mlsanityData to 14 Oct 2026

Probing deception, bias and beliefs

88 papers2 this weekLow confidence in this grouping: the papers may belong to more than one topic, or the group may changeLow confidence in this grouping: the papers may belong to more than one topic, or the group may change · About

Keywords: bias, probes, llms, linear, deception

Key facts. Probing deception, bias and beliefs: 88 papers on arXiv since 2023 (data through 14 Oct 2026). 12 of its 88 mapped papers (14%) cite “Detecting Strategic Deception Using Linear Probes” (2025, arXiv 2502.03407). 8 new papers 24 Sep–14 Oct.

What’s new

24 Sep – 14 Oct 2026

8 new papers, 24 Sep – 14 Oct; 2 this week.

  1. This week · 2
  2. Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering also fits: Activation steering methods
  3. Rule vs.Consequence: Dissociable Internal Representations of Moral Reasoning in LLMs
  4. Earlier · 6
  5. Latent space bias directions in LLMs capture confidence, not fairness
  6. Will the Judge Flip? Predicting Position-Sensitive LLM Judgments from Residual Stream Activations
  7. Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steering
  8. Feature Elicitation for SAE Steering of Deceptive Behaviour in Small LLMs also fits: Steering and probing for LLM safety
  9. Right Answer, Wrong Mechanism: Detecting Pernicious Divergence in Causal Interventions also fits: Activation steering methods
  10. The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining also fits: Hallucination detection from model internals

Foundations

Most cited by papers here

  1. Detecting Strategic Deception Using Linear Probes2025cited by 12 here
  2. Localizing Lying in Llama: Understanding Instructed Dishonesty on True-False Questions Through Prompting, Probing, and Patching2023cited by 2 here
  3. Linear Representations of Political Perspective Emerge in Large Language Models2025cited by 2 here
  4. Dissecting Bias in LLMs: A Mechanistic Interpretability Perspective2025cited by 2 here
  5. Revisiting In-context Learning Inference Circuit in Large Language Models2024cited by 1 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 9 here
  2. 2023Progress measures for grokking via mechanistic interpretabilitycited by 6 here
  3. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 7 here
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 16 here
  5. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 2 here
  6. 2024Scaling and evaluating sparse autoencoderscited by 3 here
  7. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 7 here
  8. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 4 here
  9. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 4 here
15 more it builds on, not in this chain
  1. 2022Extracting Latent Steering Vectors from Pretrained Language Modelscited by 4 here
  2. 2022Toy Models of Superpositioncited by 7 here
  3. 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 5 here
  4. 2023Steering Language Models With Activation Engineeringcited by 19 here
  5. 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 4 here
  6. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 25 here
  7. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 3 here
  8. 2023Localizing Lying in Llama: Understanding Instructed Dishonesty on True-False Questions Through Prompting, Probing, and Patchingcited by 2 here
  9. 2023Steering Llama 2 via Contrastive Activation Additioncited by 22 here
  10. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 5 here
  11. 2024Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categoriescited by 2 here
  12. 2024Programming Refusal with Conditional Activation Steeringcited by 3 here
  13. 2024Automatically Interpreting Millions of Features in Large Language Modelscited by 3 here
  14. 2025Detecting Strategic Deception Using Linear Probescited by 12 here
  15. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 3 here

The 24 most cited of 102.

Browse

30 of 88, ranked 6–35

  1. 6Tracing Moral Foundations in Large Language Models20261
  2. 7One Probe Won't Catch Them All: Towards Targeted Deception Detection20261
  3. 8Linear Probe Accuracy Scales with Model Size and Benefits from Multi-Layer Ensembling20261
  4. 9Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations20261
  5. 10Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States20261
  6. 11Elucidating Mechanisms of Demographic Bias in LLMs for Healthcare20251
  7. 12Unsupervised Concept Vector Extraction for Bias Control in LLMs20251
  8. 13FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering20251
  9. 14Evaluate Bias without Manual Test Sets: A Concept Representation Perspective for LLMs20251
  10. 15When Thinking LLMs Lie: Unveiling the Strategic Deception in Representations of Reasoning Models20251
All 30
  1. 16When Truthful Representations Flip Under Deceptive Instructions?20251
  2. 17Activation Steering for Bias Mitigation: An Interpretable Approach to Safer LLMs20251
  3. 18Mechanistic Interpretability with SAEs: Probing Religion, Violence, and Geography in Large Language Models20251
  4. 19The Hypocrisy Gap: Quantifying Divergence Between Internal Belief and Chain-of-Thought Explanation via Sparse Autoencoders20260
  5. 20White-Box Sensitivity Auditing with Steering Vectors20260
  6. 21The Confidence Manifold: Geometric Structure of Correctness Representations in Language Models20260
  7. 22How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing20260
  8. 23Understanding the Dynamics of Demonstration Conflict in In-Context Learning20260
  9. 24Understanding Moral Reasoning Trajectories in Large Language Models: Toward Probing-Based Explainability20260
  10. 25Quantitative Introspection in Language Models: Tracking Emotive States Across Conversation20260
  11. 26Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure20260
  12. 27Closing the Confidence-Faithfulness Gap in Large Language Models20260
  13. 28CogBias: Measuring and Mitigating Cognitive Bias in Large Language Models20260
  14. 29Moral Sensitivity in LLMs: A Tiered Evaluation of Contextual Bias via Behavioral Profiling and Mechanistic Interpretability20260
  15. 30In-Context Fixation: When Demonstrated Labels Override Semantics in Few-Shot Classification20260
  16. 31Linear Separability of Activation Representations after Supervised Fine-Tuning on Incorrect Responses: A Study of Synthetic Dishonesty in Large Language Models20260
  17. 32The Amplifying Mirror: Locating and Steering the Partisan Direction inside a Large Language Model20260
  18. 33Frame-Conditioned Moral Computation in LLaMA 3.1-8B-Instruct: A Mechanistic Interpretability Audit of Ethical Reasoning20260
  19. 34Comparing Linear Probes with Mahalanobis Cosine Similarity20260
  20. 35Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models20260