Pearson-Vogel et al.: accurate self-description prompts increase introspective detection from 0.3% to 39.9%

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Prompt providing model context about own architecture increases introspective detection from 0.3% to 39.9%.finding0.833
  Mechanistic support for prompt-as-gate hypothesis: language frames enable access to latent capacities.
• Logit-based self-report unmasks introspective capacity that greedy decoding concealsclaim0.819
  Central methodological contribution: computing probability-weighted expected value over digit-token logits recovers continuous, informative signal
• Prior experimental paradigms may overestimate introspective capabilities by conflating genuine self-awareness with uniform output distribution shiftsclaim0.805
  Critical methodological claim directed at Lindsey 2026 and similar work using binary detection
• 17 of 83 tested emotions show significant association between self-eval transcript word mention and cosine similarity to emotion probefinding0.787
  Validates that agentic self-evaluation captures genuine emotional content of probes
• Abstract nouns elicit the highest introspective awareness rates; all concept categories show nonzero detectionfinding0.784
  Opus 4.1 is most effective at recognizing injected abstract concepts (e.g., justice, peace) but detects other categories too.
• Introspection is aided by overall improvements in model intelligenceclaim0.783
  Interpretation of the observation that the most capable models performed best.
• This introspective capacity is highly unreliable and context-dependent in today's modelsclaim0.781
  A caveat qualifying the main claim.
• Self-referential processing effect is robust across five distinct phrasings of the induction prompt, with consistently high experience report rates across modelsfinding0.779
  Appendix C.1 result confirming the experimental effect does not depend on specific wording