acceptodds
Under review as a conference paper at ICLR 2027

If Inputs Conflict, When Does Learning Help?

Abstract

A clear image and a clear audio clip can tell different stories. If one of them belongs to another example, confidence alone cannot tell a multimodal model which story to follow. Learning a correction is an appealing response, but when does it help? We answer this question with a controlled replacement protocol and one default-preserving method, ConflictCalibrated Evidential Fusion (CEF). CEF keeps the average of source logits as a transparent default, proposes alternative predictions, and accepts one only when a cross-fitted selector predicts a rescue rather than a harm. The method is a measuring instrument as much as a competitor: it lets us score, example by example, when overriding a strong default pays. Three findings trace its benefits and limits. When three of five PolyMNIST views remain target-consistent, coordinate-wise median reaches 97.3% on the official test split, outperforming frozen CEF at 93.6%. With two conflicting sources there is no majority to follow: on AV–MNIST, CEF reaches 54.4% donor accuracy on validation against 48.1% for the best simple rule, and on a reserved test split it raises the default from 48.1% to 55.8% while clean accuracy stays at 96.8%. Finally, a correct answer can be available yet out of reach: with the candidate pool fixed, a selector trained on two replacements reaches only 6.4% on three despite a 99.8% candidate oracle, and retraining on three raises it to 84.3% at a cost of 1.3 points in clean accuracy. The value of learned correction therefore depends on the remaining redundancy, the availability of a correct candidate, and whether the selector generalizes to the corruption regime.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
multimodal fusion, robust aggregation, selective prediction, modality corruption, trusted multi-view learning, conflict-aware fusion

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

This work studies 'confident source conflict': some modalities are replaced by clean observations of another example, so neither confidence nor input quality shows the corruption. To solve the problem, the authors proposed CEF, which keeps average-logit fusion as default, builds candidates (bounded residual or 31 source-subsets), and routes each example by a cross-fitted selector trained on 'candidate correct but default wrong', with a preservation-constrained threshold. On PolyMNIST and AV-MNIST, median wins when a majority remains, routing helps when no majority exists, and the selector transfers poorly across corruption counts.

Critical strengths
  1. The authors are honest about CEF's limitation: it reports its own method losing to a zero-parameter median (97.3% vs 93.6%, Table 1B) and makes this the main finding.

  2. The evaluation is well-designed: cross-fitting (Alg. C.1), a reserved AV-MNIST test used once, and a donor-assignment sensitivity check.

  3. The rescue/harm decomposition (Fig. 2c, Table 3) clearly shows routing, not the correction, gives +6.1 points.

Critical weaknesses
  1. There is no theory on when routing provably keeps clean accuracy, the authors provided only empirical evidence. I look forward to the author providing deeper theoretical analysis and guidance.

  2. The AV-MNIST headline (+7.7pp, CI [5.1,10.2]) is on a 400-example test split. The CI is a t-interval over 5 seeds (Sec. 5.4) and ignores binomial noise, therefore I suspect the precision is overstated.

Questions for authors
  1. Does the AV-MNIST CI consider the binomial variance on 400 examples, for example a paired per-example test?

  2. Does CEF beat median-as-default routing at k=2 and k=3?

  3. Please report the Eq. 7 baseline and the trainable fusion results in detail. Besides, in Eq. 7, how is the conflict c_m defined, and what value is kappa?

1 comment

I have carefully read the authors' responses to the questions raised by me, and I appreciate the detailed information provided, especially the theoretical analysis part. I will keep my rating unchanged.

Reviewer_Pu7COpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multimodal prediction when some inputs are replaced by observations from another class, creating confident disagreement between sources. It proposes Conflict-Calibrated Evidential Fusion (CEF), which keeps source encoders frozen, uses averaged logits as a default, and constructs alternatives through residual corrections or source subsets. A learned selector decides when to override the default, with a threshold calibrated to limit performance loss on clean and other specified conditions. Experiments on PolyMNIST and AV–MNIST show that selective correction can improve recovery, while simple median fusion remains stronger when most sources are correct.

Critical strengths
  • The paper examines how individually confident sources can disagree because some inputs belong to a different example. The controlled replacement protocol provides a useful setting for studying this problem.
  • The ablation separates candidate quality from routing quality, showing that selective correction can outperform unconditional correction. The paper also reports negative results, including the superiority of median fusion in the majority-consistent setting and poor transfer across corruption regimes
Critical weaknesses
  • The methodological novelty is insufficiently established. CEF’s central mechanism selects between a default and an alternative using comparative outcomes. This is closely related to existing learning-to-defer formulations, which already accommodate fallible alternatives. The paper’s claimed distinction based on candidate fallibility is therefore unconvincing. The specific combination of candidate construction and preservation-constrained admission may be useful, but its contribution beyond established routing approaches needs clearer justification. Refer to ``Mozannar, H.; Sontag, D.. (2020). Consistent Estimators for Learning to Defer to an Expert. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research''. Meanwhile, the related-work discussion weakens the novelty argument. It omits ECML, which explicitly addresses conflicting views, and understates DBF, which directly measures confident disagreement. These methods require a substantive comparison that distinguishes conflict detection, uncertainty estimation, and target recovery.
  • The evaluation lacks controlled comparisons with the closest alternatives. Section 5.3 excludes DBF and other trainable fusion methods from the main comparisons because their training and model-selection protocols were not matched. This leaves the central comparative question unresolved. The claim that the heuristic discounting baseline in Eq. (7) upper-bounds conflict-aware weighting is also unsupported.
Questions for authors
  • What does CEF’s default-preserving admission mechanism add beyond existing learning-to-defer approaches, which already allow fallible alternatives? Please clarify which part of the formulation constitutes the methodological contribution.
  • Can the authors compare CEF with ECML and DBF, and assess what is actually new in CEF’s default-preserving admission mechanism? Can you provide controlled comparisons with ECML and DBF using matched source representations, donor-training data, and tuning budgets, with any necessary adaptations documented? Section 5.3 leaves these comparisons unresolved, and Eq. (7) does not establish an upper bound on conflict-aware fusion.
  • What benefit does the preservation constraint provide beyond selective routing itself? Since it never binds in the calibration experiment reported in Sec 6.4, could you compare the same candidates and scorer with and without this constraint, reporting donor recovery and clean-data degradation?
Reviewer_huPiOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles the problem of learning from multimodal signals when there could be replacement of the correct data with incorrect ones from another label. The proposed method, Conflict Calibrated Evidential Fusion (CEF), scores the example to make a decision about overriding the default decision. Findings regarding when the majority view is consistent and when there is no majority to follow are obtained.

Critical strengths

Performance metrics (accuracy) along with calibration metrics (NLL, Brier score) are used in the experiment.

Critical weaknesses
  • The paper is hard to follow across multiple regions.

    • Throughout the introduction, the paper directly jumps to starting with dataset, when readers do not understand what the concept being explained is.

    • It’s not clear what’s being trained.

    • Most methodology components are discussed without introducing what they mean (example J) and are sometimes discussed without rationale.

  • How realistic is the setting of the paper (donor replacement)? It is suggested to make the research motivation clear. The connection with multiview is also not correct, as the replacement changes the semantic meaning of that label rather than providing a different view (as shown in Figure 1a for label 3 using images of 3 vs. 7 is a completely different semantic concept than a different view).

  • The proposed method does not exhibit superior performance.

  • As the authors mention, the term “evidential” can create confusion. It might be more meaningful to use the term that accurately describes the methodology.

Questions for authors

Please see weakness section

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.