DECOR: Auditing LLM Deception via Information Manipulation Theory

2026-05-18

AI Summary

This work addresses the challenge of detecting subtle deception in large language models (LLMs), which often manifests through omission of critical facts, topic shifting, or semantic obfuscation—behaviors poorly captured by existing black-box detection methods lacking fine-grained interpretability. To this end, the study introduces DECOR, a novel multi-agent framework that operationalizes information manipulation theory for LLM deception detection. DECOR decomposes inputs into atomic information units and evaluates each along four distinct dimensions to quantify manipulative behavior in model responses, yielding an interpretable manipulation profile that is aggregated into a global deception index. Validated across 15 mainstream LLMs and multiple real-world benchmarks, DECOR achieves state-of-the-art performance in both single- and multi-turn dialogues. Ablation studies confirm the contribution of each component, demonstrating a fine-grained, interpretable, and generalizable auditing mechanism for deceptive behaviors.

Abstract

Large language models can deceive by subtly manipulating truthful information -- omitting key facts, shifting focus, or obscuring meaning -- making such behavior difficult to detect. Existing black-box methods rely on coarse-grained judgments, offering limited interpretability and failing to pinpoint which facts were distorted and how. We introduce DECOR, a multi-agent framework grounded in Information Manipulation Theory for fine-grained auditing of strategic deception in LLM responses. DECOR decomposes input contexts into atomic informational units and scores each unit against the response across four dimensions of manipulation, producing interpretable manipulation profiles that are aggregated into a global deception index. We comprehensively evaluate DECOR on both single-turn and multi-turn deception detection benchmarks spanning real-world domains, and show that DECOR achieves state-of-the-art performance on both, outperforming competitive baselines. The framework generalizes across 15 frontier models, and ablation studies confirm the contribution of each key design component. Our findings demonstrate that fine-grained, theory-grounded auditing of information manipulation offers an effective and interpretable path for LLM deception detection.

Problem

• LLM deception
• information manipulation
• deception detection
• interpretability
• truthful AI

Innovation

• Information Manipulation Theory
• fine-grained auditing
• multi-agent framework
• deception detection
• interpretable manipulation profiles