Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations

Hugging Face Daily Papers Papers

Summary

This paper systematically tests linear probes for deception detection in large language models, finding they fail under distributional shifts but style-augmented probes recover performance, and revealing that deception is encoded through distributed sub-threshold features.

Linear probes trained on LLM activations are increasingly proposed as deception-detection metrics, yet report AUROC exceeding 0.96 on clean benchmarks while collapsing under distributional shift. This paper systematically pressure-tests probe-based metrics across the Gemma 3 model family (1B-27B parameters), diagnosing why they fail rather than merely documenting that they fail. We test four hypotheses about deception encoding: (1) single linear direction, (2) multi-dimensional subspace, (3) convex conic hull, (4) entropy proxy. Our design includes cross-domain transfer matrices, multi-dimensional probe analysis with permutation null baselines, entropy-residualization tests, and distractor evaluations across 8 stylistic shifts. We find that: (a) probes achieve near-perfect AUROC (>=0.998) on clean data but collapse under stylistic shifts; style-augmented probes recover near-perfect detection (mean AUROC 0.979-0.983) on unseen styles; (b) the single-direction hypothesis is rejected (k=1 captures only 0.61-0.80 AUROC), with cross-domain transfer failure confirmed as geometric rather than layer-mismatch-driven; (c) the entropy-proxy hypothesis is rejected (max |rho|=0.454, max Delta-AUROC after residualization=0.004); and (d) deception does not form a significant linear subspace (per-domain k*=0), yet multi-dimensional probes (k>=5) recover the signal through distributed sub-threshold features. Probe fragility reflects distributional narrowness rather than an architectural limitation: style-augmented probes recover near-perfect detection at both 4B and 27B, establishing that the inverse scaling pattern is a training-distribution artifact rather than a genuine scale-dependent phenomenon.
Original Article
View Cached Full Text

Cached at: 06/03/26, 07:36 AM

Paper page - Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations

Source: https://huggingface.co/papers/2605.27958

Abstract

Linear probes for deception detection in large language models fail under distributional shifts despite high performance on clean data, revealing that deception is encoded through distributed sub-threshold features rather than simple linear directions.

Linear probestrained onLLM activationsare increasingly proposed asdeception-detection metrics, yet reportAUROCexceeding 0.96 on clean benchmarks while collapsing underdistributional shift. This paper systematically pressure-tests probe-based metrics across theGemma 3 model family(1B-27B parameters), diagnosing why they fail rather than merely documenting that they fail. We test four hypotheses about deception encoding: (1) single linear direction, (2) multi-dimensional subspace, (3)convex conic hull, (4)entropy proxy. Our design includescross-domain transfermatrices,multi-dimensional probe analysiswithpermutation null baselines,entropy-residualizationtests, and distractor evaluations across 8stylistic shifts. We find that: (a) probes achieve near-perfectAUROC(>=0.998) on clean data but collapse understylistic shifts; style-augmented probes recover near-perfect detection (meanAUROC0.979-0.983) on unseen styles; (b) the single-direction hypothesis is rejected (k=1 captures only 0.61-0.80AUROC), withcross-domain transferfailure confirmed as geometric rather thanlayer-mismatch-driven; (c) the entropy-proxy hypothesis is rejected (max |rho|=0.454, max Delta-AUROCafter residualization=0.004); and (d) deception does not form a significant linear subspace (per-domain k*=0), yet multi-dimensional probes (k>=5) recover the signal through distributed sub-threshold features. Probe fragility reflects distributional narrowness rather than an architectural limitation: style-augmented probes recover near-perfect detection at both 4B and 27B, establishing that the inverse scaling pattern is a training-distribution artifact rather than a genuine scale-dependent phenomenon.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2605\.27958

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.27958 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.27958 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.27958 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Scaling Trends for Lie Detector Oversight in Preference Learning

arXiv cs.AI

This paper scales the SOLiD lie-detector oversight method to larger LLMs (up to 405B parameters) and evaluates it in realistic preference-learning settings, finding that undetected deception decreases with model scale but that the method is sensitive to distribution shift between training data.

"Did you lie?" Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms

arXiv cs.AI

This paper evaluates four lie detection methods for language models across prompted lying and trained model organisms, finding that activation- and logprob-based detectors drop sharply on trained model organisms while a chain-of-thought judge remains strong. It introduces new testbeds and the Did-You-Lie (DYL) follow-up probe method, releasing datasets and model organisms.