Technical Summary: FUSE – Quantifying Uncertainty in Vision-Language Models

Problem Statement

Vision-Language Models (VLMs) have become integral to applications ranging from robotics to healthcare, yet they remain prone to unreliable outputs such as hallucinations and overconfident reasoning. A critical gap exists in Uncertainty Quantification (UQ) for these models. Conventional methods relying on next-token likelihoods or logit-based confidences fail to capture the semantic ambiguity inherent in cross-modal inputs or the variability of generative outputs. Existing approaches often treat uncertainty heuristically rather than as a principled probabilistic quantity, and they frequently overlook uncertainty embedded in the model's latent representations.

The paper identifies that VLM uncertainty arises from two distinct, complementary sources:

1. Aleatoric Uncertainty: Intrinsic ambiguity in the input data and cross-modal semantics (e.g., underspecified questions or visually ambiguous regions).
2. Epistemic Uncertainty: Variability in the model's predictions due to limitations in knowledge, weak grounding, or distribution shifts, reflected in the diversity of generated responses.

The goal is to develop a framework that fuses these two sources into a single, calibrated scalar measure of uncertainty ($u^\star$) to predict output correctness and enable selective prediction (abstaining from uncertain queries).

Methodology: The FUSE Framework

The authors propose FUSE (Fused Uncertainty with Semantic Evidence), a probabilistic framework that employs a Bayesian fusion mechanism to combine aleatoric and epistemic uncertainties.

1. Aleatoric Uncertainty: Data Representation Prior

The framework treats the input data level as a prior.

• Probabilistic Embeddings: While standard VLMs (e.g., CLIP) produce deterministic embeddings, FUSE introduces a Gaussian Process (GP) adapter. This adapter maps deterministic embeddings ($z_I, z_T$) into probabilistic distributions ($N(\mu, \Sigma)$) by learning a shared latent variable $x$ representing the "true" meaning of the data.
• Sparse Variational GP: To ensure scalability, the authors use a sparse variational Gaussian Process Latent Variable Model (GPLVM) with inducing points.
• Prior Formulation: The dispersion of these probabilistic embeddings (quantified via log-determinant of covariances) is aggregated into a scalar statistic $s$. This statistic defines a Gaussian prior over a latent uncertainty variable $u$:
  $$u \sim N(m_0, \sigma_0^2), \quad \text{where } m_0 = \alpha_0 + \beta_0 s$$
  This captures the inherent ambiguity of the input query before generation occurs.

2. Epistemic Uncertainty: Semantic Response Evidence

The framework treats the model's generative output as evidence.

• Response Diversity: The VLM generates $n$ stochastic responses. These are embedded into vectors $\{r_i\}$.
• Semantic Scatter: A semantic scatter matrix $S$ is constructed from these embeddings. The semantic evidence is defined as the log-determinant of this matrix: $\tilde{V} = \log \det S$.
• Weighted Evidence: To account for varying reliability among samples, the method introduces uncertainty weights based on the model's per-token log-probabilities. Lower confidence samples receive higher weights.
• Likelihood Model: The authors prove that under mild assumptions, the semantic evidence $\tilde{V}$ follows a linear-Gaussian likelihood with respect to the latent uncertainty $u$:
  $$\tilde{V} | u \sim N(a + n u, \tau^2)$$
  This analytical relationship allows for a closed-form Bayesian update.

3. Bayesian Fusion

The core innovation is the analytical fusion of the input prior and the response evidence.

• Closed-Form Posterior: Combining the Gaussian prior and the linear-Gaussian likelihood yields a closed-form posterior distribution for the uncertainty $u$:
  $$u | \tilde{V} \sim N(m_{post}, \sigma_{post}^2)$$
• Fused Statistic: A single uncertainty score $u^\star$ is derived from the posterior mean and variance (e.g., as a z-score bound).
• Calibration: A global threshold $\theta^\star$ is calibrated on a held-out dataset to control the conditional risk (error rate) for answered queries, enabling selective answering.

Key Contributions

1. Conceptual Formalization: The paper formally disentangles VLM uncertainty into aleatoric (input-level) and epistemic (generation-level) sources, providing a foundation for quantifying both within a unified framework.
2. Unified Probabilistic Model: It introduces a training-free (for the base VLM) Bayesian framework that analytically combines these sources. Unlike heuristic methods, it yields an interpretable, calibrated posterior.
3. Mathematical Justification: The authors provide proofs demonstrating that the semantic evidence (log-determinant of response scatter) is linearly dependent on the latent uncertainty, justifying the use of a closed-form Bayesian update.
4. Empirical Superiority: The method achieves state-of-the-art (SOTA) performance in uncertainty calibration and selective answering across multiple benchmarks.

Experimental Results

The authors evaluated FUSE on three Visual Question Answering (VQA) benchmarks: VQAv2, OKVQA, and AdVQA, using LLaVA-v1.5-13b as the backbone.

• Baselines: FUSE was compared against training-free methods (UMPIRE, Neighborhood Consistency), entropy-based metrics (Semantic Entropy, LN-Entropy), and conformal prediction approaches (LoFreeCP, CLM).
• Discrimination: FUSE achieved the highest AUROC (0.883 on VQAv2) in distinguishing correct from incorrect answers, outperforming all baselines.
• Calibration: The method demonstrated superior calibration, with the lowest Expected Calibration Error (ECE) and highest correlation between predicted uncertainty and empirical error rates.
• Selective Answering: In terms of the Area Under the Rejection–Accuracy Curve (AURAC), FUSE consistently outperformed baselines, indicating its ability to effectively identify and abstain from uncertain queries while maintaining high accuracy on answered ones.
• Ablation Studies: Experiments confirmed that fusing both uncertainty sources is critical; removing either the GP prior or the weighted evidence resulted in performance degradation.

Significance and Claims

The paper claims that FUSE addresses a critical safety and reliability gap in deploying VLMs. By moving beyond token-level likelihoods to a joint Bayesian formulation that integrates input representation priors with generation-level evidence, FUSE provides a rigorous, interpretable measure of uncertainty.

The authors emphasize that their approach:

• Does not require retraining the massive VLM backbone, only training lightweight GP adapters on the frozen encoder.
• Provides a single scalar score that reliably predicts model correctness.
• Enables risk-controlled decision-making, allowing systems to abstain from answering when uncertainty exceeds a calibrated threshold, which is essential for high-stakes domains like healthcare and autonomous systems.

The work positions itself as a step toward making VLMs more trustworthy by explicitly modeling what the model "does not know" through both the ambiguity of the input and the variability of its own generation.