mlsanityData to 14 Oct 2026

Linear Probe Accuracy Scales with Model Size and Benefits from Multi-Layer Ensembling

arXiv 2604.13386 ↗First posted 15 Apr 2026cs.LGcited by 4

On the map: LLM alignment / RLHF / preference optimisation → Safety alignment and jailbreak attacks

also in 1 more
Abstract (arXiv)

Linear probes can detect when language models produce outputs they "know" are wrong, a capability relevant to both deception and reward hacking. However, single-layer probes are fragile: the best layer varies across models and tasks, and probes fail entirely on some deception types. We show that combining probes from multiple layers into an ensemble recovers strong performance even where single-layer probes fail, improving AUROC by +29% on Insider Trading and +78% on Harm-Pressure Knowledge. Across 12 models (0.5B--176B parameters), we find probe accuracy improves with scale: ~5% AUROC per 10x parameters (R=0.81). Geometrically, deception directions rotate gradually across layers rather than appearing at one location, explaining both why single-layer probes are brittle and why multi-layer ensembles succeed.

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 13 of its 17 references that are in the map. First posted on arXiv 15 Apr 2026.

Builds on

13 of 17 references.

References · 13

  1. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  2. 2023AI Deception: A Survey of Examples, Risks, and Potential Solutionscited by 188
  3. 2025Agentic Misalignment: How LLMs Could Be Insider Threatscited by 158
  4. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 115
  5. 2024Universal Neurons in GPT2 Language Modelscited by 102
  6. 2024Truth is Universal: Robust Detection of Lies in LLMscited by 79
  7. 2025Detecting Strategic Deception Using Linear Probescited by 68
  8. 2021The Neural Data Router: Adaptive Control Flow in Transformers Improves Systematic Generalizationcited by 62
  9. 2026Building Production-Ready Probes For Geminicited by 47
  10. 2026Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreakscited by 41
  11. 2025Liars' Bench: Evaluating Lie Detectors for Language Modelscited by 22
  12. 2025Difficulties with Evaluating a Deception Detector for AIscited by 13
1 more
  1. 2025AI Deception: Risks, Dynamics, and Controlscited by 7

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 4.

4 papers.

  1. 2026Tool-Call Dependency Structure is Linearly Decodable in LLM Agent Residual Streamscited by 1
  2. 2026HalluTracer: Pre-Decoding Truthfulness Prediction via Depth-Averaged Probe-Logitcited by 0
  3. 2026Linear Probing Provides Robust and Efficient Detection of Machine-Generated Textcited by 0
  4. 2026Accuracy and Uncertainty in Machine Deception Detection (Research White Paper)cited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception
  2. 2025Caught in the Act: a mechanistic approach to detecting deception
  3. 2026One Probe Won't Catch Them All: Towards Targeted Deception Detection
  4. 2026Probe Generalization as Subspace Selection for OOD Deception Detection
  5. 2026Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition
  6. 2026Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations
  7. 2026Alignment via Training Against Probes Without Losing Monitorability
  8. 2025The Secret Agenda: LLMs Strategically Lie and Our Current Safety Tools Are Blind