Simulated eval reliably detects direction of model gains, not magnitude
SRSchmidgall · x · 2026-10-31
Across 60 scenarios, the autorater agreed only moderately with psychiatrists (ρ=0.57-0.58), and replacing human patients with simulated ones inflated the measured gain (0.34 vs 0.21).
The authors' conclusion is an important methodological caveat: simulated evaluation reliably detects the direction of between-model gains, but not their magnitude. Figures show scenario-level delta correlation dropping from ρ=0.40 (human rater, human patient) to 0.14 (AI rater, AI patient).
Related event: Google trains Gemini for psychiatric intake via simulated-patient RL(11 posts)→
More from Research
- Researcher: capability transfer hits a ceiling when training against less human-like models — saurabh_shah2 · 2026-10-31
- Compaction study: agent constraints vanish 6x more often than task goals in long sessions — maksym_andr · 2026-10-31
- CMU's Gym-Anything uses coding agents to auto-build 200 robot RL environments; best model scores 0% — AkariAsai · 2026-10-31
- Cherimoya ChIP-seq models ship as-is: bad underlying data means bad predictions — jmschreiber91 · 2026-10-31
- When chromatin models disagree: null accessibility shift but a disrupted CEBPG binding site — jmschreiber91 · 2026-10-31
- Whole-Body Dexterity turns the humanoid body into one large dexterous manipulator — chris_j_paxton · 2026-10-31