2026-06-01原文2 min readUpdated: 2026-06-01

When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception

This study introduces a multi-model paradigm to study synthetic deception via LoRA fine-tuning of five transformer models. Linear probes detect deception with near-perfect AUC in early layers, and logistic regression probes outperform MLP probes, supporting the Linear Representation Hypothesis. Probes generalize across domains with minimal loss. Different models exhibit distinct representational regimes: collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2. The results show that robust, domain-invariant deception representations can be rapidly entrenched through modest supervised fine-tuning, with implications for activation-based monitoring.

SourcearXiv Machine LearningAuthor: Vahideh Zolfaghari

Article intelligence

EngineersAdvanced

Key points

Linear probes on mean-pooled hidden states detect synthetic dishonesty with near-perfect AUC (≥0.99) as early as layers 1-3 in four architectures. Logistic regression consistently matches or outperforms MLP probes.
Probes trained on TruthfulQA generalize with near-zero loss (ΔAUC≈0) to held-out MMLU subjects. Late-layer representations show strong robustness to Gaussian noise.
Mechanistic analysis reveals two regimes: representational collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2.
The dishonesty direction consolidates progressively in deeper layers, with optimal calibration (ECE<0.01 except Pythia) achievable in layers 1-4.

Why it matters

This matters because linear probes on mean-pooled hidden states detect synthetic dishonesty with near-perfect AUC (≥0.99) as early as layers 1-3 in four architectures. Logistic regression consistently matches or outperforms MLP probes.

Technical impact

May affect model selection, inference cost, product capability, and evaluation benchmarks.

[2605.30381] When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception

[Submitted on 28 May 2026]

Title:When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception

View a PDF of the paper titled When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception, by Vahideh Zolfaghari

View PDF

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.

Subjects:

Machine Learning (cs.LG); Artificial Intelligence (cs.AI)

Cite as: arXiv:2605.30381 [cs.LG]

(or arXiv:2605.30381v1 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2605.30381

arXiv-issued DOI via DataCite

Submission history

From: Vahideh Zolfaghari [view email] [v1] Thu, 28 May 2026 01:20:06 UTC (4,181 KB)

Full-text links:

Access Paper:

View a PDF of the paper titled When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception, by Vahideh Zolfaghari

View PDF

view license

Current browse context:

cs.LG

new | recent | 2026-05

Change to browse by:

cs cs.AI

References & Citations

NASA ADS

Google Scholar

Semantic Scholar

Data provided by:

Bibliographic Tools

Bibliographic and Citation Tools

Bibliographic Explorer Toggle

Bibliographic Explorer (What is the Explorer?)

Connected Papers Toggle

Connected Papers (What is Connected Papers?)

Litmaps Toggle

Litmaps (What is Litmaps?)

scite.ai Toggle

scite Smart Citations (What are Smart Citations?)

Code, Data, Media

Code, Data and Media Associated with this Article

alphaXiv Toggle

alphaXiv (What is alphaXiv?)

Links to Code Toggle

CatalyzeX Code Finder for Papers (What is CatalyzeX?)

DagsHub Toggle

DagsHub (What is DagsHub?)

GotitPub Toggle

Gotit.pub (What is GotitPub?)

Huggingface Toggle

Hugging Face (What is Huggingface?)

ScienceCast Toggle

ScienceCast (What is ScienceCast?)

Demos

Replicate Toggle

Replicate (What is Replicate?)

Spaces Toggle

Hugging Face Spaces (What is Spaces?)

Spaces Toggle

TXYZ.AI (What is TXYZ.AI?)