cs.CLAug 17, 2026
SaveUncertainty-Aware Decision Making in Multimodal Large Language Models
Organizations: Khalifa University of Science and Technology, Abu Dhabi, United Arab Emirates
Abstract
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.
Figures & tables
Figure 1: Source–signal–calibration–action framework for uncertainty-aware MLLMs. Multimodal inputs and domain context give rise to uncertainty sources. These sources are observed through model-internal, output-level, evidence-grounding, verbalized, verifier, or judge signals. Calibration and risk-control methods convert raw signals into decision-relevant estimates. The final policy selects an action: answer, hedge, abstain or say IDK, clarify, request better input, retrieve evidence, self-check, or route the case to a stronger model or human expert.
| Survey family | Main coverage | Gap for this survey |
|---|---|---|
| Text-only LLM uncertainty | Uncertainty quantification, confidence calibration, semantic uncertainty, and selective prediction in language-only settings [ 44 , 28 , 58 , 16 ] . | Limited treatment of visual evidence quality, grounding, cross-modal conflict, and multimodal answerability. |
| Text-only abstention and IDK | Knowledge boundaries, refusal, and selective answering in language-only settings [ 54 ] . | Abstention is not tied to visual sufficiency, image quality, or multimodal grounding. |
| MLLM hallucination and safety | Hallucination symptoms, attack and defense settings, benchmark construction, and safety risks [ 3 , 26 , 29 , 5 ] . | Often organized around failure detection or safety risk rather than uncertainty signals, calibration, and action policies. |
| Broad MLLM and evaluation surveys | Architectures, training data, capabilities, benchmarks, and application coverage [ 65 , 73 ] . | Too broad to compare uncertainty sources, signal access assumptions, calibration targets, and response actions in detail. |
| This survey | Source-aware uncertainty, observable signals, calibration, risk control, selective answering, abstention, clarification, retrieval, self-checking, escalation, and action-aware evaluation. | Integrates multimodal uncertainty estimation with decision policy. |
Table 1: Positioning against adjacent survey families. The table separates related coverage from the gap addressed by this survey. Citations in the first four rows are representative anchors for adjacent survey families; the final row states the positioning of the present survey.
Figure 2: From evidence conditions to uncertainty-aware actions. The same user question can require different model behavior depending on the quality, availability, and grounding of the visual evidence. With clear evidence, direct answering is appropriate, and unnecessary abstention becomes a failure mode. With hazy or partially degraded evidence, the model should hedge or request better input rather than over-specifying an uncertain visual attribute. With occluded evidence, a confident answer may reflect weak grounding or distractor bias, so the safer behavior is to state uncertainty or abstain. With referent ambiguity, the visual input contains multiple plausible targets, so the model should clarify or disambiguate the referred object instead of answering from a salient distractor. The examples are illustrative rather than exhaustive.
Figure 3: Propagation and interaction of MLLM uncertainty sources. The figure illustrates how uncertainty can cascade across sensory, grounding, reasoning, and policy layers in multimodal language models. In the sensory–grounding chain, ambiguous visual input can lead to perceptual uncertainty, weaken the alignment between evidence and claims, and allow language priors to fill visual gaps. These failures can become entangled with cross-modal conflict, where text and image evidence disagree, and with reasoning uncertainty, where inference becomes unstable under weak or conflicting evidence. System-level uncertainty can also arise from distributional or adversarial conditions that cause calibration failure, while missing or absent evidence creates answerability uncertainty and should trigger abstention or clarification rather than forced answering. Black arrows denote primary causal pathways, gray arrows denote entanglement between uncertainty sources, and colors distinguish perceptual-chain components, reasoning components, systemic sources, and model actions. The diagram is a schematic synthesis rather than an exhaustive causal model.
| Source | What goes wrong | Possible signal | Likely action |
|---|---|---|---|
| Sensory and perceptual uncertainty [ 53 , 13 , 63 , 27 ] | The relevant evidence is degraded, ambiguous, occluded, noisy, too small, temporally sparse, or misrecognized. | Visual perturbation instability, low perceptual confidence, inconsistent captions, OCR or layout failure. | Request better input, clarify the target, route to a specialized perception module, or abstain. |
| Grounding and cross-modal uncertainty [ 37 , 7 , 1 , 40 ] | The answer is weakly supported by the evidence, or the text, image, retrieved context, or other modalities disagree. | Grounding score, attribution mismatch, hallucination detector, semantic disagreement, verifier or judge score. | Self-check, cite evidence, retrieve, abstain, or escalate. |
| Language-prior and reasoning uncertainty [ 17 , 25 , 4 , 42 ] | The model fills visual gaps with prior knowledge, or an unstable reasoning path changes the conclusion. | Sampling disagreement, reasoning-step confidence, semantic variation, verifier score, answer variation. | Self-check, sample alternatives, use tools, retrieve, or escalate. |
| Answerability uncertainty [ 33 , 53 , 46 , 77 ] | The question cannot be answered from the supplied input or requires missing evidence. | Answerability classifier, IDK score, prediction-set size, refusal consistency, unsupported-evidence flag. | Abstain, say IDK, ask for missing evidence, or clarify. |
| Distributional, adversarial, and deployment uncertainty [ 6 , 59 , 52 , 49 ] | The model faces domain shift, modality shift, corrupted input, misleading prompts, adversarial cues, or high-risk deployment conditions. | Calibration drift, conformal risk estimate, perturbation disagreement, shift detector, safety or domain-risk score. | Apply risk control, abstain, retrieve, route, or escalate. |
Table 2: Multimodal uncertainty sources and decision implications. This table defines the source taxonomy used in later sections. Citations in the source column are representative anchors; the signal and action columns are author synthesis from the cited evidence and surrounding discussion.
Figure 4: Choosing uncertainty signals under model-access constraints. The figure maps model access to observable uncertainty signals, diagnosed targets, and supported actions. White-box access enables model-internal signals; gray-box access supports output-distribution signals; and black-box access relies on prompts, responses, tools, and judge-based signals. These signals diagnose confidence, semantic instability, grounding failure, or answerability risk, which then guide whether the system should proceed, seek additional evidence, or defer. Arrows indicate typical rather than exclusive pathways.
| Signal family | Typical access | Primary diagnosis | Action value | Main caveat |
|---|---|---|---|---|
| Token, logit, and entropy signals [ 74 , 17 , 8 ] | Logits or probabilities | Decoding uncertainty, hidden knowledge boundary, label ambiguity | Confidence reporting, selective answering, refusal triggers | Often unavailable in closed models and weak at separating uncertainty sources. |
| Hidden-state and visual-token signals [ 13 , 39 , 20 , 63 ] | Open weights, embeddings, attention, or visual-token features | Perception, visual grounding, OCR or region-level uncertainty | Request better input, route to perception module, abstain from ungrounded answers | Requires model access and may be difficult to compare across architectures. |
| Sampling and semantic disagreement [ 69 , 68 , 4 , 67 , 55 ] | Repeated outputs, perturbed prompts, or equivalent inputs | Response instability, reasoning uncertainty, epistemic uncertainty | Self-checking, selective prediction, abstention, verifier selection | Expensive and sensitive to sampling settings; repeated agreement can still be confidently wrong. |
| Grounding, attribution, and perturbation checks [ 37 , 61 , 70 , 42 , 15 ] | Region evidence, attribution, grounding model, or input perturbation | Evidence support, hallucination risk, cross-modal conflict | Evidence citation, retrieval, better-input requests, abstention, escalation | Depends on the quality of the grounding or perturbation procedure. |
| Verbalized confidence and IDK scores [ 60 , 8 , 53 , 46 , 41 ] | Prompt access and generated text | Self-reported confidence, answerability, knowledge boundary | IDK, clarification, user-facing uncertainty communication | Verbal confidence may be miscalibrated or strategically phrased. |
| Verifier, judge, ensemble, and agentic signals [ 24 , 38 , 71 , 9 , 30 ] | Extra model, judge, tool, or controller | Cross-model disagreement, route quality, answer support | Retrieval, self-checking, routing, escalation | The judge or controller also requires calibration and can introduce bias. |
Table 3: Signal and method taxonomy for uncertainty-aware MLLMs. The table is organized by what information is available, what uncertainty source the signal mainly diagnoses, and what action it can support. Citations in the signal-family column provide representative support for each method family.
| Action | When it is appropriate | Signals that can trigger it | Potential failure mode |
|---|---|---|---|
| Answer [ 11 , 59 , 37 ] | Evidence is sufficient and calibrated risk is within the accepted range. | High grounding support, stable samples, calibrated confidence, low shift signal. | Confident hallucination if the threshold is too loose. |
| Hedge [ 60 , 8 , 56 ] | Evidence is partial but still useful, and the limitation can be stated precisely. | Moderate confidence, mild disagreement, incomplete but relevant support. | Vague caution that hides a lack of evidence. |
| Abstain or IDK [ 33 , 53 , 46 , 77 , 66 ] | The required evidence is missing, incompatible, outside the input, or beyond the model boundary. | Answerability score, IDK probability, large prediction set, unsupported-evidence flag. | Excessive refusal and reduced task usefulness. |
| Clarify [ 19 , 43 , 57 ] | The user’s question, referent, intent, or preference is underspecified. | Ambiguous reference, multiple plausible interpretations, conflicting user context. | Unnecessary interaction when the answer was already clear enough. |
| Request better input [ 27 , 43 , 57 ] | The query is clear but the sensory evidence is degraded or incomplete. | Blur, occlusion, low OCR confidence, visual perturbation instability, missing video segment. | Burdening the user when the model should use available evidence. |
| Retrieve, self-check, or verify [ 9 , 76 , 38 , 71 , 70 ] | External knowledge, evidence checking, or reasoning validation is needed before answering. | Reasoning disagreement, weak factual support, verifier uncertainty, tool-need signal. | Added latency, retrieval noise, or over-reliance on an imperfect verifier. |
Table 4: Action-policy taxonomy for uncertainty-aware MLLMs. The table maps evidence conditions and uncertainty signals to response actions. Citations attached to action labels are representative examples; the trigger and failure-mode columns are synthesized policy implications.
Figure 5: Evaluating uncertainty-aware behavior. Evaluation should test whether the model’s uncertainty leads to an appropriate action for a labeled or controlled evidence condition. Clear and grounded evidence should support proceeding with an answer, degraded or ambiguous evidence should trigger evidence-seeking actions such as clarification, better-input requests, retrieval, or self-checking, and missing evidence or high-risk contexts should support deferral through abstention, IDK, or escalation. The evaluation target therefore depends on the selected action: prediction quality for proceeding, evidence improvement for seeking additional information, and safety or risk control for deferral. When evidence-condition labels are unavailable, reliable source- and action-aware evaluation remains an open benchmark challenge.
| Evaluation question | Benchmark family | Metric family | What can remain hidden |
|---|---|---|---|
| Is confidence aligned with correctness? | General VQA, chart QA, medical VQA, multimodal reasoning, open-ended QA [ 6 , 59 , 10 , 22 , 60 ] . | ECE, Brier score, NLL, reliability diagrams, confidence-accuracy gap. | Whether the answer is grounded, answerable, or safe to act on. |
| Can uncertainty rank risky outputs? | Error detection, hallucination detection, selective prediction, verifier screening [ 11 , 69 , 55 , 70 , 7 ] . | AUROC, AUPRC, F1, TPR at fixed FPR, risk-coverage curves. | Which action should follow after a risky output is detected. |
| Does the model know when evidence is insufficient? | Unanswerable VQA, visual IDK, video refusal, embodied QA, audio-visual QA [ 33 , 53 , 46 , 77 , 66 , 50 ] . | Answerability accuracy, abstention precision, refusal helpfulness, IDK calibration. | Whether abstention is too conservative or insufficiently informative. |
| Can the system control risk? | Conformal MLLM/VLM benchmarks, prediction-set evaluation, risk-controlled response selection [ 52 , 49 , 64 , 2 , 15 ] . | Empirical coverage, violation rate, prediction-set size, set efficiency, controlled risk. | Robustness under domain shift, prompt shift, and calibration-data mismatch. |
| Does uncertainty improve user-facing behavior? | Grounded QA, clinical or document workflows, interactive agents, high-risk settings [ 47 , 19 , 27 , 9 , 76 ] . | Action utility, escalation appropriateness, human ratings, explanation quality, support precision. | Whether users understand the uncertainty source and can act on it safely. |
Table 5: Evaluation guide for uncertainty-aware MLLMs. The table links evaluation questions to benchmark families, metric families, and hidden failure modes. Citations in the benchmark-family column are representative examples of the evaluation setting.
Figure 6: Research roadmap for uncertainty-aware MLLMs. The roadmap organizes future work by time horizon. Near-term work should build the foundations for source-aware uncertainty analysis, including evidence-condition labels and action-aware benchmarks. Mid-term work should make uncertainty estimates reliable under realistic access and deployment conditions through black-box uncertainty estimation, calibration under shift, source-action policies, and evaluation beyond accuracy. Long-term work should support deployment in broader modalities, multi-turn and agentic settings, and human-facing applications, with appropriate communication and governance mechanisms.
More in this topic
Multimodal Large Language Models
Multimodal large language models combine language models with image, video, audio, and other inputs for grounded understanding and generation.
See all 1,091 papers