PaperlayerView exact revision on arXiv

Source-linked AI summary

Revealing Epistemic Uncertainty in MLLMs via Causal-Invariant Masking

Haoyang Luo, Linwei Tao, Jie Gui, Xinghao Chen, Chang Xu, Jianyuan Guo, Minjing Dong

arXiv:2610.02887v1cs.CVcs.AI

TL;DR

MLLM uncertainty methods can miss epistemic uncertainty associated with model limitations, complicating detection of hallucinations driven by spurious correlations. CIM compares semantic predictions on original and causally focused views, and its EED proxy estimates semantic shift in embedding space. The paper reports state-of-the-art performance and 50% lower computational latency without compromising accuracy.

  • Problem

    Existing MLLM uncertainty methods tend to quantify aleatoric uncertainty while underestimating epistemic uncertainty related to model limitations.

    Source: §1 · p. 2

  • Method

    CIM measures semantic shift between original predictions and a causally focused view, while EED estimates that shift directly in embedding space.

    Source: §1 · p. 2, §3.4 · p. 6

  • Results

    CIM achieves state-of-the-art performance, including 76.8% AUROC on AdVQA, while EED reduces computational latency by 50% without compromising accuracy.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    CIM supports hallucination detection and efficient adaptive inference for multimodal reasoning.

    Source: §5 · p. 10

    CIM depends on the quality of RoIs and equivalence estimation, whose failures can affect results, though their empirical impact is limited.

    Source: §5 · p. 10

Abstract

from arXiv · show

Multimodal Large Language Models (MLLMs) suffer from hallucinations, creating a critical need for Uncertainty Quantification (UQ) to ensure reliable deployment. However, existing approaches struggle to detect uncertainty caused by superficial associations, especially when the query-relevant signal is weak. We mainly attribute this issue to their bias toward aleatoric uncertainty arising from data ambiguity, overlooking epistemic uncertainty stemming from model limitations. To further decompose uncertainty types for a comprehensive UQ, we propose Causal-Invariant Masking (CIM), which measures the semantic shift between the original predictions and those conditioned on a causally-focused view. Based on this framework, we introduce Semantic Divergence as our core metric for UQ and provide theoretical evidence that it converges to the variance of model's sensitivity to non-causal correlations, establishing its ability to capture MLLM's limitation. To accelerate UQ in MLLMs, we further propose Expected Embedding Drift (EED), a fast geometric proxy metric that estimates semantic shift directly within the hyperspherical embedding space. Experiments show that our method achieves state-of-the-art performance on various benchmarks, while the proposed EED accelerates by nearly 50% with comparable performance.

1 Introduction

MLLM hallucinations undermine reliable deployment, while existing uncertainty methods under-detect epistemic uncertainty associated with model limitations. CIM compares predictions on the original input and a causally focused view, with Semantic Divergence as its core metric and EED as a faster proxy.

1 Introduction

  • Hallucinations threaten reliable MLLM deployment because misleading visual or linguistic factors can produce confidently stated false content.Examples include non-existent objects, attributes, or events in responses.Source: §1 · pp. 1–2
  • UQ estimates model confidence so systems can abstain or flag potential hallucinations; approaches include predictive entropy and sampling-based semantic uncertainty.Semantic Entropy aggregates probabilities over clusters of similar generations.Source: §1 · p. 2
  • Existing MLLM uncertainty methods are biased toward aleatoric uncertainty, leaving epistemic uncertainty from model limitations underestimated.Epistemic uncertainty reflects limits in perception or comprehension, while aleatoric uncertainty reflects poor data quality or incomplete information.Source: §1 · p. 2
  • CIM measures semantic shifts between original predictions and a causally focused view to quantify epistemic uncertainty, with EED as a faster proxy.The focused view retains causal features; EED estimates the shift directly in embedding space.Source: §1 · p. 2

2 Related Work

Prior work spans uncertainty estimation for LLMs and MLLMs, including sampling-based consistency and perturbation approaches. In MLLMs, CIM differs by using query-conditioned, RoI-preserving perturbations to retain answer evidence while suppressing irrelevant context.

2.1 Uncertainty Estimation in LLMs

  • LLM uncertainty work includes intrinsic, verbalized, and consistency-based methods, while general LLM approaches can be challenged by vision-language tasks.Examples include token-level perplexity, verbalized confidence, semantic entropy, and embedding-space dispersion measures.Source: §2.1 · p. 2, §2.2 · p. 2

2.2 Uncertainty Estimation in MLLMs

  • MLLM uncertainty work includes sampled-response consistency methods and perturbation-based approaches for hallucination detection and mitigation.Examples include UMPIRE, VL-Uncertainty, and Uncertainty-o.Source: §2.2 · pp. 2–3
  • Unlike global perturbations under full context, CIM uses query-conditioned, RoI-preserving perturbations to retain answer evidence while suppressing irrelevant context.Its comparison is intended to probe reliance on relevant versus irrelevant features, rather than variance under general information loss.Source: §2.2 · p. 3

3 Method

CIM measures semantic changes after masking query-irrelevant visual context, and EED approximates this shift in embedding space. The analysis connects Semantic Divergence to model sensitivity to non-causal features.

3.1 Preliminary

  • Semantic Entropy groups sampled outputs by meaning but does not separate uncertainty types; Figure 1 shows Semantic Divergence stays robust to blur and spikes under hallucinative background mixup.Source: Figure 1 · p. 3, §3.1 · p. 3

3.2 Epistemic Uncertainty Isolation via Causal Masking

  • CIM masks non-causal context with a query-conditioned RoI view and measures the KL shift between original and focused semantic distributions; large shifts indicate reliance on spurious context.Source: §3.2 · pp. 3–4

3.3 Theoretical Analysis

  • Under an energy decomposition with invariant causal energy, Semantic Divergence cancels ambiguity in causal features and, under bounded shifts, approximates variance in sensitivity to non-causal features.Source: §3.3 · p. 5

3.4 Fast Quantification via Expected Embedding Drift

  • EED estimates semantic shift through embedding drift, avoiding discrete clustering; under a spherical semantic-distribution assumption, it estimates KL divergence up to scale and an additive constant.The method aggregates token-level drift over sampling passes, with computational overhead reduced to O(K′) from O((K + K′)^2) entailment judgments.Source: §3.4 · pp. 5–6

4 Experiment

Experiments evaluate Usem and UEED on InternVL2-8B and Qwen2.5-VL-7B-Instruct across general VQA and hallucination-targeted benchmarks, with equal maximum sampling-time budgets for fair comparison. Their mixed uncertainty score interpolates the proposed values with Semantic Entropy to capture both aleatoric and epistemic uncertainty.

4.1 Results on Hallucination Detection

  • CIM leads the reported VQA, InternVL POPE, and MME comparisons; on AdVQA it reaches 76.8% AUROC versus Semantic Entropy (75.5%) and EigenScore (67.1%).On InternVL, the explicit method achieves approximately 75% AUROC across POPE splits and significantly outperforms Semantic Entropy (approximately 68–72%).Source: §4.1 · pp. 7–8
  • On QwenVL POPE Random, EED scores 83.2% versus 66.2% for the explicit method, while EED is often within 1% of Usem on VQA benchmarks.The authors attribute EED’s relative advantage on QwenVL POPE to operating in continuous embedding space, which is less affected by unstable discrete semantic clustering.Source: §4.1 · pp. 7–8

4.2 Discussions

  • RoI Crop achieves 81.3% AUROC on InternVL VQAv2, outperforming random and reversed crops by preserving relevant evidence while removing spurious context.Reversed Crop performs significantly worse than Random Crop in most tested cases; the authors say background-only input leaves no anchor for correct reasoning.Source: §4.2 · p. 8
  • In the SE-free pure-SD setting, RoI Crop improves over Random Crop by 6.5/9.2 AUROC on InternVL VQAv2/AdVQA and 11.6/14.0 on QwenVL.On GQA, RoI Crop is comparable to ground-truth RoI Crop, and both outperform the other crop variants.Source: §4.2 · p. 8
  • A balanced 40%–60% sampling allocation maximizes performance; skewed allocations degrade it, while the optimal proportion remains stable as the total budget increases.The figure shows QwenVL results on AdVQA and MME.Source: §4.2 · p. 8, Figure 3 · p. 9
  • UEED reduces per-sample computational cost by approximately 50% across both MLLM architectures and all four benchmarks, including QwenVL AdVQA dropping from approximately 9.6s to 4.8s.EED avoids NLI-based O((K + K′)^2)-time clustering for the perturbed input by operating directly in embedding space.Source: Figure 4 · p. 9, §4.2 · p. 9
  • Uncertainty-triggered re-answering yields accuracy gains across VQAv2, OKVQA, and AdVQA; on AdVQA, InternVL rises from 47.9% to 51.1% and QwenVL from 47.5% to 50.0%.The method re-generates answers when Usem exceeds a threshold.Source: Figure 5 · p. 9, §4.2 · p. 9
  • With adaptive resolution, processing QwenVL’s most confident 50% of samples at maximum resolution 256 cuts latency from 0.65s to 0.50s and raises accuracy from approximately 75.2% to 76.1%.The figure covers InternVL and QwenVL; reported useful thresholds span the 50th–70th percentiles.Source: Figure 6 · p. 9, §4.2 · pp. 9–10

5 Conclusion

The conclusion frames existing MLLM uncertainty quantification as biased toward aleatoric uncertainty and presents CIM as a way to target epistemic uncertainty from non-causal features. It reports state-of-the-art performance, faster EED, and a limitation tied to RoI quality and equivalence estimation.

5 Conclusion

  • CIM targets epistemic uncertainty by measuring semantic shifts under causally invariant perturbations, with a theoretical guarantee.Semantic Divergence is presented as targeting reliance on non-causal features.Source: §5 · p. 10
  • The authors report state-of-the-art performance and say EED reduces computational latency by 50% without compromising accuracy.Source: §5 · p. 10
  • The conclusion identifies bias toward aleatoric uncertainty as a gap in existing MLLM uncertainty quantification.Source: §5 · p. 10
  • CIM's results can be affected by the quality of its RoIs and equivalence estimation, although their empirical impact is reported as limited.Source: §5 · p. 10

A Proofs of Equations and Propositions

The appendix proofs compare original and masked-input distributions, relate Semantic Divergence to a Jensen gap and its variance approximation, and derive a KL expression for vMF distributions.

A.1 Proof of Lemma 3.2 (Aleatoric Cancellation)

  • The Lemma 3.2 proof defines distributions for the original and masked inputs, then subtracts their log-probability expressions.Source: §A.1 · p. 14

Part 1: Equivalence to Jensen Gap

  • The Jensen-gap proof identifies the divergence with a Jensen gap that is zero exactly when the shift is constant almost surely.Under a bounded shift, the Taylor remainder is bounded by O(ϵ^3), supporting the variance approximation.Source: Part 1: Equivalence to Jensen Gap · pp. 14–15
  • The proof states that the exact divergence is best interpreted in a low-hallucination setting where |∆s(s)| is relatively small.Source: Part 1: Equivalence to Jensen Gap · p. 14

A.3 Proof of Lemma 3.5 (KL Divergence of vMF Distributions)

  • For the vMF-distribution KL expression, the proof substitutes the expected vector E[z] = A_d(κ)µ_Q into the integral.Source: §A.3 · p. 15

A.4 Proof of Theorem 3.6 (Geometric Proxy Equivalence)

The proof relates Expected Embedding Drift to semantic divergence through an affine transformation. At high concentration, the relation becomes exactly linear, so minimizing EED strictly minimizes the divergence.

A.4 Proof of Theorem 3.6 (Geometric Proxy Equivalence)

  • At high concentration, the affine relation between EED and semantic divergence becomes exactly linear, so minimizing EED strictly minimizes the divergence.The offset depends only on concentration κ and dimension d.Source: §A.4 · p. 15

B Detailed Experiment Setup

The detailed setup fixes evaluation subsets and a shared decoding budget, then specifies CIM's sampling, interpolation, RoI transformation, and prompt procedures.

B Detailed Experiment Setup

  • The setup uses four NVIDIA RTX A6000 GPUs and supplies prompts for answering questions, extracting key objects, and clustering answer pairs by meaning.Source: B Detailed Experiment Setup · p. 16
  • GroundingDINO generates the causal-invariant RoI, with a central crop as fallback; task-specific or query-conditioned detectors may also be substituted.The reported setup uses a 0.3 detection threshold and rejects boxes covering more than 95% of the image.Source: B Detailed Experiment Setup · p. 16
  • Evaluation uses full POPE and MME test sets and the first 2,000 validation examples from each of VQAv2, OKVQA, AdVQA, and GQA.Figure 7 reports uncertainty estimation performance on MME.Source: B Detailed Experiment Setup · p. 16, Figure 7 · p. 17
  • All sampling-based methods use a total decoding budget of 32 sequences per query; CIM splits this into 16 original-view and 16 filtered-view samples.Source: B Detailed Experiment Setup · p. 16
  • CIM interpolates its uncertainty score with Semantic Entropy estimated from the 16 original-view samples, using λ = 0.05 to retain aleatoric sensitivity alongside CIM's epistemic signal.Source: B Detailed Experiment Setup · p. 16

C.1 Additional MME Results

CIM performs best on the additional MME hallucination-detection result and identifies confident errors that standard semantic entropy misses. Further analyses support evidence-preserving RoI masking, including under degraded detections, relational questions, and an alternative detector.

C.1 Additional MME Results

  • 74.7%: CIM achieves the best performance on MME hallucination detection, extending its reported generalization beyond the VQA datasets in the main tables.The result is reported in Figure 7.Source: §C.1 · p. 17

C.2 Case Study

  • CIM uncertainty is high on illustrated hallucinated samples even when standard semantic entropy remains low.Examples include confident misidentification from background text, color hallucination, and complex-object counting with a stable answer.Source: Figure 8 · p. 17, §C.2 · p. 17

C.3 Gaussian Noise Perturbation

  • Gaussian-noise perturbation performs close to random masking and below RoI masking, which supports preserving causal evidence rather than merely changing the image.The comparison reports final mixed uncertainty scores on VQAv2 and AdVQA.Source: Table 4 · p. 18, §C.3 · p. 18

C.4 Detector Failure Sensitivity

  • AUROC drops steadily as valid RoIs are replaced with random crops, while the detector fallback rate is low across benchmarks.Table 5 examines degraded RoIs on GQA; Table 6 reports fallback rates, where fallback uses a central crop when no valid RoI is detected.Source: Table 5 · p. 18, Table 6 · p. 18, §C.4 · p. 18

C.5 Relational and Global GQA Questions

  • CIM's gain is maintained on relational and global GQA questions, suggesting the union box preserves enough contextual structure for these questions.The union box retains the spatial region between query-relevant objects.Source: §C.5 · p. 18, Table 7 · p. 19

C.6 Alternative Grounding with Florence-2

  • Replacing GroundingDINO with query-conditioned Florence-2 produces only small changes, indicating CIM is not tied to a particular RoI detector.The comparison reports AUROC/AURAC.Source: Table 8 · p. 19, §C.6 · p. 19

C.7 Empirical Validations for Energy Assumptions

The empirical checks support the paper's energy assumptions: measured shifts are small relative to decision margins, and POPE output entropy differs between the evaluated models.

C.7 Empirical Validations for Energy Assumptions

  • The measured causal-energy shift is relatively small compared with the median top-1/top-2 logit gap, supporting the decomposition assumption.The comparison is reported for crops versus vacant-background views.Source: Table 9 · p. 19, §C.7 · p. 19
  • 0.50–1.00: median |∆s| on hallucinated samples is smaller than the 3.62–4.81 median decision margin, supporting the small-shift approximation.The authors estimate ϵ ≈ 1.0 and report that both ∥∆c∥ and ∥∆s∥ are substantially smaller than the corresponding decision margins.Source: §C.7 · p. 19, Table 10 · p. 20

C.8 POPE Entropy Distributions

  • On POPE, QwenVL has a sharper average token-level output distribution than InternVL, which the authors link to less stable discrete semantic clustering and EED's particular effectiveness for QwenVL.Table 11 reports average token-level entropy during decoding.Source: Table 11 · p. 20, §C.8 · p. 20

C.9 NLI Model Ablation

The NLI-model ablation compares the default Qwen-7B entailment model with DeBERTa-small. The reported changes are slight, and SD remains competitive and outperforms SE under the same NLI model.

C.9 NLI Model Ablation

  • The ablation compares the default Qwen-7B entailment model with a smaller DeBERTa-small model.Table 12 defines the reported SD difference as the metric-value difference between those two entailment-model settings.Source: Table 12 · p. 20, §C.9 · p. 20
  • Using DeBERTa-small slightly changes performance relative to the default entailment model.The comparison is reported in Table 12.Source: Table 12 · p. 20, §C.9 · p. 20
  • SD remains competitive and continues to outperform SE when both use the same NLI model.Source: §C.9 · p. 20

C.10 EED Proxy and Distance Metrics

The appendix validates EED as a proxy for semantic distribution shift and compares divergence choices, then records the reuse terms for evaluated assets.

C.10 EED Proxy and Distance Metrics

  • Mean EED increases with Semantic Divergence, supporting EED as an efficient proxy for semantic distribution shift.Figure 9 left groups samples by Usem and plots corresponding EED values.Source: §C.10 · p. 21
  • Performance is not highly sensitive to the tested distribution distance, though forward KL performs best in the experiments.Figure 9 right compares squared L2, Jensen-Shannon divergence, reverse KL, and forward KL.Source: §C.10 · p. 21

D Licenses for Existing Assets

Loading 2610.02887v1…