finding
active
finding:trained-models-improve-substantially-over-base-on-held-out-wildchat-and-petri-investigations-within-0-03-auroc-of-opus-referenceTrained models improve substantially over base on held-out WildChat and PETRI investigations, within ±0.03 AUROC of Opus reference
In- and out-of-distribution generalization.
Source paper
extracted_from(2026) · Adam Karvonen · Euan Ong · Subhash Kantamneni · Samuel Marks
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Pipeline retargetable to other behavior classes.
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.
- Trained models slightly exceed Opus reference on unfiltered bank (+0.015 Qwen3-8B, +0.033 Qwen3.5-397B)finding0.815Gains concentrated in low-concreteness claims.
- Generalization to hint setting.
- Models trained directly with asynchronous updates would exhibit even greater robustness than synchronously trained modelshypothesis0.783Hypothesis that motivated the asynchronous robustness comparison experiment
- Core evidence that model withholds pro-animal-welfare responses during training
- Post-trained models repurpose user token activations to represent the Assistant's thoughts.hypothesis0.768Tested by J-lens comparison.
- Scaling laws for dictionary learning are unknown and needed to assess feasibility on frontier models
- Core goal statement.
- Demonstrates the sharp drop in factual truth generalization at the counting boundary.