acceptodds
Under review as a conference paper at ICLR 2027

Confidence Needs Evidence: Reliable Confidence Estimation for Multimodal Models

Abstract

Reliable deployment of multimodal models requires confidence estimates that distinguish trustworthy predictions from those that should be rejected. However, false confidence—the assignment of high confidence to incorrect predictions—remains difficult to mitigate. Our experiments show that many false-confidence cases in multimodal question answering arise from language priors without corresponding task evidence, indicating that confidence is unreliable when it lacks evidence support. Motivated by this finding, we propose , a simple, training-free confidence estimator. With only two forward passes, measures evidence support through the full-versus-blind probability lift of the fixed full-input prediction and combines this signal with normalized predictive-entropy confidence to produce an evidence-weighted confidence score. We evaluate with multiple models on four benchmarks covering video QA and joint audio–video QA. consistently outperforms state-of-the-art baselines and achieves the highest correctness AUROC in all 16 benchmark–model settings, demonstrating its reliability and generality for multimodal confidence estimation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Confidence Estimation, Multimodal Question Answering, Evidence Support, Video Understanding, Audio-Visual Understanding

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_15CtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors argue that existing confidence estimators in MLLMs, which rely on output distributions, repeated sampling or hidden representations, may inherit language biases from the models, resulting in high-confidence predictions that are not adequately supported by multimodal evidence. To address this issue, they propose EviConf, a training-free confidence estimation method. Specifically, EviConf computes internal confidence using normalised predictive entropy while fixing the answer selected from the multimodal input. It then measures content evidence support by calculating the difference in probability between the full input and a question-only input, and temporal evidence support by determining the proportion of K temporal shuffles that decrease the probability of the selected answer. The product of these two signals is mapped to the range [0, 1] to obtain an evidence score, which is then combined with the internal confidence score using a geometric mean. Experiments across four benchmarks and eight models demonstrate that EviConf achieves the highest AUROC for correctness among the estimators compared in all evaluated settings. The authors also report reductions in confident error rates of 55–92% for video QA and 19–65% for audio-visual QA.

Critical strengths

1.The paper addresses the important issue of false confidence in multimodal models and provides an intuitive motivation for incorporating multimodal evidence into confidence estimation.

2.The method is evaluated across four benchmarks and eight models, consistently achieving the highest AUROC among the compared baselines. The ablation studies further investigate the contributions of different components.

Critical weaknesses

1.The proposed temporal support measures how frequently shuffling video frames decreases the probability of the selected answer. However, such a decrease does not necessarily indicate that the original temporal order provides correct evidence. For instance, an incorrect prediction may also become less confident after temporal shuffling. Therefore, the proposed metric may primarily capture sensitivity to temporal order rather than the reliability of temporal reasoning.

2.Equation 3 combines content and temporal support multiplicatively, which may be inappropriate for questions that do not require temporal reasoning. For example, when answering a static question such as "What color is the man's shirt?", temporal shuffling may not decrease the answer probability, resulting in zero temporal support. Consequently, the positive contribution of content evidence is completely eliminated, regardless of how strongly the visual input supports the answer. The paper does not adequately justify this design choice.

Questions for authors

1.Why was the number of temporal shuffles fixed at K=4? Could the authors provide a sensitivity analysis with different values of K?

2.Figure 1 appears inconsistent with Equation 3. For example, the content lift (+0.50) and temporal contribution (+0.27) are added to obtain +0.77, whereas Equation 3 uses multiplicative fusion. Additionally, the internal confidence for the Violin example differs between the left (0.91) and right (0.92) panels.

3.The content evidence score is defined as the probability difference between the full input and the question-only input. However, if the blind-input distribution is nearly uniform or exhibits systematic answer-position biases, this score may largely reflect the full-input prediction probability rather than independent evidence support.

  1. The Related Work section lacks a comprehensive discussion of multimodal uncertainty estimation methods based on input perturbation and sampling. For example, FESTA (Findings of EMNLP 2025) explores uncertainty estimation through functionally equivalent and complementary multimodal input samples, which appears closely related to EviConf's input intervention strategy.

5.The paper does not sufficiently explain why the proposed Full-vs-Blind Contrast and Temporal Shuffle strategies should be more effective than existing multimodal input-sampling approaches.

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

The paper proposes EviConf, a training-free confidence estimator for multiple-choice video and audio-visual question answering. It combines predictive-entropy confidence with the selected answers probability increase over a question-only input, modulated by how frequently temporal shuffling decreases that answer’s probability. Evaluation covers eight models and four benchmarks with the highest reported result in all 16 model–benchmark settings.

Critical strengths
  1. A relevant problem and an interpretable method. The motivation is clear and practical: a model can be confident without adequately using the supplied evidence.
  2. The experiments include video and audio-visual tasks, multiple model sizes, and cross-family evaluations.
Critical weaknesses
  1. The closest related method is cited but absent from the experiments, where VAUQ already proposes training-free self-evaluation by combining predictive entropy with a measure of visual evidence contribution, including a full-versus-missing-visual-input formulation and a stronger masking-based variant.
  2. The temporal component requires stronger validation relative to its cost. Tables 5 and 7 report an average improvement of 0.9 AUROC points overall and 1.7 points on temporal questions, based on 200-example subsets. The component adds four multimodal passes, increasing the total from two to six. Its effects are heterogeneous: for example, overall AUROC decreases by 2.7 points for Qwen3-Omni-30B on DailyOmni.
  3. The evidence supports improved correctness ranking more strongly than calibrated confidence: ECE improvements are inconsistent, and evaluation is restricted to multiple-choice QA.
Questions for authors

See Weaknesses Above.

Reviewer_9LdnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work deals with reliable confidence estimation in multimodal prediction. Specifically, it considers multimodal large language models for video and audio-visual question answering. These models may confidently give wrong answers based on language priors, ignoring evidence from their input. This work addresses this challenge by introducing EviConf, a training-free confidence estimator. EviConf estimates an MLLM's confidence in its prediction for a given query by combining its internal confidence with its sensitivity to evidence. It does so by comparing the predictive probability with and without the evidence (content term) and by testing whether shuffling the frames lowers it (temporal term). These terms form an evidence score, which is combined with internal confidence through a geometric mean. EviConf is evaluated on 4 video and audio-visual QA benchmarks across 8 models. It is compared with probability-based, sampling-based, and supervised baselines. Performance is quantified as correctness AUROC and calibration error (ECE). Results show that EviConf achieves the best AUROC in all settings and substantially reduces confident errors.

Critical strengths
  • The proposed method is simple, and straightforward to implement
  • Empirical evaluation is extensive, comparing EviConf against 7 methods on 4 benchmark datasets and 8 models.
Critical weaknesses
  • While EviConf achieves high predictive performance (AUCROC), it has high calibration error.
  • The evaluation is performed on subsamples of the original benchmark datasets, and the results seem to be from single runs rather than aggregates over multiple runs.
Questions for authors
  • Can you explain why EviConf has low calibration despite high predictive performance?
  • Can you clarify whether the reported metrics are aggregates over multiple runs or over a single run? It'd be hard to clearly compare the performance of different methods without disentangling it from inference & data related stochasticity

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.