Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Semantic Energy: Detecting LLM Hallucination Beyond Entropy

Huan Ma, Jiadong Pan, Jing Liu, Yan Chen, Joey Tianyi Zhou, Guangyu Wang, Qinghua Hu, Hua Wu, Changqing Zhang, Haifeng Wang

TL;DR

This work tackles hallucinations in large language models by advancing uncertainty estimation beyond entropy-based semantics. It introduces Semantic Energy, an energy-based framework operating on penultimate-layer logits and semantic clusters, anchored by a Boltzmann-inspired formulation to capture epistemic uncertainty. Empirical results on CSQA and TriviaQA across Qwen-3-8B and ERNIE-21B-A3B demonstrate superior hallucination-detection signals and better calibration versus semantic entropy, including think-mode scenarios. The approach enhances reliability for downstream uncertainty-driven tasks and underscores the value of leveraging internal model states to quantify uncertainty more faithfully.

Abstract

Large Language Models (LLMs) are being increasingly deployed in real-world applications, but they remain susceptible to hallucinations, which produce fluent yet incorrect responses and lead to erroneous decision-making. Uncertainty estimation is a feasible approach to detect such hallucinations. For example, semantic entropy estimates uncertainty by considering the semantic diversity across multiple sampled responses, thus identifying hallucinations. However, semantic entropy relies on post-softmax probabilities and fails to capture the model's inherent uncertainty, causing it to be ineffective in certain scenarios. To address this issue, we introduce Semantic Energy, a novel uncertainty estimation framework that leverages the inherent confidence of LLMs by operating directly on logits of penultimate layer. By combining semantic clustering with a Boltzmann-inspired energy distribution, our method better captures uncertainty in cases where semantic entropy fails. Experiments across multiple benchmarks show that Semantic Energy significantly improves hallucination detection and uncertainty estimation, offering more reliable signals for downstream applications such as hallucination detection.

Semantic Energy: Detecting LLM Hallucination Beyond Entropy

TL;DR

This work tackles hallucinations in large language models by advancing uncertainty estimation beyond entropy-based semantics. It introduces Semantic Energy, an energy-based framework operating on penultimate-layer logits and semantic clusters, anchored by a Boltzmann-inspired formulation to capture epistemic uncertainty. Empirical results on CSQA and TriviaQA across Qwen-3-8B and ERNIE-21B-A3B demonstrate superior hallucination-detection signals and better calibration versus semantic entropy, including think-mode scenarios. The approach enhances reliability for downstream uncertainty-driven tasks and underscores the value of leveraging internal model states to quantify uncertainty more faithfully.

Abstract

Large Language Models (LLMs) are being increasingly deployed in real-world applications, but they remain susceptible to hallucinations, which produce fluent yet incorrect responses and lead to erroneous decision-making. Uncertainty estimation is a feasible approach to detect such hallucinations. For example, semantic entropy estimates uncertainty by considering the semantic diversity across multiple sampled responses, thus identifying hallucinations. However, semantic entropy relies on post-softmax probabilities and fails to capture the model's inherent uncertainty, causing it to be ineffective in certain scenarios. To address this issue, we introduce Semantic Energy, a novel uncertainty estimation framework that leverages the inherent confidence of LLMs by operating directly on logits of penultimate layer. By combining semantic clustering with a Boltzmann-inspired energy distribution, our method better captures uncertainty in cases where semantic entropy fails. Experiments across multiple benchmarks show that Semantic Energy significantly improves hallucination detection and uncertainty estimation, offering more reliable signals for downstream applications such as hallucination detection.

Paper Structure

This paper contains 24 sections, 14 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Estimating LLM Uncertainty with Token-Level Entropy
  4. Semantic Entropy and Response Clustering
  5. Modeling Uncertainty via Semantic Energy
  6. Limitations of Entropy-Based Uncertainty Estimation
  7. Energy-Based Confidence Estimation
  8. Boltzmann Distribution
  9. Specific Implementation in LLMs
  10. Experiments
  11. Setup
  12. Model & Baseline.
  13. Datasets & Metrics.
  14. Main Results
  15. Ablation Studies
  16. ...and 9 more sections

Figures (3)

  • Figure 1: An intuitive comparison between Semantic Entropy and Semantic Energy in their ability to characterize uncertainty. Both approaches first sample and perform semantic clustering over distinct clusters of activity. The difference lies in the computation: Semantic Entropy is calculated based on normalized probabilities, while Semantic Energy is derived from logits. This enables Semantic Energy to distinguish cases when Semantic Entropy fails.
  • Figure 2: Comparison of semantic vs. non-semantic uncertainty modeling on TriviaQA and CSimpleQA datasets.
  • Figure 3: Comparison of semantic entropy vs. semantic energy on CSQA datasets with think mode on.