PaperlayerView exact revision on arXiv

Source-linked AI summary

Uncertainty-Aware Decision Making in Multimodal Large Language Models

Abderrahmene Boudiaf, Irfan Hussain, Sajid Javed

arXiv:2608.17084v1cs.CL

TL;DR

MLLM uncertainty can arise from inadequate, conflicting, or shifted multimodal evidence, yet confidence alone does not determine whether an answer is appropriate. This survey organizes uncertainty sources, signals, calibration, and actions, concluding that uncertainty-aware systems should adapt behavior through answering, abstention, clarification, retrieval, verification, or escalation.

  • Problem

    Uncertainty in MLLMs spans input quality, perception, grounding, reasoning, distribution shift, and answerability, but these sources are often not distinguishable from a confidence value alone.

  • Method

    The survey synthesizes a source–signal–calibration–action taxonomy linking multimodal uncertainty sources and observable signals to calibrated response policies.

  • Results

    The surveyed methods show that uncertainty can guide calibrated answering, abstention, clarification, retrieval, self-checking, verification, and escalation rather than merely report confidence.

  • Takeaways & Limitations

    Uncertainty should be evaluated by whether it changes system behavior appropriately when multimodal evidence is insufficient, conflicting, shifted, or high-risk.

  • Takeaways & Limitations

    Risk-control guarantees depend on the calibration distribution, score function, and assumed relationship between calibration and deployment cases.

Abstract

from arXiv · show

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.

1. Introduction

This survey frames uncertainty in multimodal large language models as a source–signal–calibration–action problem rather than a single confidence score. It argues that uncertainty awareness is demonstrated when calibrated estimates change system behavior, including answering, hedging, abstaining, clarifying, requesting better input, or retrieving evidence.

  • Introduction: MLLM uncertainty can originate in input quality, perception and grounding, language-based completion of missing evidence, cross-modal conflict, reasoning, answerability, or distribution shift.Examples include low-quality images, omitted video moments, noisy audio, incorrect object recognition, wrong-region attention, and weak visual support.
  • Introduction: Text-only uncertainty methods provide foundations but cannot establish whether multimodal evidence was visible, agreed across modalities, correctly grounded, or sufficient for clarification.The survey therefore treats multimodal uncertainty as requiring evidence-sensitive analysis beyond text-only confidence.
  • Introduction: Uncertainty is decision-centered: reliable MLLMs should answer when evidence is sufficient, hedge when confidence is limited, abstain when evidence is missing, and clarify underspecified queries.The framework also supports requesting better input, retrieving evidence, and self-checking.
  • Introduction: The survey’s organizing chain maps multimodal inputs and context to uncertainty sources, observable signals, calibrated confidence or controlled risk, and an action policy.Signals may be model-internal, output-level, evidence-grounding, verbalized, verifier, or judge signals.
  • Introduction: The survey contributes a taxonomy of multimodal uncertainty sources, connects estimation to response actions, and positions its scope around MLLM uncertainty estimation, calibration, risk control, answerability, abstention, and action selection.Text-only, hallucination, safety, and broad MLLM surveys are used selectively as background or to clarify the survey’s boundary.

2. Scope, Positioning, and Evidence Base

The survey defines uncertainty-aware MLLM research as linking uncertainty estimation and calibration to decisions such as answering, hedging, abstaining, clarifying, retrieving, verifying, or escalating. It focuses on multimodal evidence and positions itself around the source-to-action chain, supported by 67 focused evidence rows.

  • Scope: Uncertainty-aware decision making includes estimating, expressing, calibrating, and controlling uncertainty to improve behavior under multimodal evidence.The downstream policy determines whether to answer, hedge, abstain, clarify, request better input, retrieve evidence, verify, or escalate.
  • Decision framework: The framework maps multimodal evidence, task, and context to an uncertainty signal, calibrated confidence or risk estimate, and action policy.Possible actions include direct answers, hedged answers, IDK responses, clarification, retrieval, or routing to another system or expert.
  • Key concepts: The survey distinguishes uncertainty sources, observable signals, calibration, and policies that connect estimated confidence or risk to correctness, cost, and action.Signals include entropy, logit margin, semantic disagreement, perturbation instability, grounding, verbalized confidence, prediction-set size, answerability, and verifier scores.
  • Inclusion boundaries: The survey includes work where uncertainty informs confidence, calibration, selective answering, conformal prediction, error detection, answerability, refusal, self-checking, retrieval, or escalation.Adjacent hallucination, safety, robustness, and evaluation work is central only when it links evidence uncertainty to calibrated confidence, risk control, or action selection.
  • Positioning and evidence base: The survey addresses a gap left by text-only uncertainty and abstention, MLLM hallucination and safety, and broad MLLM evaluation surveys by centering the multimodal source-to-action chain.Its core evidence base contains 67 focused evidence rows from MLLM, LVLM, VLM bridge, domain-specific, and directly relevant survey studies.

3. Sources of Uncertainty in MLLMs

MLLM uncertainty arises from multiple interacting sources spanning input quality, perception, grounding, reasoning, distribution shift, and answerability. Because these sources require different responses, systems should identify uncertainty before committing to an answer.

  • Taxonomy: A source-aware taxonomy distinguishes uncertainty from the input, perceptual encoder, cross-modal alignment, reasoning path, deployment domain, and task answerability.Different sources call for different actions, such as better input, clarification, or abstention.
  • Sensory and perceptual uncertainty: Sensory and perceptual uncertainty arises from degraded or ambiguous evidence and may be amplified rather than repaired by additional language reasoning.Examples include blur, low resolution, cropping, occlusion, noise, temporal sparsity, object misrecognition, unreadable image text, and confused chart axes.
  • Grounding and cross-modal uncertainty: Grounding uncertainty concerns whether answers are supported by multimodal evidence, while cross-modal uncertainty concerns disagreement among modalities.A plausible answer may remain unanchored to the relevant region or conflict with visual, textual, or retrieved evidence.
  • Reasoning uncertainty: Language priors can produce fluent but unsupported completions, while reasoning uncertainty appears when intermediate steps or sampled reasoning paths yield unstable conclusions.These priors help when evidence is complete but become risky when evidence is ambiguous or missing.
  • Answerability uncertainty: Answerability uncertainty occurs when the requested information is absent from the image, video, chart, or supplied context, making abstention or requests for missing evidence appropriate.The correct response may be no answer or a question for additional evidence rather than ordinary error correction.
  • Interactions and shift: Calibration can fail under distribution shift or adversarial inputs, and uncertainty sources can cascade from poor input through perception and grounding to reasoning and policy.Missing evidence can make both confidence and explanations misleading.

4. Observable Signals and Estimation Methods

Observable uncertainty signals convert latent failures into actionable evidence, but their usefulness depends on model access, the uncertainty source they approximate, and the action they support. Internal, sampling-based, grounding, self-assessment, and verifier signals each offer benefits while requiring calibration and risk control before guiding decisions.

  • Signal selection: Uncertainty signals make latent conditions observable and should be evaluated by required model access, approximated uncertainty source, and supported action.Signals may use token probabilities, hidden states, repeated samples, perturbations, grounding checks, verbalized confidence, verifiers, judges, or conformal scores.
  • Model-internal signals: Internal signals such as logits, entropy, hidden representations, and visual-token measures can support withholding, abstention, and vision-grounding control when model access is available.Their limitation is that many deployed MLLMs expose neither logits nor hidden states, and internal scores may not identify whether uncertainty arises from perception, grounding, reasoning, or answerability.
  • Sampling-based signals: Sampling-based signals use disagreement across repeated, paraphrased, equivalent, or perturbed inputs, providing practical closed-model access but incurring cost and ambiguity.Correct paraphrases can appear diverse, while repeated identical errors can appear stable; grounding, answerability, or calibration checks improve reliability.
  • Grounding and cross-modal evidence: Grounding signals test whether answers are supported by supplied multimodal evidence, while contradiction signals detect disagreement across modalities that may warrant clarification, retrieval, self-checking, or abstention.Examples include region grounding, attribution, perturbation sensitivity, OCR and chart checks, cross-view consensus, and contradiction detection.
  • Self-assessment signals: Verbalized confidence and IDK-style answerability signals expose self-assessment and knowledge boundaries, but explicit confidence can diverge from correctness and implicit confidence.Answerability-oriented methods connect more directly to abstention, refusal, and clarification than confidence statements alone.
  • Verification and calibration: Verifiers, judges, ensembles, and controllers enable richer decision policies, yet their outputs require calibration because ranking ability does not guarantee calibrated confidence.A signal can rank risky outputs well while remaining miscalibrated or calibrated for a different target than the action requires.

5. Calibration and Risk Control

Calibration gives uncertainty scores their intended meaning, while risk control links them to acceptable error, violation, or unsupported-answer rates. For MLLMs, effective calibration must define an action-paired target, preserve uncertainty sources, and account for failure modes such as overconfidence and distribution shift.

  • Calibration and risk control: Calibration and risk control are both necessary because raw uncertainty signals are not decision-ready.A score may rank errors usefully while remaining poorly calibrated, and verbal confidence may be empirically unreliable.
  • Target definition: MLLM calibration must specify its target, which may include correctness, evidence support, answerability, prediction-set coverage, hallucination risk, or downstream action cost.These targets are not interchangeable: calibration on multiple-choice VQA accuracy can coexist with overconfidence on unanswerable images.
  • Action-paired calibration: Pairing calibration targets with actions prevents one generic confidence value from being reused across selective answering, IDK, better-input requests, and escalation.Selective answering targets incorrect accepted-answer risk, IDK targets evidence sufficiency or answerability, and better-input requests target perceptual adequacy.
  • Source-aware calibration: Source-aware calibration separates visual, reasoning, and answerability uncertainty so each can trigger a different policy.Low visual confidence can prompt better input or perception routing; low reasoning confidence can prompt self-checking or verification; low answerability can prompt IDK or clarification.
  • Operational risk control: Conformal and risk-control methods convert calibrated scores into explicit acceptance, set-valued, abstention, or review policies under stated assumptions.Their guarantees depend on the calibration distribution, score function, and assumptions, while major failure modes include overconfidence, underconfidence, source conflation, verbalization mismatch, and shift sensitivity.

6. From Uncertainty to Action

Uncertainty-aware MLLMs should select actions according to evidence conditions, calibrated risk, user needs, and error costs rather than applying one response policy universally. The action space spans answering, abstaining, clarifying, improving inputs, retrieving, self-checking, routing, and escalation, with evaluation focused on decision appropriateness.

  • Action policy: Action selection is part of uncertainty handling: systems may answer, hedge, abstain, clarify, request better input, retrieve, self-check, route, or escalate.The appropriate action depends on evidence condition, calibrated risk, user need, and cost of error.
  • Answering and abstention: Selective answering permits responses only when expected risk is below a threshold, while abstention should be calibrated rather than automatic.Selective methods are evaluated using coverage, risk, and risk-coverage curves, and multimodal work extends these ideas across visual, audio-visual, and shifted settings.
  • Answerability and IDK: IDK should reflect unanswerability from missing, incompatible, insufficient, or context-external evidence, including partial observability and underspecified embodied tasks.This frames IDK as an interaction policy under partial evidence rather than merely a language behavior.
  • Clarification and input quality: Clarification addresses ambiguous questions, whereas better-input requests address inadequate sensory evidence such as blur, incomplete crops, noise, or missing video moments.These actions preserve usefulness by avoiding forced direct answers when the question or evidence needs repair.
  • Evidence improvement: Retrieval, self-checking, and verification improve the evidence pathway when external knowledge or unstable reasoning requires additional support before answering.Agentic systems can use uncertainty to call tools, consult verifiers, compare candidates, refine answers, or flag findings for review.
  • Routing and escalation: Tiered routing answers low-risk cases, retrieves or self-checks medium-risk cases, and escalates high-risk cases when error costs are unacceptable.Evaluation should test whether the selected action fits the evidence condition and deployment risk, not merely whether uncertainty predicts errors.

7. Evaluating Uncertainty-Aware Behavior

Uncertainty-aware MLLM evaluation should connect evidence conditions, uncertainty signals, calibration, and action policies rather than relying on answer accuracy or a single confidence metric. Benchmarks and reporting should test whether uncertainty improves behavior under insufficient, conflicting, shifted, or high-risk evidence.

  • Evaluation protocol: Evaluation must link task and evidence conditions to the uncertainty signal, calibration rule, and both answer and action quality.Evidence conditions may be labeled or controlled, while signals must specify required model access and decisions should reflect calibrated thresholds.
  • Evidence conditions: Controlled or naturalistic evaluations should make evidence conditions explicit through interventions, annotations, metadata, or expert judgment.Examples include blur, crop, occlusion, distractors, missing evidence, modality conflict, object visibility, answerability, grounding support, and high-risk labels.
  • Benchmark families: Benchmark families should separately assess confidence–correctness, unanswerability and refusal, hallucination and grounding, risk control, and domain- or modality-specific calibration.These families expose different failure modes, including unsupported content, unavailable evidence, and domain-specific behavior that generic image–text benchmarks may miss.
  • Metrics: Metric choice should match the target: calibration metrics assess confidence alignment, detection metrics rank risk, selective metrics measure action tradeoffs, and conformal metrics enforce coverage or risk constraints.No single metric is sufficient, and conformal performance can degrade under distribution shift.
  • Reporting: Studies should report model version, modality, benchmark, prompt, access level, signal, calibration data, target, thresholds, action policy, costs, and failure cases.Explicit reporting is needed for meaningful comparison, especially for black-box MLLMs relying on repeated sampling, verbalized confidence, or judge scores.

8. Deployment Landscape

Deployment changes the meaning and acceptable use of uncertainty: low-risk systems may hedge or selectively answer, whereas medical, embodied, and safety-critical systems require conservative, inspectable, and escalation-aware policies. Because modalities produce distinct uncertainty-action patterns, human-facing systems should communicate specific causes and actionable next steps while becoming more source-aware, action-aware, shift-robust, and reproducible.

  • Risk-sensitive deployment: Medical, clinical, embodied, and safety-critical deployments require domain-calibrated, inspectable grounding and escalation integrated into the workflow.These settings emphasize confidence calibration, uncertainty propagation, grounded evidence, and risk-controlled reporting because unsupported fluent responses can cause harm.
  • Risk-sensitive deployment: Deployment risk should determine uncertainty policy: low-risk tasks can tolerate wider answer regions, while high-risk tasks need stricter thresholds, audit trails, and escalation.High-risk systems should not always refuse, because over-refusal can reduce usefulness.
  • Modality-specific deployment: Different modalities require different uncertainty actions, including reporting unseen video events, unreadable chart axes, supporting document pages, or audiovisual conflicts.Image-text, video, audio, chart, OCR, and document systems face distinct ambiguity, answerability, acoustic, timing, extraction, layout, scale, and numerical uncertainties.
  • Uncertainty communication: Human-facing systems should distinguish uncertainty causes and provide actionable next steps such as clearer inputs, missing context, retrieval, expert consultation, or tentative use.Users need to know why the model is uncertain and what to do next, rather than receiving only a vague hedge, refusal, or confidence score.
  • Open deployment gaps: Deployment pressures expose gaps requiring more source-aware, action-aware, shift-robust, and reproducible methods across models and modalities.These requirements connect uncertainty estimation to deployment behavior rather than treating uncertainty as only a confidence number.

9. Open Problems and Research Agenda

The research agenda moves from comparable, source-aware uncertainty evaluation toward robust deployment and decision-centered integration in interactive, tool-using, and human-facing multimodal systems. Progress depends on evaluating whether uncertainty improves actions under realistic access constraints, distribution shift, modality diversity, and human use.

  • Research roadmap: Near-term research should standardize evidence-condition labels and action-aware benchmarks, while later work targets source-aware estimates, calibration under shift, and evaluation beyond accuracy.The roadmap organizes these priorities by time horizon and emphasizes realistic access and deployment conditions.
  • Source-aware uncertainty: Uncertainty methods should separate perception, grounding, reasoning, answerability, and shift because each source can require a different system action.Visual ambiguity may require better input, grounding failure evidence checking, reasoning instability self-verification, and answerability failure abstention or clarification.
  • Action-aware benchmarks: Benchmarks should compare answers, abstentions, clarifications, retrieval, prediction sets, self-checks, and escalations using explicit action costs and downstream utility.The agenda shifts evaluation beyond correlation with correctness toward whether uncertainty produces better behavior.
  • Deployment robustness: Black-box uncertainty estimation is central because deployed MLLMs may expose only text outputs, while calibration can fail across domains, input quality, modalities, languages, prompts, and users.Feasible signals include repeated sampling, semantic disagreement, visual perturbation, verbalized confidence, external verifiers, and judge scores.
  • Modality coverage: Future work must broaden evaluation beyond image-text settings to audio, video, multi-image reasoning, dialogue, documents, charts, embodied systems, and tool-using agents.These settings introduce distinct challenges including temporal absence, acoustic noise, layout and numerical extraction, partial observability, and compounding decisions.
  • Reproducibility and human use: Reporting standards should document models, modalities, benchmarks, prompts, access, calibration, signals, targets, policies, thresholds, costs, metrics, and failures, while testing human-centered communication.Releasing prompts, calibration splits, evaluator instructions, and evidence matrices would make results more cumulative and reproducible.

10. Limitations of This Survey

The survey is limited by a time-bounded and uneven evidence base, heterogeneous studies that preclude direct numerical comparison, and its role as a structured synthesis rather than a quantitative meta-analysis. These constraints reinforce that the source–signal–calibration–action framework organizes and audits uncertainty claims but does not replace domain-specific validation.

  • Evidence base: The evidence base is time-bounded because relevant MLLM uncertainty, abstention, calibration, and risk-control work is rapidly evolving.Findings may change as models, benchmarks, venues, and paper versions are updated.
  • Evidence base: The survey avoids fixed leaderboard claims and single-number comparisons unless they are made within the original study.This limits conclusions that depend on results likely to change across updated studies.
  • Coverage: Coverage is uneven across modalities, with image-text settings better represented than audio, video, multi-image dialogue, embodied, omni-modal, and long-horizon agentic systems.The imbalance reflects the current evidence base rather than the relative importance of these settings.
  • Synthesis method: Heterogeneity in model access, prompts, calibration data, scores, evaluation targets, thresholds, and action policies makes direct numerical comparison difficult.The survey therefore organizes concepts, assumptions, and decision logic through tables and figures instead of quantitative meta-analysis.
  • Framework role: The framework organizes and audits uncertainty claims but does not substitute for domain-specific validation.These limitations clarify the framework’s role rather than weakening its necessity.

11. Conclusion

MLLM uncertainty extends beyond generated-sentence confidence to failures involving evidence quality, grounding, modality conflict, reasoning stability, distribution shift, and risk. The survey frames uncertainty through a source–signal–calibration–action chain and argues that progress depends on evaluating and improving behavior under uncertainty.

  • Conclusion: MLLM uncertainty can arise from blurred or absent evidence, visual misreading, weak grounding, cross-modal contradiction, unstable reasoning, distribution shift, or excessive autonomous-use risk.Fluent responses may conceal these distinct failure causes, which generic confidence measures do not explain.
  • Conclusion: The survey organizes uncertainty-aware MLLMs as a source–signal–calibration–action chain linking unreliability sources to observable signals, decision meaning, and system responses.Signals include probabilities, representations, samples, perturbations, grounding checks, verbalized confidence, verifier or judge scores, and conformal scores.
  • Conclusion: Evaluation should prioritize behavior under uncertainty rather than confidence quality alone, supported by source-aware decomposition, action-aware benchmarks, shift-robust calibration, broader modalities, and user-centered communication.The proposed scope includes audio, video, documents, charts, embodied agents, and interactive settings, alongside black-box calibration.

Appendix A. Evidence Mapping and Traceability · Appendix B. Main-Text and Supplement Boundary

The appendices preserve procedural and traceability material needed to audit how the survey’s evidence base was assembled, while the main text remains focused on synthesis. They also separate representative citations and verification practices from the fuller evidence matrix and reproducibility artifacts.

  • Appendix A. Evidence Mapping and Traceability: Appendix A retains search sources, query families, screening counts, evidence-tier definitions, extraction fields, and corpus-refresh details for auditability.These materials document how the evidence matrix was assembled without making the search procedure central to the survey.
  • Appendix A. Evidence Mapping and Traceability: The evidence mapping distinguishes 67 focused evidence rows from background survey citations and non-counted bridge material.The distinction clarifies which materials constitute the focused evidence base and which are retained as contextual support.
  • Appendix A. Evidence Mapping and Traceability: Main-table citations serve as representative row-level anchors rather than replacements for the full evidence matrix.The appendix therefore preserves the broader mapping needed to trace individual claims beyond the condensed main tables.
  • Appendix A. Evidence Mapping and Traceability: A final source-verification pass flags venue status and withdrawn or actively revised preprints.This check helps prevent the main text from relying on unsupported publication metadata.
  • Appendix B. Main-Text and Supplement Boundary: The main text is limited to five synthesis tables and six conceptual figures.The survey keeps its central presentation compact while moving additional documentation elsewhere.
  • Appendix B. Main-Text and Supplement Boundary: The appendices or companion supplement retain corpus composition, evidence tiers, evidence-audit summaries, search-log fields, representative excerpts, and the evidence-field schema.These artifacts provide procedural and evidentiary traceability beyond the main text.
  • Appendix B. Main-Text and Supplement Boundary: Supplementary materials also preserve metric-family intersection counts, source-action heatmap counts, application-domain risk diagnostics, and count-heavy research-roadmap diagnostics.These count- and risk-oriented artifacts support reproducibility without expanding the main synthesis tables.
Loading 2608.17084v1…