Training 54 Model Organisms Shows Interpretability Audit Scores May Be a Lottery
gsarti_ · x · 2026-10-14
The paper "How to train your model organism" (with David Bau and Byron C. Wallace) warns that model organisms commonly used as testbeds for mechanistic interpretability can make stress-test results unreliable. The team trained 54 model organisms and found their interpretability scores vary widely depending on how they were trained — meaning interpretability benchmark results may be a lottery — and argue for more realistic model organisms in interpretability auditing.
More from Research
- NTU team has Claude write robot control code, hitting 70.24% success in sim — aigclink · 2026-10-14
- Six training strategies compared: can LLMs reason over graphs at scale? — burkov · 2026-10-14
- VisInteract: NeurIPS Spotlight reframes text-to-visualization as multi-turn intent retrieval — jiqizhixin · 2026-10-14
- WorldRecon reconstructs large-scale 3D scenes from unposed multi-view images, cutting Chamfer distance 43% — ZhejiangUniversity · 2026-10-14
- RSI-Nav: recursive self-improvement for navigation agents hits 68.6% SR on R2R-CE — OpenGVLab · 2026-10-14
- NUS proposes Jeva, a decision-native architecture with near-perfect null detection at thousands of candidates — NationalUniversityofSingapore · 2026-10-14