FLaG: Fine-Grained Latent Grouping for Hallucination Detection

2026-05-29

AI Summary

This work addresses the challenge of hallucination detection in large language models, which stems from heterogeneous error mechanisms that cannot be effectively captured by a single global uncertainty measure. The authors formulate hallucination detection as a mechanism-aware evidence aggregation problem and propose a lightweight framework based on energy-based routing and log marginal likelihood aggregation. Their approach softly assigns inputs to multiple latent evidence groups and integrates group-conditional reliability signals. Theoretically, this is the first study to characterize the structure of the Bayes-optimal detection statistic under heterogeneous error mechanisms and to provide an efficient approximation with controllable error. Experiments demonstrate that the proposed framework achieves state-of-the-art performance across multiple benchmarks and large language models, exhibits strong generalization across datasets and models, and remains effective even under low-supervision conditions.

Abstract

Hallucinations in large language models (LLMs) arise from heterogeneous failure mechanisms, making reliable detection difficult for any single global uncertainty score. In this work, we formulate hallucination detection as a mechanism-aware evidence aggregation problem, where diverse representation- and token-level signals must be interpreted under multiple latent explanations. We propose FLaG, a lightweight hallucination detection framework that models correctness through a set of latent evidence groups. Each instance is softly associated with multiple groups via an energy-based routing mechanism, and group-conditional reliability signals are combined through a principled log-marginal aggregation. This design enables FLaG to capture heterogeneous hallucination patterns while remaining invariant to decision thresholds and evaluation metrics. The framework operates as a frozen-model head, requires no modification to the underlying language model, and incurs minimal computational overhead. We further provide a theoretical perspective that connects FLaG to optimal evidence aggregation under heterogeneous error mechanisms, showing that the Bayes-optimal test statistic necessarily admits a log-marginal form and that FLaG constitutes a tractable approximation with a controllable error bound. Extensive experiments across multiple benchmarks and LLM backbones demonstrate that FLaG consistently achieves SOTA performance, while exhibiting robust transfer across datasets and models, and remaining effective under limited supervision.

Problem

• hallucination detection
• large language models
• heterogeneous failure mechanisms
• uncertainty estimation
• evidence aggregation

Innovation

• latent grouping
• hallucination detection
• evidence aggregation
• energy-based routing
• log-marginal likelihood