ニュース
D-Score: A Spectral Hidden-State Signal for Hallucination Detection in Large Language Models
arXiv cs.AI · 公開日 · 読了3分
30秒で要点
- 何が起きたか
- Researchers introduced D-Score, a spectral method to detect hallucinations in large language models by analyzing hidden activation patterns from a single forward pass.
- なぜ重要か
- Matters for engineers building LLM applications who need hallucination detection without external verifiers, retrieval systems, or multiple model generations.
- 注意点
- Paper is under review and not yet peer-reviewed. Evaluation limited to two datasets. Requires per-model and per-layer calibration with tolerance parameters.
この要約を音声で聴く
記事より
-->
Computer Science > Computation and Language
arXiv:2607.24586v1 (cs)
[Submitted on 27 Jul 2026]
Title: D-Score: A Spectral Hidden-State Signal for Hallucination Detection in Large Language Models
Authors: Bianca Raimondi , Davide Evangelista , Maurizio Gabbrielli , Elena Loli Piccolomini
View a PDF of the paper titled D-Score: A Spectral Hidden-State Signal for Hallucination Detection in Large Language Models, by Bianca Raimondi and 3 other authors
View PDF HTML (experimental)
Abstract: Large Language Models can produce fluent text that is false, unsupported by the available evidence, or inconsistent with information that appears to be internally represented by the model. We study hallucination detection from the geometry of hidden activations and introduce the D-Score, a simple spectral statistic computed from a single forward pass. For a fixed model, layer, and tolerance parameter, the D-Score counts how many singular directions of the hidden activation matrix have singular values that remain close to the leading one. We use this quantity as a hallucination score, classifying an input text as hallucinated when its D-Score is larger than a pre-defined quantity. The motivation is that, when a model processes a text that conflicts with information available in its own internal state, the hidden representation may encode both the asserted content and some form of count
原文からの抜粋です。全文は配信元でお読みください。
- language model
The Agent Architect
1つのパターン、1つのトレードオフ、1つの本番障害事例。エージェントシステムを構築する人のための短い週刊ブリーフィング。
週1回のメール、ワンクリックで購読解除できます。アドレスはブリーフィングの送信のみに使用します。