Skip to content

When Do Supervised UQ Ensembles Improve LLM Hallucination Detection? A Robustness Study

Aug 2026 · 0 citations · 38 references
Computer Science

TL;DR

A supervised ensembling framework that trains a classifier over heterogeneous UQ-based scorer outputs on a small, domain-specific dataset of labeled LLM responses, then applies it to out-of-sample hallucination classification without retrieval, tools, or reference documents is studied.

Abstract

Uncertainty quantification (UQ) methods are widely used for hallucination detection in large language models (LLMs) in closed-book settings where ground-truth evidence is unavailable at inference time. Prior work has proposed combining UQ signals via learned ensembles, but empirical investigations into the robustness of these ensembles are limited. We study a supervised ensembling framework that trains a classifier over heterogeneous UQ-based scorer outputs on a small, domain-specific dataset of labeled LLM responses, then applies it to out-of-sample hallucination classification without retrieval, tools, or reference documents. Across four LLMs, nine datasets, and three generation regimes (short-form QA, long-form generation, and code generation), we provide a systematic robustness analysis along three axes: sample efficiency, in-domain dataset transfer, and generation regime dependence. We find that supervised ensembles outperform the best individual scorer in 30 of 32 settings, with gains realized from as few as 100 labeled instances. Ensembles retain most of their advantage in cases of in-domain transfer under distribution shift, outperforming the best non-ensemble scorer in 23 of 28 transfer settings. Sampling-based black-box ensembles are nearly as effective as full ensembles, while single-generation white-box ensembles offer limited benefit.

View source

Similar papers

Book Open access Aug 2026

Training a Generalist Hallucination Detector across Multiple Domains via Adaptive Layer Aggregation

Stability-Aware Adaptive Layer Aggregation (SALA) defines a layer-wise instability index computed from the training data, which measures how separability varies across domains and provides an upper bound on the possible separability drop from training to test.

Xinyi Li, Zhen Fang, Yadan Luo et al. · 0 citations
Preprint Aug 2026

Beyond Global Editing: Per-Instance Disentangled Subspaces for Training-Free Hallucination Mitigation in LVLMs

This work proposes a training-free hallucination mitigation framework for dynamic, per-instance suppression at test time, and proposes a dynamically combined projection that selectively suppresses the most probable hallucination directions while preserving image-grounded semantics.

Ali Cheraghian, Hamidreza Dastmalchi, Hamed Barzamini et al. · 0 citations
Preprint Aug 2026

UHP Detection: LVLMs have their Unique Hallucination Pattern in the Consistency Space

Large vision--language models (LVLMs) demonstrate strong multimodal reasoning capabilities but remain prone to hallucination, where model predictions are not grounded in visual evidence, so a fully black-box framework that models hallucination as a structured uncertainty pattern is proposed.

Amir Mohammad Ezzati, Kiyan Rezaee, Bardiya Kariminia et al. · 0 citations
Preprint Aug 2026

Test-Time Hallucination Control in Large Vision-Language Models

Object Hallucination in large vision-language models (LVLMs), where models generate non-factual content about input images, remains a critical barrier to their reliability in real-world applications. Existing mitigation strategies can be categorized into training-based and training-free methods. Training-based methods often achieve strong performance but are costly, requiring extensive computational resources, large-scale data, and time-consuming fine-tuning. Training-free approaches are particularly appealing due to their efficiency. However, existing training-free methods either require multiple decoding rounds, which adds computational overhead, or modify internal states in a model-specific way that risks degrading pretrained knowledge. We propose Test-Time Hallucination Mitigation (TTH) method, a novel training-free method that addresses both limitations. TTH introduces a token-validator module, implemented as a zero-shot Multi-Modal Classifier (MMC), to generate auxiliary logits grounded in the input image. These logits are fused with the original LVLM outputs at the token level for object tokens selected from a candidate pool. An entropy-based weighting scheme is then applied to enable robust and accurate predictions. Extensive experiments across multiple LVLM families and diverse benchmarks demonstrate that TTH consistently improves accuracy and robustness, underscoring its generalizability and practical effectiveness. Code is released at https://github.com/Mehran-TAM/TTH

Mehran Tamjidi, Hamidreza Dastmalchi, Ali Cheraghian et al. · 0 citations
Conference Jul 2026

HaloGuard: A Practical Hybrid Framework for Hallucination Detection in Retrieval-Augmented Generation Systems : Combining Multi-Judge Agreement, Uncertainty Quantification, and Lexical Features Without Embedding Dependencies

The issue of hallucination detection for Retrieval-Augmented Generation (RAG) systems is an ongoing reliability challenge that arises with the scale up of LLM deployment. We introduce the HaloGuard: a systems-integration study that integrates three types of signals, multi-judge LLM consensus, uncertainty quantification through temperature sampling, and lexical text features, and evaluate with LibreEval (10,871 samples; GPT-4o, Claude-3.5-Sonnet, Llama-3.1-8B). They all rely on previous work; the novel aspect of the present paper is their principled combination, empirical decomposition through ablation, calibration analysis, and deployment characterization over operating thresholds, hallucination types, and six languages At a conservative threshold, HaloGuard achieves AUC-ROC = 0.986 [95% CI: 0.981-0.991] and F1 = 0.665 [0.639-0.690]; at the balanced max-F1 threshold, F1 = 0.769 (Precision = 0.812 / Recall = 0.731), directly addressing the precision-1.000 operating-point artifact. A fine-tuned variant (HaloGuard-FT, DeBERTa-v3-base) reaches F1 = 0.847 [0.831-0.863] and AUC = 0.994 [0.991-0.997]. Weight sensitivity analysis confirms F1 is stable across judge weight alpha in [0.5, 0.7] (F1 range <= 0.016), ruling out weight cherry-picking. Additional analyses quantify a 2.15x novelty differential (p < 0.001), source-attribution correlation (r = -0.319), and systematic cross-lingual F1 degradation (EN: 0.665 -> JA: 0.543). Calibration is well-behaved (ECE = 0.043). All results are scoped to LibreEval; cross-dataset validation on RAGTruth is the stated priority next step. Full judge prompts, hyperparameters, and code are documented for reproducibility.

Harshath R, G. S, S. Thangavel et al. · 0 citations

Related blog posts