reliable prediction among the known classes. Rank supervision learns the relationship between sample reliability and the evidence within ID data, guiding the model to form a tendency to assign smaller evidence strengths to samples with weaker support for their true classes. At test time, the same association can apply to OOD inputs that exhibit these learned reliability cues, even when the classification head still favors a known class. This is the learning tendency we aim to encourage through rank supervision on ID samples to improve uncertainty estimation on OOD samples.

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

The paper proposes Conformity Supervised Evidential Deep Learning (CSEDL), a dual-head evidential learning framework. A direction head is supervised by ground-truth labels, while an uncertainty head is supervised by the empirical rank of a nonconformity score computed from the model’s predictive probability on the true class. The two heads are combined to reconstruct Dirichlet parameters. Experiments are conducted on OOD detection, tumor referral prediction, autonomous-driving trajectory risk forecasting, trusted multi-view classification, and LLM confidence calibration.

Critical strengths

1、The motivation is clear: conventional EDL can over-accumulate evidence on unreliable samples, and the paper attempts to provide an explicit sample-wise signal for uncertainty. 2、The dual-head parameterization is simple and easy to implement. The reconstruction formula is clean and preserves a valid Dirichlet parameterization. 3、The empirical evaluation covers several application domains, and the reported gains on OOD detection and PCam are non-trivial.

Critical weaknesses
  1. The relationship between empirical rank supervision and conformal prediction is insufficiently justified. The use of the term conformity supervised suggests a meaningful connection to conformal prediction. However, the proposed procedure appears to use empirical ranks computed from a FIFO cache of historical nonconformity scores, rather than a conventional conformal calibration procedure with an explicit coverage guarantee. This distinction is important. Standard conformal prediction provides statistical guarantees under specified assumptions, whereas learning an uncertainty predictor from empirical score ranks does not, by itself, establish finite-sample coverage or distribution-free validity. The cache also contains scores produced by models at potentially different training stages. Consequently, these scores may not be exchangeable or identically distributed with scores generated by the current model. If the score distribution changes during training, the resulting empirical ranks may be systematically biased or stale. The manuscript does not adequately establish whether CSEDL is intended to provide a formal conformal guarantee or merely to borrow the idea of ranking nonconformity scores as a training signal. 2、Rank uniformization does not establish better uncertainty estimation. The paper motivates rank-based supervision partly through the desirable properties of normalized ranks. However, a transformation of scores into empirical ranks does not automatically improve uncertainty estimation. 3、The experiments do not establish that the learned uncertainty generalizes across distinct failure modes
Questions for authors

Q1. What exactly is guaranteed by the conformity-based formulation? Does CSEDL provide a formal conformal coverage guarantee, or does it use empirical ranks only as auxiliary training targets? If a guarantee is claimed, what assumptions are required when the score cache contains outputs from different model states?

Q2. Why is rank supervision better than simpler alternatives? Does CSEDL outperform auxiliary supervision based on raw nonconformity scores, alternative score transformations, and conventional uncertainty targets under matched experimental conditions?

Q3. What does the predicted uncertainty measure across tasks? How do the authors distinguish classification difficulty, probability of error, distributional novelty, and operational risk? What evidence supports the claim that the same training principle is useful for these distinct objectives?

Q4. What is the isolated contribution of the dual-head architecture? How much improvement remains when the architecture is controlled for and only the rank-supervision component changes?

Q5. Is CSEDL competitive with standard calibration methods? In particular, how does it compare against temperature scaling and other established approaches when evaluated with the same backbone, data split, and model-selection budget?

Q6. How sensitive is the method to cache staleness? Can the authors quantify the difference between historical cache-derived ranks and ranks computed from current-model scores, and relate this difference to downstream performance?

Q7. Do the downstream results establish practical utility? For medical referral and trajectory risk, do the reported gains translate into improvements at fixed referral budgets or safety-relevant operating points, rather than only improvements in aggregate metrics?

3 comments

Dear Reviewer Wvrh,

Thanks again for your valuable comments. We provided a rebuttal to your review comments four days ago. Would you please kindly check whether your concerns are well addressed by those responses? Thanks.

Regards from the authors.

Thanks for your response.

Dear Reviewer Wvrh,

Thanks for your response. Please indicate whether your concerns have been fully resolved or if any remain. If all the concerns are solved, please consider raising your overall rating. If some concerns remain, we are more than happy to provide further clarification.

Your precise response will significantly improve the quality of this paper. Thanks again for your help.

Regards from the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.