claim
active
claim:gemma-represents-a-strong-ordering-of-interventions-in-activations-but-cannot-output-it-in-token-space-when-instructedGemma represents a strong ordering of interventions in activations but cannot output it in token space when instructed.
Interpretation of probe vs judge gap.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- At n=256 best-of-n, probe scoring outperforms Gemma LLM-judge scoring by nearly 15 percentage pointssupportsTest-time scaling comparison.
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.
- Fig. 2a,e.
- Post-trained models repurpose user token activations to represent the Assistant's thoughts.hypothesis0.770Tested by J-lens comparison.
- Practical implication.
- Limitation acknowledgment about the adequacy of the linear representation assumption
- Layer localization.
- General principle derived from the Mountain Car experiment: curved manifold-following yields coherent manipulation, linear shortcuts fail.
- Scaling trends at larger scale.
- Long-context generalization.