The paper introduces Generative Score Inference (GSI), a flexible framework for uncertainty quantification in multimodal tasks like image captioning and LLM hallucination detection. GSI utilizes conditional deep generative models—specifically diffusion models—to estimate conditional score distributions, achieving state-of-the-art (SOTA) performance in constructing statistically valid and informative prediction sets.
TL;DR
Reliable decision-making in AI requires more than just high accuracy; it requires knowing when to trust a model. Generative Score Inference (GSI) is a new powerhouse framework that uses conditional diffusion models to quantify uncertainty. By accurately mapping the distribution of "prediction errors" (scores), GSI provides tighter, sample-specific confidence intervals for everything from medical tabular data to LLM-generated text and image captions.
The Problem: The "Average" is Not Enough
Most modern AI models are evaluated on average performance. However, in safety-critical fields, we need conditional guarantees.
Standard Conformal Prediction (CP) ensures that if a model says it is 90% confident, it will be right 90% of the time on average. But for a specific, difficult input, it might only be 20% right. This "marginal coverage" is a major loophole. Multimodal data (images + text) makes this worse because the relationship between inputs and outputs is non-linear, high-dimensional, and full of heterogeneous noise.
The Insight: Modeling the "Score," Not the "Data"
The authors of GSI propose a clever pivot. Instead of trying to model the complex multi-modal output space directly to find uncertainty, they model the Score Function $s(y, \hat{f}(x))$.
A score measures the "dissimilarity" between the truth ($y$) and the prediction ($\hat{y}$). By training a Conditional Diffusion Model to learn the distribution of these scores given an input $x$, GSI can:
- Generate synthetic error samples: "Imagine" 1,000 possible error levels for a specific input.
- Calculate Local Quantiles: Use those 1,000 samples to find the exact 95th percentile error for this specific image or this specific question.
Methodology: Why Diffusion?
While GSI is generator-agnostic, the authors emphasize Conditional Diffusion Models. Unlike GANs (which suffer from mode collapse) or VAEs (which produce blurry results), Diffusion models excel at capturing "heteroscedastic" noise—situations where some parts of the data are much noisier than others.
The Workflow:
- Split: Use part of the data to train the predictor and the rest for "calibration."
- Score: Calculate residuals (errors) on the calibration set.
- Train Generator: Train a diffusion model to predict these error scores based on the input features.
- Infer: For a new test case, sample the diffusion model to define the "Safe Zone" (Prediction Set).
Experiments: Catching the Hallucination
One of GSI's most impressive feats is in LLM Hallucination Detection. Current methods like "Semantic Entropy" look for consistency—if an LLM gives three different answers, it's probably hallucinating. But what if the LLM is consistently wrong?
GSI anchors its scores to verified references. In the WikiQA benchmark, GSI outperformed Semantic Entropy (SE) because it could detect cases where the model was confident but factually misaligned with the truth.
Figure: GSI (blue) maintains better Type I error control and significantly higher statistical power than baselines.
Key Results:
- Tabular Data: GSI produced intervals 10-20% narrower than CUQR while providing better coverage for "difficult" subgroups.
- Image Captioning: On the MS-COCO dataset, GSI exhibited higher statistical power in selecting "reliable" captions compared to the Conformal Alignment baseline.
Critical Analysis & Future Outlook
GSI is a significant step toward Trustworthy AI, but it comes with a "Generation Tax."
- Computational Cost: Sampling from a diffusion model at inference time is slower than a simple regression pass. However, as the authors note, it is still faster than generating 10+ LLM responses for entropy-based methods.
- Reference Dependency: For hallucination detection, GSI still needs a "calibration set" with some ground truth.
The Takeaway: GSI demonstrates that generative models aren't just for making pretty pictures or chatting; they are sophisticated statistical tools that can provide rigorous "safety certificates" for AI outputs. As diffusion samplers become faster (e.g., DPM-Solver++), GSI could become the standard layer for real-time uncertainty quantification in foundation models.
Conclusion
By bridging the gap between deep generative modeling and classical statistical inference, GSI offers a path forward for multimodal systems that are not just "smart," but provably reliable.
Reference: Tian, X. & Shen, X. (2026). Generative Score Inference for Multimodal Data. University of Minnesota.