analiza aiqbench
This research proposes a training-free method that enhances hallucination detection in Large Language Models by perturbing model parameters or hidden unit activations during sampling. By leveraging Bayesian uncertainty rather than standard token distribution sampling, the approach significantly improves inference-time reliability across diverse architectures. This technique offers a practical, low-overhead solution for mitigating the risk of plausible but incorrect outputs in deployed systems.
Enhancing Hallucination Detection through Noise Injection
arXiv:2502.03799v4 Announce Type: replace Abstract: Large Language Models (LLMs) are prone to generating plausible yet incorrect responses, known as hallucinations. Effectively detecting hallucinations is therefore crucial for the safe deployment of LLMs. Recent research has linked hallucinations to model uncertainty, suggesting that hallucinations can be detected by measuring dispersion over answer distributions obtained from multiple samples drawn from a model. While drawing from the distribution over tokens defined by the model is a natural way to obtain samples, in this work, we argue that it is suboptimal for the purpose of detecting hallucinations. We show that detection can be improved significantly by taking into account model uncertainty in the Bayesian sense. To this end, we propose a very simple, training-free approach based on perturbing an appropriate subset of model parameters, or equivalently hidden unit activations, during sampling. We demonstrate that our approach significantly improves inference-time hallucination detection over standard sampling across diverse datasets, model architectures, and uncertainty metrics.
Abstract:Large Language Models (LLMs) are prone to generating plausible yet incorrect responses, known as hallucinations. Effectively detecting hallucinations is therefore crucial for the safe deployment of LLMs. Recent research has linked hallucinations to model uncertainty, suggesting that hallucinations can be detected by measuring dispersion over answer distributions obtained from multiple samples drawn from a model. While drawing from the distribution over tokens defined by the model is a natural way to obtain samples, in this work, we argue that it is suboptimal for the purpose of detecting hallucinations. We show that detection can be improved significantly by taking into account model uncertainty in the Bayesian sense. To this end, we propose a very simple, training-free approach based on perturbing an appropriate subset of model parameters, or equivalently hidden unit activations, during sampling. We demonstrate that our approach significantly improves inference-time hallucination detection over standard sampling across diverse datasets, model architectures, and uncertainty metrics.
Computation and Language (cs.CL); Systems and Control (eess.SY)
arXiv:2502.03799 [cs.CL]
(or arXiv:2502.03799v4 [cs.CL] for this version)
Text integral sub licența deschisă a publicației · 169 cuvinte
Subiecte conexe
Mitigating Hallucination in Large Language Models: A Capability-Oriented Survey on RAG, Reasoning, and Agentic Systems
arXiv · 30 sept. 2026
UniPASE: A Generative Model for Universal Speech Enhancement with High Fidelity and Low Hallucinations
arXiv · 5 aug. 2026
StuPASE: Towards Low-Hallucination Studio-Quality Generative Speech Enhancement
arXiv · 5 aug. 2026
Understanding and Mitigating Hallucination Escape in Tool-Using LLM Agents
arXiv · 7 oct. 2026