mlsanityData to 7 Oct 2026

Latent Introspection: Models Can Detect Prior Concept Injections

arXiv 2602.20031 ↗First posted 23 Feb 2026cs.AI cs.LGcited by 12

Not on the map. Why?

Abstract (arXiv)

We uncover a latent capacity for introspection in a Qwen 32B model, demonstrating that the model can detect when concepts have been injected into its earlier context and identify which concept was injected. While the model denies injection in sampled outputs, logit lens analysis reveals clear detection signals in the residual stream, which are attenuated in the final layers. Furthermore, prompting the model with accurate information about AI introspection mechanisms can dramatically strengthen this effect: the sensitivity to injection increases massively (0.3% -> 39.9%) with only a 0.6% increase in false positives. Also, mutual information between nine injected and recovered concepts rises from 0.61 bits to 1.05 bits, ruling out generic noise explanations. Our results demonstrate models can have a surprising capacity for introspection and steering awareness that is easy to overlook, with consequences for latent reasoning and safety.

Key facts. Cited by 12 papers in the map (data through 7 Oct 2026). Builds on 12 of its 17 references that are in the map. First posted on arXiv 23 Feb 2026.

Builds on

12 of 17 references.

References · 12

  1. 2022Language Models (Mostly) Know What They Knowcited by 1,783
  2. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
  3. 2024Qwen2.5-Coder Technical Reportcited by 1,151
  4. 2023Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Promptingcited by 1,081
  5. 2023Steering Language Models With Activation Engineeringcited by 824
  6. 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 733
  7. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 673
  8. 2024A StrongREJECT for Empty Jailbreakscited by 352
  9. 2024The Remarkable Robustness of LLMs: Stages of Inference?cited by 138
  10. 2024Looking Inward: Language Models Can Learn About Themselves by Introspectioncited by 105
  11. 2026Emergent Introspective Awareness in Large Language Modelscited by 73
  12. 2025Privileged Self-Access Matters for Introspection in AIcited by 20

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 12.

12 papers.

  1. 2026Emergent Introspection in AI is Content-Agnosticcited by 5
  2. 2026Mechanisms of Introspective Awarenesscited by 4
  3. 2026Language models recognize dropout and Gaussian noise applied to their activationscited by 3
  4. 2026Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervisioncited by 3
  5. 2026Quantitative Introspection in Language Models: Tracking Emotive States Across Conversationcited by 1
  6. 2026Beyond Confidence: Rethinking Self-Assessments for Performance Prediction in LLMscited by 1
  7. 2026Self-Recognition Finetuning can Prevent and Reverse Emergent Misalignmentcited by 1
  8. 2026Traces of Social Competence in Large Language Modelscited by 0
  9. 2026Operational Proto-Introspection in Looped Language Models: Process-Quality Taps, Executable Branching, and the Readout-Control Boundarycited by 0
  10. 2026Open-Weight Masked Introspection: Measuring What Language Models Can Report About Their Own Computationcited by 0
  11. 2026Measuring Activation Control in Large Language Modelscited by 0
  12. 2026Language Models Act on Hidden Valencecited by 0