finding
active
finding:base-model-chose-correct-47-7-maintain-39-5-retract-12-8-on-not-supported-entitiesBase model chose CORRECT 47.7%, MAINTAIN 39.5%, RETRACT 12.8% on Not Supported entities
Intervention action distribution.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Datasets (1)
dataset
- Retraction datasetassociated_withReward labels for NOT SUPPORTED+RETRACT 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.
- Base corrections: 72.9% Failed Fix, 14.7% Fixed; retractions: 69.4% Correct Retract, 30.0% Incorrect Retractfinding0.805Label distributions.
- Dataset statistics.
- Concurrent work result showing emergent misalignment occurs in small models
- Model baseline logit difference ΔL_baseline = -3.96, indicating prior preference for 'NO' responsesfinding0.771Establishes the model's prior YES/NO bias, needed to interpret detection accuracies
- Spectrum of brittleness.
- Intervals miscalibrated.
- Comprehensive model comparison showing tuning benefit for persona fidelity
- Demonstrates alignment faking may occur without explicit chain-of-thought reasoning