arXiv cs.CL· J\'er\'emie Dentan, Alexi Canesse, Mahammed El Sharkawy, Sonia Vanier·· 13 小时前AI 评分34
Kernel Token Contradiction：一种快速且有原则的 LLM 声明级不确定性量化方法
Kernel Token Contradiction: a Fast and Principled Approach for LLM Claim Uncertainty Quantification
AI 导读
研究者提出 Kernel Token Contradiction（KTC），一种在白盒条件下计算 LLM 声明级不确定性量化的轻量方法，将候选 token 表示为半正定核并结合 token 矛盾分数，用 Von Neumann 熵量化不确定性。
正文 · 原文
Abstract:Claim-level Uncertainty Quantification (UQ) aims to mitigate the lack of reliability of Large Language Models (LLMs) by evaluating the factuality of each claim in their outputs. We introduce Kernel Token Contradiction (KTC), a lightweight approach to compute claim-level UQ under realistic white-box conditions. KTC represents the candidate tokens involved in LLM generation as a positive semi-definite kernel that integrates both the LLM's conditional distribution and a token contradiction score. We then use the Von Neumann entropy to quantify the uncertainty of this kernel. To estimate token contradiction, we develop a new approach based on frequency statistics from the Wikipedia corpus. Although CPU-only, our approach achieves over an 8.2x speedup compared to state-of-the-art GPU-accelerated methods based on cross-encoders, and over a 65x speedup compared to CPU-only methods with comparable performance. Our evaluation spans two benchmarks across four European languages and 16 different models. KTC not only matches the average performance of existing methods but also outperforms them in high-precision regimes. This combination of computational efficiency and accuracy makes real-time monitoring of LLM outputs practical in production.
| Comments: | Accepted for publication at NeurIPS 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2608.22506 [cs.CL] |
| (or arXiv:2608.22506v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2608.22506 arXiv-issued DOI via DataCite |
Submission history
From: Jérémie Dentan [view email]
[v1]
Sun, 23 Aug 2026 16:55:02 UTC (474 KB)
[v2]
Tue, 27 Oct 2026 10:31:11 UTC (569 KB)
来源：arXiv cs.CL · arxiv.org