acceptodds
Under review as a conference paper at ICLR 2027

Error Detectability in Model Merging: Diagnosis and Constrained Repair

Abstract

Model merging can preserve a classifier's decisions while degrading the confidence ranking used to detect its errors. We isolate this effect on expert–merge agreement subsets, where predictions, and hence correctness labels, coincide: on the five CLIP ViT-B/32 tasks with at least ten shared errors, expert macro AUROC on these subsets is 86.4%, versus 69.2% for TIES. A three-class linear construction shows that parameter averaging can reverse a perfect error ranking while preserving every prediction, through an example-dependent Jensen gap in the competing logits. We then introduce PED (Preserving Error Detectability), a local task-vector correction that optimizes error ranking under margin constraints on protected correct fitting examples and retains only candidates whose macro fixed-label fitting AUROC is at least that of the base. With 128 fitting and 128 selection labels per task, PED raises eight-task held-out AUROC from 84.7% to 88.2% and accuracy from 71.7% to 73.4%, exceeding LMN and ProbSurgery by 2.8 and 2.3 points in direct paired comparisons. On ViT-L/14, it improves AUROC by 3.9 points on TIES and by 4.2 points on task arithmetic. With correctness labels held fixed, PED recovers 72.1% of the expert–TIES ranking gap on shared ViT-B/32 decisions, so its gains include better ranking of the base model's own errors.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Model Merging, Misclassification Detection, Confidence Ranking

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 4 reviews, above 85% of the others.
Reviewer_7hakOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether model merging preserves error detectability (the correctness AUROC of confidence scores), not only accuracy. To separate confidence changes from label changes, it compares experts and merged models on agreement subsets, where predictions and correctness labels coincide. On five CLIP ViT-B/32 tasks, experts reach 86.4% macro agreement-subset AUROC versus 69.2% for TIES. A three-class linear construction shows that parameter averaging can fully reverse a perfect error ranking without changing any prediction, through an example-dependent Jensen gap in the competing-class log-sum-exp. The paper then proposes PED, a post-merge correction in a 24-dimensional task-vector subspace. PED minimizes a pairwise ranking surrogate under margin floors on correctly predicted fitting examples. It keeps only candidates whose macro fixed-label fitting AUROC is at least the base’s. With 128 fitting and 128 selection labels per task, PED raises TIES' AUROC from 84.7% to 88.2% and accuracy by 1.7 points, outperforming LMN and ProbSurgery. It also recovers 72.1% of the expert–TIES gap on fixed base decisions and gains 3.9–4.2 AUROC points on ViT-L/14.

Critical strengths
  1. The agreement-subset evaluation separates changes in confidence ranking from changes in which examples are classified correctly. The three-class setup takes this further by showing that rankings can change even if all predictions stay the same. This is an important idea, since typical model-merging evaluations focus only on accuracy and might miss this type of issue.
  2. PED outperforms TIES in both AUROC and accuracy, and the correction baselines used in the main experiments. These improvements remain consistent on ViT-L/14 and with task arithmetic. The fixed-label experiment shows that the gains are not just due to different predictions. Since the correction is built directly into the merged weights, there is no need for an inference-time module.
Critical weaknesses
  1. The experiments cover eight image-classification tasks, two CLIP backbones, and two merging methods. It remains unclear how well the method generalizes to other architectures, domains, or tasks. More importantly, fitting and selection examples are drawn from the official held-out evaluation data, although they are kept disjoint from the audit examples. This is a valid separated evaluation protocol, but it leaves open how practical PED would be when labeled examples from the target deployment distribution are scarce or unavailable.
  2. The evidence does not fully isolate the source of PED’s gains. The controls suggest that the complete constrained optimization procedure is effective, but the paper acknowledges that the separate contributions of the optimizer, margin constraints, and eligibility test are not identified. In addition, although the linear analysis identifies a precise mechanism for ranking reversal, the nonlinear decomposition includes a residual whose empirical role remains unresolved. These limitations weaken the mechanistic explanation for the observed improvements, though they do not invalidate the reported results.
Questions for authors
  1. Since fitting and selection labels are drawn from the official test split, can the authors clarify how the method would be deployed when target-task labels are available only from a separate validation set? Do the gains persist when adaptation uses a distinct training or validation source and the final audit comes from a separate evaluation split?
  2. Can the authors isolate the contributions of the protected-margin constraints and the fitting-AUROC eligibility test from the choice of optimizer? In particular, how much of the improvement persists when ranking optimization, preservation constraints, and candidate selection are changed individually while keeping the coefficient space and data budget fixed?
  3. Can the authors quantify how much of the observed ranking degradation in the nonlinear models is associated with the Jensen-gap term versus the residual between weight merging and the reference logit ensemble? This would help establish how directly the linear explanation accounts for the empirical phenomenon.
  4. (Not affecting my rating) Is there a reason why AdaMerging is not included in the reporting?

My rating is most likely to be affected by a strong answer to question 2.

Reviewer_VcF2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies whether model merging preserves not just predictions but the confidence ranking used to detect errors (correctness AUROC). One of the contributions is an agreement-subset evaluation that compares expert and merged AUROC on examples where predictions. Another one, Prop.1 showing that parameter averaging can reverse a perfect error ranking while preserving every prediction, via an example-dependent Jensen gap in the competing logits, it's an interesting finding.

Critical strengths

Prop. 1 is hand-verifiable, the ≥3-class requirement is a sharp observation, and Eq. (7) cleanly separates averaged expert log-odds from the Jensen offset. The fixed-label evaluation (anchoring base class and base correctness labels) is a genuinely good idea and supports the claim that part of the gain is better separation of existing decisions, not relabeling.

Critical weaknesses
  1. PED's only formal guarantee (Prop. 2) is on fitting examples. More specifically, that initially correct fitting predictions stay correct and fitting macro AUROC does not decrease. But the headline claims are held-out AUROC and accuracy gains, and the authors themselves note "aggregate accuracy gains do not measure retention of every base decision" (Section 7). It sounds like a overstatement based on the title and method name.

  2. The ablation does not isolate PED's components (Section 6.2) but it still seems to attribute the gain to the combination? Perhaps, the authors could think of doing ablation on ranking loss+margin constraints+eligibility gate+forward-check acceptance.

  3. Eight image-classification tasks, two CLIP backbones, fixed expert checkpoints, known task identity at inference, fixed zero-shot text heads. The last two are significant simplifying assumptions: real deployment rarely knows task identity with certainty, and the "task heads are frozen" assumption is doing real work in the derivations (the anchor h_i and the fixed label set). Are these tasks representative of the merging literature or is the confidence-ranking effect an artifact of CLIP's particular logit geometry?

  4. Minor: I have a few concerns on some key objects are introduced without enough scaffolding for a general reader. This is partly my own background (I don't work in VLMs), I think is also a presentation issue.

Questions for authors
  1. As mentioned earlier, have you done a factorial ablation isolating (a) ranking loss only, (b) + margin constraints, (c) + forward-pass acceptance, (d) + eligibility gate? Given that Rank-L (+0.2) and SoftSafe-L (+0.5) are far below PED (+3.5), which component accounts for the jump?

  2. Table 3 reports 72.1% recovery at d=24, 128 labels. How does recovery scale with d=40 and 384 labels? Is the residual 4.8-point gap to the expert reducible by more capacity/labels, or is it a floor?

  3. Is the expert–TIES ranking gap specific to CLIP zero-shot heads, or does it appear with trained heads and in language-model merging?

Reviewer_M9jKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether model merging Maintain the confidence ranking required to differentiate correct predictions from errors. It evaluates experts and merged models on agreement subsets, where predictions and correctness labels coincide. A three-class linear construction showcases that averaging can reverse perfect error ranking without changing any prediction, through an example-dependent Jensen gap in competing logits. The proposed PED method optimizes a pairwise ranking objective in a local task-vector subspace. It verifies margin constraints on initially correct fitting examples and applies a separate fitting-AUROC eligibility test. Experiments cover eight classification tasks, two CLIP backbones, and TIES and task-arithmetic merge bases, with both own-prediction and fixed-label evaluations.

Critical strengths
  1. a convincing diagnostic contribution. On the five eligible expert–TIES agreement subsets, expert AUROC is 86.4 compared with 69.2 for TIES Table 8. These comparisons hold instances and correctness labels fixed. Proposition 1 supplies a correct construction showing complete ranking reversal despite unchanged predictions. Together, these results demonstrate a meaningful limitation of evaluating merged models primarily through classification accuracy.

  2. consistent evidence for the complete repair procedure. PED improves ViT-B/32 TIES macro AUROC from 84.7 to 88.2 and accuracy from 71.7% to 73.4% Table 1. Its AUROC point estimate improves on every task Table 2. Direct paired comparisons favor PED over LMN by 2.8 points and ProbSurgery by 2.3 points, with intervals excluding zero Table 9. Gains extend to ViT-L/14 and task arithmetic. Table 3 additionally demonstrates improved ranking of fixed base decisions, addressing the concern that gains could arise entirely from changes in which examples are counted as errors.

Critical weaknesses

the distinctive constrained-repair components are not individually validated. Rank-L uses Adam, whereas PED changes the proposal optimizer, feasibility checks, acceptance procedure, and eligibility filtering. SoftSafe-L also changes the optimization formulation and has only one reported adaptation repeat. These comparisons establish the effectiveness of the complete procedures but leave the contribution of the margin constraints and AUROC gate unresolved. Section 6.2 acknowledges this explicitly.

this matters because the guarantees concern fitting examples, while the practical benefit of protecting predictions should be assessed on held-out data. Aggregate accuracy gains do not reveal how often initially correct audit predictions become incorrect. A controlled ablation with decision-retention measurements would clarify whether the safeguards provide a generalization benefit.

Questions for authors

Request for evidence: On the principal ViT-B/32 TIES setting, can you provide matched variants using the same local optimizer, fitting splits, selection rule, and comparable candidate budgets, while separately removing the protected-margin constraints and fitting-AUROC eligibility test? Please report held-out AUROC, accuracy, and the fraction of initially correct audit predictions that become incorrect. This would directly address the unresolved contribution of the safeguards.

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

The paper studies the degradation of error detectability after model merging. Even when predictions remain unchanged, confidence may become less effective at separating correct predictions from errors. The authors use agreement-subset evaluation to diagnose this issue and explain ranking reversal in linear models through the competing-logit Jensen gap. PED optimizes ranking in a task-vector subspace while imposing margin constraints on protected correct fitting examples.

Critical strengths

The paper distinguishes prediction accuracy from error detectability, addressing a practically relevant problem. Agreement-subset evaluation with correctness labels held fixed helps separate changes in confidence ranking from changes in which examples are errors. The linear construction also clearly demonstrates that confidence ranking can reverse despite unchanged predictions.

Critical weaknesses
  1. The separate benefits of the components are not sufficiently isolated. PED combines margin constraints, the proposal in Equation (13), forward checks, and a fitting-AUROC eligibility test. Rank-L and PED differ in both optimization and candidate acceptance, so the comparisons mainly support the complete procedure. It remains unclear whether simpler ranking optimization with the same eligibility test would be sufficient. Forward differences and repeated forward checks also add training costs, but actual running times are not reported.

  2. The gains over directly distilling expert outputs remain unclear. The experiments do not sufficiently establish how much additional benefit the ranking objective and margin constraints provide compared with directly matching expert outputs.

Questions for authors
  1. Could the authors include a distillation comparison in the same coefficient space, with comparable data, expert access, optimization budgets, and selection rules, to better distinguish the contributions of the ranking objective, margin constraints, and optimization procedure?

  2. Could the authors report the proportion of initially correct predictions that remain correct on the audit examples, together with the proportions that change from correct to incorrect and from incorrect to correct? To further assess practical value, are fewer errors missed when the same proportion of predictions is sent for human review? At the same coverage, is the error rate among retained predictions lower?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.