finding
active
finding:policy-rollout-dendrograms-cluster-intervention-entity-and-action-in-separate-sub-branches-absent-in-base-rolloutsPolicy rollout dendrograms cluster intervention, entity, and action in separate sub-branches; absent in base rollouts
Preliminary mechanism evidence.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- A policy partially becomes factual by being primed for acting in a factually precise manner.supportsInterpretation of merged dendrogram clusters.
Hypotheses (1)
hypothesis
- Mechanism hypothesis for future work.
Methods (1)
method
- Preliminary interpretability of interventions.
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.
- Used to test event alignment.
- Motivating statement.
- What mechanisms underlie the changes to the policy, including the reduction when interventions are inlined?question0.728Open interpretability question.
- Practical implication.
- Key property of collective intelligence: emergent behaviors at the group level.
- A pair of query and key subcomponents distributed across attention heads performs previous-token behaviorfinding0.707VPD recovers an attention algorithm for attending to the previous token, distributed across multiple heads.
- Observed by Anima Labs in untrained base models; not present in training data, implying computational origin of self-reported parallel processing.
- Interpretation of KL asymmetry.