hypothesis
active
hypothesis:sycophancy-or-hint-use-features-may-explain-why-the-model-attends-to-the-human-s-answer-hintSycophancy or hint-use features may explain why the model attends to the human's answer hint.
Conjectured but not shown in graphs.
Neighborhood — ranked by edge-count
Papers (1)
paper
Questions (1)
question
- Missing mechanistic piece.
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.
- Author's interpretation from SAE decomposition of the sycophancy persona vector
- Neutral hint: suggestion from a person.
- Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models (Denison et al. 2024)concept0.799Related work on LLMs generalizing to reward hacking; methodology used for RL experiments
- The model attends to the J-space on the user prompt when determining its introspective answer, so concepts there are reportable.hypothesis0.798Motivated by elephant observation.
- SAEs uncover safety-relevant representations that might be monitored or controlled.
- The paper's honest statement of the residual interpretive ambiguity after all controls
- Specific risk identified in spiritual use of AI.
- Conditional prediction about how a well-informed dialogue agent would handle questions of personal identity