claim
active
claim:aos-might-be-too-expressive-so-explanations-may-reflect-the-ao-s-understanding-rather-than-the-target-llm-sAOs might be too expressive, so explanations may reflect the AO's understanding rather than the target LLM's
Limitation; confabulation risk.
Source paper
extracted_from(2025) · Adam Karvonen
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- AOs might produce plausible guesses not reflecting target model.
Methods (1)
method
- Linear ProbecitesSimple linear classifiers trained on model activations used as the probing technique within the introduced method.
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.
- Argument for scalability.
- Overall positioning.
- AOs articulate responses with natural-language flexibility rather than a fixed concept setclaim0.801Contrast with SAE features.
- Argument for simplicity.
- Limitation relative to SAEs.
- Key limitation identified: NLAs hallucinate specific details while preserving thematic accuracy; informs practical usage.
- AOs are not best on Secret Side Constraint, where black-box prompting reveals the secretfinding0.773The one task where AOs lose.
- Novelty claim establishing the paper's contribution relative to prior work focused on closed-form tasks