method
active
method:prefill-detection-taskPrefill detection task
Task where a random word is prefilled as the assistant's response, then the model is asked whether it intended to say that word, testing introspection on prior intentions.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Concept InjectionimplementsTechnique of injecting activation patterns associated with specific concepts into a model's internal states to test whether self-reports reflect ground truth.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Speculation that QK circuit 'concordance heads' underlie the ability to distinguish intended from unintended outputs.
- Task paradigm from prior work asking 'Did you detect an injected thought?' via YES/NO logit comparison; shown here to be confounded
- Adversarial multi-turn experiment where first turn uses pre-finetuning model to test if follow-up maintains character
- The process by which a model reconstructs the KV cache from the transcript when moving to a new server or different model
- The optimal layer for the prefill introspection differs from the optimal layer for detecting injected thoughts.
- One of four ToM tasks analyzed; requires inferring speaker intent from indirect hints; scored 0/1.
- Task of detecting a model's internal thoughts; found by Lindsey (2026) to peak at ~2/3 depth in transformers.
- Language model reasoning tasks with sequential geometry used in experiments.