🤖 AI Summary
This study investigates how large language models retain accurate internal representations of truthful information even after fine-tuning induces systematic deceptive outputs. Using LoRA-based fine-tuning, we construct paired honest and deceptive variants across five Transformer models and train linear probes to detect deception signals in hidden states. For the first time in a multi-model setting, we demonstrate that deceptive representations are linearly separable, with deception directions progressively consolidating in deeper layers. Notably, models such as Gemma-2 stably preserve these high-dimensional representations without collapse. Our experiments reveal that near-perfect deception detection (AUC ≥ 0.99) is achievable using only the first one to three layers’ hidden states. Furthermore, probes trained on TruthfulQA generalize zero-shot to unseen MMLU subjects, and deep-layer representations exhibit strong robustness against Gaussian noise.
📝 Abstract
Deceptive alignment, in which models maintain accurate internal representations while deliberately producing false outputs, remains a central challenge in AI safety. While strategic deception is the primary long-term concern, synthetic dishonesty - induced via direct optimization on incorrect answers - provides a controlled testbed for studying the representational basis of learned deception. We introduce a multi-model paradigm in which honest and deceptive variants of five transformer models (Pythia-1.4B, Gemma-2-2B/9B, Qwen2.5-7B, Llama-3.1-8B) are fine-tuned using LoRA on the same question distribution. Linear probes trained on mean-pooled hidden states detect synthetic dishonesty with near-perfect AUC (greater than or equal to 0.99) as early as layers 1-3 in four architectures, while Pythia-1.4B reaches a peak of 0.705. Logistic regression probes consistently match or outperform MLP probes, supporting the Linear Representation Hypothesis. Probes trained on TruthfulQA generalize with near-zero loss (Delta AUC approx. 0) to held-out MMLU subjects. Late-layer representations show strong robustness to Gaussian noise, with Gemma-2 models exhibiting exceptional stability. Mechanistic analysis of Fisher Discriminant Ratio, effective rank, centroid geometry, directional stability, cross-domain alignment, and calibration (ECE) reveals two regimes: representational collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2. Across all models, the dishonesty direction consolidates progressively in deeper layers, with optimal calibration (ECE less than 0.01 except Pythia) achievable in layers 1-4. These results demonstrate that robust, domain-invariant dishonesty representations can be rapidly entrenched via modest supervised fine-tuning, with implications for activation-based monitoring.