acceptodds
Under review as a conference paper at ICLR 2027

Conditional Confidence Calibration for Foundation Model Decisions

Abstract

Foundation models increasingly serve as decision components, where discrete outputs require reliable confidence estimates. In safety applications, vision-language and language models make fine-grained judgments about image or text content. These multi-label tasks decompose naturally into binary decisions, each requiring a calibrated confidence estimate. The same need arises whenever a model score is used to accept, reject, or route an output. Standard post-hoc calibration typically learns a global or category-level mapping from model scores to confidence, potentially missing systematic reliability differences across input contexts. We propose conditional confidence calibration, which estimates the probability of a defined binary outcome from the model score, the decision being calibrated, and context available at decision time. Standard global and per-category calibration use only the score and, optionally, the decision category. Our adaptive method additionally fits a contextual calibrator using input descriptors such as text load, visual realism, or request domain, then learns from calibration data how much it should modify the better-supported decision-level estimate. The learned combination can default to that estimate when context is unhelpful, without updating the base foundation model. We evaluate across seven image, image-text, and text safety datasets and three non-safety datasets using Gemma 4 models, with E4B throughout and matched 12B evaluations on two image-text datasets. Across 14 safety comparisons, adaptive calibration improves Macro balanced accuracy over per-category confidence in 12, with mean relative gains of 5.3% on image and image-text tasks and 5.7% on text tasks, and lowers Brier score by 6.3% on average. On the non-safety datasets, it lowers Brier score by 3.2–12.1% and improves balanced accuracy by 2.0–2.8%. The gains also hold across model families: with Qwen3-VL and InternVL3.5 at 4B and 8B, adaptive calibration improves both metrics in all eight comparisons. These results demonstrate that adaptive combination improves both decision and probability quality over non-contextual confidence baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
confidence calibration, foundation models, uncertainty estimation, multimodal safety, attribute-guided calibration, adaptive calibration, vision-language models

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_wVM1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes post-hoc confidence estimation for frozen foundation models by adaptively combining decision-specific and context-conditioned probability estimates. The safety implementation uses a two-estimator mixture, while the non-safety experiments combine three experts. Across seven safety datasets, the method improves Brier score in all 14 primary comparisons and Macro balanced accuracy in 12. Additional evaluations cover other model families and prediction-correctness estimation on three non-safety benchmarks.

Critical strengths
  1. Broad evidence of improved probability quality: Adaptive calibration reduces Brier score in all 14 primary safety comparisons, with 12 significant improvements after Holm correction. Additional results across Qwen3-VL, InternVL3.5, and three non-safety tasks strengthen the evidence that the approach is useful beyond one backbone or application.

  2. Controls that meaningfully test the proposed explanation: The capacity-matched score-diagnostic, shuffled-context, and random-feature controls help distinguish useful context from additional estimator flexibility. Real visual context outperforms shuffled and random features in all six comparisons, with five significant improvements against each control. The authors also explicitly acknowledge weaker text results, making the scope of the evidence clearer.

Critical weaknesses
  1. The incremental value of adaptive mixing needs clearer evidence. The framework builds on established contextual calibration and learned estimator combinations. Constant candidates are included, and non-safety experiments compare constant and adaptive gates, but the reported results do not clearly quantify the benefit of example-dependent mixing over the contextual estimator alone or a tuned constant mixture. Explicit comparisons would help isolate the contribution of the gate.

  2. The added value of semantic text context is limited under the stronger control. In Table 20, real-context calibration beats the score-diagnostics-only gate in only 2/8 text comparisons, with no significant wins and a mean relative Brier reduction of −0.14%. Visual context has stronger support against shuffled and random features. Although the authors acknowledge this disparity, it limits how broadly the results support semantic conditioning as the source of improvement.

  3. The evidence for improved within-context calibration remains incomplete. Brier score measures both calibration and resolution, so its improvement alone does not establish reduced subgroup miscalibration. Figure 2 already provides aggregate reliability curves against a calibrated baseline. However, Tables 22–23 evaluate a separate ECE-targeted attribute method against raw scores. Subgroup comparisons of the main adaptive estimator against calibrated decision-specific probabilities would more directly test the conditional-calibration claim.

Questions for authors
  1. Could you report the selected contextual estimator alone, a tuned constant mixture, and the adaptive gate side by side? For the safety experiments, does constant-mixture selection include interior weights beyond (w=0,1)?

  2. What accuracy–cost tradeoff does model-assisted text attribute extraction offer relative to rule-based attributes and the score-diagnostics-only gate?

  3. How does the main Brier-selected adaptive estimator compare with calibrated per-category probabilities on subgroup calibration, particularly for low-support groups?

  4. Grounding-based confidence calibration (Padhi et al., 2025) [1] is relevant prior work on using cross-modal evidence to improve multimodal confidence estimates. Please clarify how the proposed context-conditioned calibration differs from grounding-based approaches like these.

[1] Padhi, Trilok, Ramneet Kaur, Adam D. Cobb, Manoj Acharya, Anirban Roy, Colin Samplawski, Brian Matejek, Alexander M. Berenbeim, Nathaniel D. Bastian, and Susmit Jha. "Calibrating uncertainty quantification of multi-modal llms using grounding." arXiv preprint arXiv:2505.03788 (2025).

Reviewer_YrGqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies confidence calibration for foundation-model decisions. The model first produces a score. The method then uses the score, the decision category, and the input context to produce two probability estimates: one based only on the score and one that also uses context. A gate learns how to combine the two estimates.The paper evaluates the method on safety datasets and non-safety tasks using Macro BA and Brier score. It also compares real context with shuffled context and random features.

Critical strengths
  • The paper does not compare only against raw scores. It includes Platt, beta, isotonic, and histogram calibration methods, and covers multiple models and tasks.
  • Appendix Table 20 controls for real context, shuffled context, and random features. The image-side context gains are relatively clear, while the paper also reports weaker text-side results.
Critical weaknesses
  • The methodological novelty is incremental. The main method learns a mixture weight between two calibration estimates. Its independent value over a direct context-aware calibrator with the same information and capacity is not sufficiently established.
  • An average improvement does not imply reliable behavior everywhere. A lower Brier score indicates better average probability quality, but does not guarantee better predictions for every sample or subgroup. In Table 20, real text context performs slightly worse than the score-diagnostics-only gate on average, with a relative Brier change of -0.14%. In addition, the image setting does not impose the same hard fallback for unseen contexts as the text setting.
Questions for authors
  1. Could you compare the proposed gate with a direct context-aware calibrator that uses the same context features, model capacity, and calibration data? This would show whether the gain comes from the gate or simply from the additional features.
  2. Please clarify the scope of the text-context claim. Since real text context does not improve over the score-diagnostics-only gate on average, does the simpler score-diagnostics approach retain most of the benefit?
  3. Could you report the performance of the main method on unseen or sparsely supported contexts relative to the strongest score-only calibration baseline, and explain whether the image fallback mechanism actually limits degradation? Results for an ECE-specific variant against raw scores should not replace this comparison.
Reviewer_zW8LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies confidence calibration for foundation model decisions by conditioning confidence estimates on model scores, decision categories, and input context. The proposed adaptive calibration method combines a decision-level calibrator with a contextual calibrator through a learned gating mechanism. The authors evaluate the method across safety and non-safety tasks using multiple foundation model families and report improvements in both decision quality and probability estimation.

Critical strengths
  1. The problem is relevant, particularly for safety-related applications where confidence estimates are used to make downstream decisions. Conditioning calibration on input context is a reasonable direction.

  2. The experimental evaluation is fairly comprehensive, covering several datasets, modalities, and model families.

Critical weaknesses
  1. The methodological novelty appears somewhat limited. Conditional calibration and input-dependent confidence estimation have been extensively studied. The proposed method combines a decision-level and a contextual calibrator using a learned gate, which is a reasonable but relatively straightforward extension. The paper would benefit from stronger comparisons with existing input-dependent calibration methods.

  2. The contribution of semantic context is not consistently supported by the text experiments. Appendix Table 20 shows that real-context calibration outperforms the score-diagnostics-only gate in only 2 of 8 text comparisons, with a slightly negative average Brier improvement. This is an important result because it suggests that the gains over conventional per-category calibration may come primarily from using a more expressive calibration model rather than from the semantic attributes themselves. While the authors acknowledge this limitation, the paper could distinguish more clearly between improvements attributable to contextual information and those arising from additional calibration flexibility.

  3. The generalization of learned contextual calibration deserves further investigation. The improvement on SPA-VL official test is relatively small. Adaptive calibration achieves 0.704 Macro BA compared with 0.701 for per-category calibration, whereas considerably larger gains are observed on some other datasets. Since this is the strictest transfer setting considered in the paper, it raises questions about whether the learned contextual corrections remain useful when the calibration and evaluation distributions differ.

  4. The distinction between confidence calibration and decision improvement could be clearer. The paper uses Macro BA as its primary decision metric and Brier score to assess probability quality. However, improvements in balanced accuracy may also arise from threshold optimization, while Brier score measures both reliability and discrimination.

Questions for authors
  1. How does the adaptive combination compare with a single contextual calibration model or input-dependent temperature scaling using identical features and model capacity?

  2. Given that semantic text attributes do not consistently outperform score-only diagnostics, what is the main source of the observed improvement on text tasks?

  3. Why are the improvements much smaller on SPA-VL official test than on the validation holdout? How sensitive is the learned gate to changes in the context distribution?

  4. Can the authors further isolate the contribution of probability calibration from threshold optimization to the reported Macro BA improvements?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.