Contrastive representation learning for self-supervised deception detection in edge LLMs
PLOS One · 14.08.2026 · 21 (8) · e0354894
DOI: 10.1371/journal.pone.0354894 ↗Kurz gefasst
Fachartikel in der Zeitschrift PLOS One, online erschienen am 14.08.2026. Verfasst von Feng An und 1 weiteren Autor:innen. Beteiligte Einrichtung: Department of Biotechnology, Suzhou Industrial Park Institute of Services Outsourcing, Suzhou, Jiangsu, China. Hauptthemen laut MEDLINE: Deception. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Jiangsu Provincial Social Science Application Research Excellence Project Special Project on Science-ethics and Research Integrity, Suzhou Municipal Science and Technology Bureau, Jiangsu Provincial Education Science Planning Project.
Abstract (Originalsprache)
Existing deceptive alignment detection schemes generally follow a three-step strategy: auto-labeling, supervised fine-tuning (SFT), and proximal policy optimization (PPO). In which, the detection is treated as a simple binary classification and rely on heavyweight teacher models for Chain-of-Thought (CoT) annotation, limiting discrimination of nuanced deceptive strategies and creating an oracle dependency that prevents autonomous operation. This paper introduces contrastive representation learning , rather than learning a hard decision boundary (BCE loss), our lightweight monitor (0.1% parameters) projects CoT hidden states into a structured semantic space where deceptive and safe reasoning form separable manifolds. Through Triplet Loss optimization, the monitor captures gradual deceptive transitions, from surface hedging to fundamental objective substitution, that elude binary classifiers. Evaluation on Sycophancy subset of DeceptionBench confirms that contrastive learning outperforms BCE classification by 2.33pp Deception Tendency Rate (DTR, lower better, 39.29% vs. 36.96%). This establishes a geometric foundation for self-supervised deception detection, transforming CoT transparency from vulnerability into forensic evidence.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Autor:innen und Institutionen
- Department of Biotechnology, Suzhou Industrial Park Institute of Services Outsourcing, Suzhou, Jiangsu, China.
- College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu, China.
ORCID-Profile: Feng An
Förderung
- Jiangsu Provincial Social Science Application Research Excellence Project Special Project on Science-ethics and Research Integrity
- Suzhou Municipal Science and Technology Bureau
- Jiangsu Provincial Education Science Planning Project
- Jiangsu Provincial Vocational Education Teaching Reform Project
- Jiangsu Province Higher Education Teaching Reform Research Project
Zitieren
APA
An, F., & Tao, W. (2026). Contrastive representation learning for self-supervised deception detection in edge LLMs. PLOS One, 21(8), e0354894. https://doi.org/10.1371/journal.pone.0354894
Vancouver
An F, Tao W. Contrastive representation learning for self-supervised deception detection in edge LLMs. PLOS One. 2026;21(8):e0354894. doi:10.1371/journal.pone.0354894
BibTeX
@article{An2026,
author = {An, Feng and Tao, Wenyin},
title = {Contrastive representation learning for self-supervised deception detection in edge LLMs},
journal = {PLOS One},
year = {2026},
volume = {21},
number = {8},
pages = {e0354894},
doi = {10.1371/journal.pone.0354894}
}
EUVIMED