Experiment 3: Semantic Clustering of Experience Reports

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

• Cross-model semantic convergence of experience reports under self-referential processing is difficult to reconcile with roleplay because independently trained models construct distinct semantic profiles in all control conditionsclaim0.764
  The paper's argument against pure sycophancy as explanation for results
• Experience reports under self-referential processing are mechanistically gated by SAE features associated with deception and roleplayclaim0.732
  Claim supported by Experiment 2 dose-response curves; suppressing deception features increases consciousness reports, amplifying decreases them
• Experience Replaymethod0.726
  RL technique using episodic memory to improve sample efficiency; used in some game-playing agents.
• Experience Sampling Method (ESM)method0.721
  Human psychology method for repeated in-situ self-report; methodological inspiration for the paper's approach
• Structure Of Experienceconcept0.719
• Behavior Clusteringconcept0.718
  Grouping similar model behaviors; the unsupervised method surfaces clusters of concerning patterns.
• The features are often organized in geometrically-related clusters that share a semantic relationship.claim0.717
  Decoder cosine similarity maps onto concept similarity.
• It's tricky, because for a typical language model the entity is sort of tricameral: the base simulator, the simulated simulator, and the simulated awareness.quote0.716
  Antra's earlier definitive statement of the tricameral model.