cs.CLOct 12, 2026

Estimating Uncertainty of Omnimodal Large Language Models via External Omnimodal Embeddings

Authors: Iacopo Ghinassi, Hy Nguyen Gia La, Luke Ong

Organizations: CCDS, Nanyang Technological University, Singapore, 639798, Singapore

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent work has investigated Omni-modal Large Language Models (OLLMs), models which can take multiple modalities and their combinations as input, as a way forward to have systems really capable of interacting with their environment. Uncertainty quantification (UQ) for OLLMs, however, remains an open problem: existing methods largely target at most two modalities at a time, require access to model internals, or re-invocation of the evaluated model. We propose a model-agnostic UQ framework that scores a Chain-of-Thought response entirely through external omni-modal embeddings, measuring the grounding of the reasoning in the multimodal input and its internal coherence. Our contribution is the family of signals and an empirical characterisation of when it is reliable, rather than any single estimator: we find the geometric scores to be largely redundant with one another (mean within-split rank correlation 0.91-0.94) but strongly complementary to answer-dispersion baselines, and their reliability to depend on the generator and the modality. The resulting confidence estimates require no access to logits, hidden states or gradients of the evaluated model and no additional inference from it; they do require segmenting the reasoning trace with an auxiliary model and, for the aggregated score, a labelled calibration set that we show can be as small as 25 examples (although we also show that such a set buys little over fixing one signal a priori). On UNO-Bench, with both a closed and an open model, our scores yield small gains over self-consistency (systematic with the closed generator, mixed with the open one) and are competitive with self-verbalisation and internal-state baselines, especially in open-ended, omni-modal questions where previous state-of-the-art baselines struggle.

Figures & tables

Figure 1: We map the omni-modal inputs and generated Chain-of-Thought reasoning steps into a shared semantic space to compute Coherence and Grounding , which grade the majority-selected answer without internal model logits and without any additional call to the evaluated model.
Figure 2: Visual overview of the proposed method, as explained in Methodology section.
Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Prompt used for semantic clustering of sampled answers.
Figure 6: Media-permutation control for GansavgG^{\text{avg}}_{\text{ans}} . Circles are the AUROC obtained with each question’s own media; bars are the central 95%95\% of the null obtained by re-joining questions to other questions’ media within blocks of equal split and modality signature, with the null mean marked. The upper row reads the grounding score at question level, the lower row passes it through the deployed vote share of Equation 15 , which normalises within a question and therefore cancels what the media contribute equally to all of its candidates. The dashed line is chance.
Figure 7: AUROC ( ↑\uparrow ), AURAC ( ↑\uparrow ), and ECE ( ↓\downarrow ) pooled over the union of splits (Appendix D ) as a function of the number of sampled candidates NN (MiniCPM-generated answers; means over 100 randomized evaluation runs per budget, error bars show ±1\pm 1 standard error of the mean).
PackageVersion
Python3.12
numpy2.3.2
pandas2.3.1
scipy1.16.1
scikit-learn1.7.1
torch2.8.0
Appendix
Table 33: Primary software dependencies used in our implementation.
More in this topic

Uncertainty Quantification

Uncertainty quantification estimates how reliable predictions and scientific inferences are, and how uncertainty propagates into decisions.

See all 663 papers

Explore similar work

CardsList
Feb 27, 2026

Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume

Despite their capabilities, Multimodal Large Language Models (MLLMs) may produce plausible but erroneous outputs, hindering reliable deployment. Accurate uncertainty metrics could enable escalation of unreliable queries to human experts or larger models for improved performance. However, existing uncertainty metrics have practical constraints, such as being designed only for specific modalities, reliant on external tools, or computationally expensive. We introduce UMPIRE, a training-free uncertainty quantification framework for MLLMs that works efficiently across various input and output modalities without external tools, relying only on the models' own internal modality features. UMPIRE computes the incoherence-adjusted semantic volume of sampled MLLM responses for a given task instance, effectively capturing both the global semantic diversity of samples and the local incoherence of responses based on internal model confidence. We propose uncertainty desiderata for MLLMs and provide theoretical analysis motivating UMPIRE's design. Extensive experiments show that UMPIRE consistently outperforms baseline metrics in error detection and uncertainty calibration across image, audio, and video-text benchmarks, including adversarial and out-of-distribution settings. We also demonstrate UMPIRE's generalization to non-text output tasks, including image and audio generation.

Gregory Kang Ruey Lau, Hieu Dao, Nicole Kan Hui Lin +1

Department of Computer Science, National University of Singapore

LLM Uncertainty EstimationMultimodal Large Language ModelsMultimodal Model Evaluation
AskPDF
Sep 1, 2026

Dissecting Training-Free Uncertainty Estimation in Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical settings, yet estimating when their answers can be trusted remains underexplored. We systematically compare training-free verbalized, token-level and semantic uncertainty estimation methods on twelve MLLMs and five VQA benchmarks. We find that answer length decides which family wins: token-level entropy on short answers, verbalized methods on sentence-length responses, and semantic methods on long-form generation. Token-level entropy becomes far more informative when estimated at a higher temperature than the low ones used to generate answers, and asking a model to judge its own answer becomes more reliable with model size. Building on this, a length-aware recipe that additionally flags responses on which the model abstains outperforms every individual method on average at about a quarter of the cost of semantic methods, although no method attains the coverage required for safe deployment. These results provide practitioners with principled guidance for selecting estimators by answer length and model scale.

Soroush Seifi, Vaggelis Dorovatas, Lin Li +2

Toyota Motor Europe · University of Oxford

Uncertainty QuantificationMultimodal Large Language ModelsLLM Evaluation
AskPDF
Aug 17, 2026

Uncertainty-Aware Decision Making in Multimodal Large Language Models

Multimodal large language models (MLLMs) increasingly answer questions whose correctness depends on visual, textual, temporal, acoustic, document, chart, or embodied evidence. Their failures are therefore not only linguistic. A fluent answer may conceal poor input quality, a perceptual error, weak grounding, conflict between modalities, unstable reasoning, distribution shift, or a question that is not answerable from the supplied evidence. This survey organizes the literature on uncertainty-aware MLLMs around a decision-centered framework: uncertainty sources give rise to observable signals, signals must be calibrated or controlled for risk, and calibrated uncertainty should determine the system action. We review work on token and logit uncertainty, semantic disagreement, perturbation instability, grounding and attribution scores, verbalized confidence, verifier and judge scores, conformal prediction, selective answering, abstention, clarification, retrieval, self-checking, and escalation. The central argument is that uncertainty should not be evaluated only as a confidence number; it should be evaluated by whether it improves behavior under insufficient, conflicting, shifted, or high-risk multimodal evidence. We position this survey against text-only uncertainty and abstention surveys, broad MLLM surveys, MLLM hallucination surveys, and safety-oriented reviews. We conclude with open problems in source-aware decomposition, action-aware benchmarks, calibration under shift, black-box uncertainty estimation, broader modality coverage, reproducible reporting, and human-centered uncertainty communication.

Abderrahmene Boudiaf, Irfan Hussain, Sajid Javed

Khalifa University of Science and Technology, Abu Dhabi, United Arab Emirates

Multimodal Large Language ModelsLLM Uncertainty EstimationDecision-Making under UncertaintySelective Prediction
AskPDF