acceptodds
Under review as a conference paper at ICLR 2027

Agreement Is Not Confidence: Trajectory Agreement Degrades as Training Converges

Abstract

When a model's logits were never logged, a cheap substitute for confidence is how often the checkpoints saved during training agree with the final prediction. The same signal drives a published selective-prediction method that weights late disagreement most heavily. We show that this signal carries less information the closer the checkpoints are to the end of training. Two elementary bounds make this precise. The calibration error of trajectory agreement is pinned to within δ_W of the model's error rate, where δ_W is the mean disagreement in the checkpoint window. Its AUROC for error detection is at most one half plus the largest class-conditional probability that some checkpoint in the window disagrees with the final prediction. As the window approaches the final checkpoint, agreement tends to the constant 1: its calibration error tends to the error rate and its AUROC to one half. On 3D visual grounding (ScanRefer and Nr3D, four training runs) and CIFAR-10 with ResNet-18 (three seeds), moving a fixed-size window later raises calibration error and lowers AUROC in 13 of 14 run-by-window-size cells, with bootstrap intervals excluding zero; in the last-checkpoint window, 89–94% of errors receive agreement 1. Temperature-scaled softmax and a three-seed deep ensemble on the same rows are calibrated (ECE 0.031 and 0.041 against 0.228 for agreement, on ScanRefer), so the failure belongs to the construction, not to the model. For the published method, late weighting helps where accuracy keeps improving and hurts where it stalls, and its late limit falls to AUROC 0.51–0.54 in every run.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
calibration, selective prediction, training dynamics, uncertainty estimation, checkpoint agreement, 3D visual grounding

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

This paper analyzes "trajectory agreement" — the fraction of training checkpoints whose prediction agrees with the final model's prediction — as a logit-free confidence proxy. The same signal underlies the published selective-prediction method SPTD (Rabanser et al., 2025, TMLR), which up-weights late disagreement by (t/T)k(t/T)^k. The authors argue that as training converges, checkpoints agree more and more, so agreement saturates at 1 and stops carrying error information ("convergence-induced saturation"). Two bounds make this precise: (i) the calibration error of agreement is pinned to within Wˉ\bar{W} of the model's error rate qq, i.e. ∣q−Wˉ∣≤CE(AW)≤q+Wˉ|q-\bar{W}| \le \mathrm{CE}(A_W) \le q+\bar{W}; (ii) AUROC(S)≤1/2+TV(P0,P1)\mathrm{AUROC}(S) \le 1/2 + \mathrm{TV}(P_0, P_1), and when the score is constant on the window-agreement event EWE_W, AUROC≤1/2+max⁡yPr⁡(EWc∣Y=y)≤1/2+KWˉ/min⁡(q,1−q)\mathrm{AUROC} \le 1/2 + \max_y \Pr(E_W^{c} \mid Y=y) \le 1/2 + K\bar{W}/\min(q, 1-q). A corollary gives the late limit: Wˉ→0\bar{W} \to 0 implies CE→q\mathrm{CE} \to q and AUROC→1/2\mathrm{AUROC} \to 1/2. Experiments use 3D visual grounding (ScanRefer, 3 seeds; Nr3D, 1 run; a 3.8M-parameter GeoClassifier over frozen Uni3D + DistilBERT features with ground-truth proposals, T=10T=10) and CIFAR-10 / ResNet-18 (3 seeds, T=80T=80). Key results: moving a fixed-size window later raises CE and lowers AUROC in 13 of 14 run-by-window-size cells; in the last-checkpoint window 89-94% of errors receive agreement 1; temperature scaling and a 3-seed ensemble are calibrated on the same predictions while agreement is badly overconfident (ECE 0.228 vs 0.031 on ScanRefer); SPTD's late weighting helps on CIFAR-10 (+0.016 to +0.020) but hurts on 3D (-0.002 to -0.016). A secondary finding is that agreement fabricates a subgroup gap.

Critical strengths
  1. Honest, reproducible execution: three seeds agree closely; scene-grouped and paired bootstraps are used; the single non-significant cell is flagged; weak "four-group" evidence is reported as descriptive only; the supplement ships scripts with unit tests and JSON number sources.
Critical weaknesses
  1. The central empirical comparison is confounded by checkpoint density. SPTD's original experiments used T∈[800,2200]T \in [800, 2200], but this paper uses T=10T=10 (3D) and T=80T=80 (CIFAR-10). SPTD's weights (t/T)k(t/T)^k depend on trajectory resolution. The authors' own thinning study (Table 7) shows the effect size moves with density — the CIFAR-10 k=2k{=}2 gain shrinks from +0.016..+0.020 to +0.007..+0.008 when thinned to T=10T=10. By symmetry, the 3D degradation observed at T=10T=10 (-0.002 to -0.016) may also change (vanish or reverse) at realistic density, and the authors state denser 3D trajectories were not stored. So the headline claim "SPTD hurts on 3D grounding" is not established at the density the method was designed for.

  2. Mismatch between the object analyzed (fixed-final agreement) and the marketed target (SPTD). The genuine failure is the k→∞k\to\infty limit, which nobody uses; at k=2k=2 (the original setting) SPTD improves AUROC on CIFAR-10, and the 3D degradation is small (≤0.01\le 0.01). The authors themselves write that their results "qualify rather than contradict" the method, yet the abstract and introduction are framed adversarially. Moreover, Proposition 2 only covers the case where the score is constant on EWE_W; it does not directly bound SPTD's actual score gkg_k, so the theory aligns with the method only at k=0k=0 and k=∞k=\infty.

Questions for authors
  1. (Key) Can you store denser 3D trajectories (e.g. T≥100T \ge 100), or run SPTD on CIFAR-10 near the original T≈800T \approx 800, to test whether the 3D AUROC degradation persists at realistic density?
  2. If SPTD's weighted-disagreement score gkg_k is recalibrated (Platt/isotonic), does it rival softmax? If so, the finding is better stated as "the raw proxy needs recalibration".
  3. Please add the within-trajectory probability-averaging baseline (average the softmax over the same TT checkpoints, report ECE/AUROC). Does it also saturate, or is saturation specific to agreement?
  4. Can Proposition 2 be extended from "constant on EWE_W" to the general weighted gkg_k, and can you give a matching lower bound to establish tightness?
  5. How does the picture change with non-ground-truth proposals (a real detector front end) and at lower error rates / larger models — does saturation still scale with qq?
  6. Can you add a small LLM experiment (e.g. self-consistency agreement across samples vs training-trajectory agreement)?
Reviewer_HHbqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper analyzes the limitations of using agreement between training checkpoints and the final model as a confidence proxy. According to the authors, agreement saturates as the checkpoint predictions converge.

Critical strengths
  • The paper is clear to read through.
Critical weaknesses
  • Weakness of the Premise . The paper does not sufficiently distinguish agreement as an error-ranking proxy from agreement as a calibrated probability of correctness. The paper criticizes the use of agreement as a confidence proxy, but does not provide sufficient evidence that prior work interprets raw agreement as a calibrated probability of correctness. The paper itself has to justify that the agreement is commonly used as a substitute to assess confidence, but does not.

  • Lack of Impact. : The main argument is largely trivial. Even if the main argument may be true, the analysis and possible impact on the field is too marginal.

  • Lack of Experimental Validation. In line 51, the paper argues that " We call this convergence-induced saturation and study it as a property of the construction, independent of any particular model." However, the experiments are done on two domains, and for CIFAR-10, done only on ResNet18. The results may differ on the domain, training method or procedure(optimizer, learning rate, self-supervised, etc), and different calibration metrics. For example, the author may conduct experiment on different size of the training set on CIFAR-10 or ImageNet and validate their statement, where the trained checkpoint will naturally achieve lower performance on smaller size.

Questions for authors

See weaknesses.

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

This paper studies whether the agreement between a window of training checkpoints and the final prediction can serve as a confidence score. The main contribution is two bounds. The first pins the calibration error of agreement to the model’s error rate, up to the window’s mean disagreement. The second caps the AUROC for error detection by the disagreement mass in the window. The theory is easy to follow, and the bounds are tight against the empirical results, so theory and experiments align well. The experiments are relatively simple, and the calibration of agreement remains far worse than temperature scaling (e.g., ECE 0.228 vs. 0.031 on ScanRefer).The problem is new and interesting: it asks whether checkpoint agreement actually reflects the model’s confidence in its prediction.

Critical strengths
  1. A clear, correct diagnosis of a published method’s failure mode. The paper isolates when late-checkpoint weighting in SPTD helps (accuracy still improving, CIFAR-10) and when it hurts (accuracy plateaued, 3D grounding), with paired bootstrap intervals across seven runs and a thinning ablation. Fig. 2b, showing that correct predictions lock in early while errors keep switching, explains the mechanism.

  2. Clean controls. Temperature scaling and the seed ensemble on identical predictions show the miscalibration belongs to the construction, not the model.

Critical weaknesses
  1. Limited theoretical and practical contribution. Both propositions are elementary and generic: the lower bound in Prop. 1 is the calibration-in-the-large gap, which holds for any score, and Corollary 1 states that a score converging to a constant has CE =q=q and AUROC =12=\tfrac12. The reported “tightness” on the last-checkpoint window is automatic because both bounds collapse to qq as δW→0\delta_W\to0. The main conclusion, that agreement saturates as checkpoints converge, follows from the definition of the score. The practical message is also narrow: on CIFAR-10 the late-window cost is 0.011–0.021 ECE, so the failure matters only on low-accuracy tasks, and where logits exist, temperature scaling is better and needs one checkpoint.

  2. The practical case against agreement is weaker than stated, and the setup limits generality. Table 4 shows that global isotonic recalibration brings agreement to ECE 0.009, better than temperature scaling; the irreducible deficit is discrimination (AUROC 0.764 vs 0.795), which the paper does not foreground. The construction is also transductive: test inputs must be available during training or all K checkpoints must be stored and re-run. The 3D results come from one model family with ground-truth proposals, and trajectories have only T=10 checkpoints, far below SPTD’s original 800–2,200.

Questions for authors

Clarifications (could change my assessment)

  1. Can you report SPTD at a density closer to the original (T in the hundreds) on at least one 3D task? Table 7 suggests denser sampling enlarges the 3D loss, which would strengthen your claim.
  2. For the three ScanRefer seeds, what is the cross-seed agreement separately on queries the final models get right vs. wrong? Please also report the seed-ensemble control on CIFAR-10, where three seeds already exist.
  3. Please report AUROC and AURC for recalibrated agreement (Table 4) across all runs, and state explicitly that recalibration fixes ECE but not ranking.
  4. What are the inference-time and storage costs for new test inputs not seen during training (K+1 passes, K checkpoints), compared with softmax, temperature scaling and the 3-seed ensemble?

Suggestions (future work)

  1. Relate the construction to weight averaging along the trajectory (Izmailov et al., 2018, SWA), which the paper does not cite.
1 comment

Thank you for the thorough response and the revisions.

My questions are addressed or scoped by authors' rebuttal. However, my remaining reservation is that the contribution is an empirical diagnosis of one published method’s failure regime.

Thus, I will keep my original rating. I won't argue for a rejection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.