History
Percentage of reversed comparative yes/no question pairs exhibiting implicit post-hoc rationalization in early-2025 frontier models
Defined and estimated
Completed definition (Arcuschin et al. v1 models, World Model-derived pair set, v1 autorater criterion). Estimate 7-33% across models, central ~15% judged. Noted later versions report much lower rates under stricter criteria. No decomposition needed. Budget was limited; per-model table and later-version figures not re-verified, hence marginal yield 0.4.
Claim entered the graph