finding
active
finding:full-model-s-90-credible-intervals-have-empirical-coverage-only-0-48-to-0-64Full Model's 90% credible intervals have empirical coverage only 0.48 to 0.64.
Intervals miscalibrated.
Source paper
extracted_from(2026) · Eric Bigelow · Amir Zur · Satchel Grant · Tal Haklay +6
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Central interpretation.
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.
- Concurrent work result showing emergent misalignment occurs in small models
- Strong scaling trend for introspective fidelity when excluding invalid steering-sign pairs
- Group-level simulation result showing generalizability of BMR benefit across agents
- Intervention action distribution.
- Models produce first-attempt mean scores 87.8-91.8/100 without steering across all model familiesfinding0.760Establishes high baseline quality confirming steering-induced degradation is the experimental signal
- Indicates lack of statistically significant differences between top methods
- The model tends to reflect more when the question is difficult, and accuracy is generally lower for harder questionshypothesis0.755Hypothesis explaining negative correlation between reflection rate and accuracy without implying reflection is harmful