COLM 2026: New causal methods localize diffuse model behaviors like sycophancy
dhadfieldmenell · x · 2026-10-07
At COLM 2026, Aruna and collaborators (Amir Zur, Atticus Geiger, David Hadfield-Menell) presented work using causal methods to localize diffuse behaviors in LLMs such as style and sycophancy. The authors argue that common localization approaches relying on single-token differences struggle with such distributed generative behaviors. Poster #80, Franciscan A.
More from Research
- COLM paper: latent reasoning traces decodable 65-93% of the time in LRMs — sarahwiegreffe · 2026-10-07
- Apple ML Research is hiring a PhD intern to make foundation models know what they know — sineadwilliamso · 2026-10-07
- Project Glasswing Reports 135K Verified Vulnerabilities, 9,333 Already Patched — ResultBackground2450 · 2026-10-07
- CAVEAT testbed exposes how merchants can steer your shopping AI agent — ZacharyHuang12 · 2026-10-07
- Pi and mini-swe-agent Passed 9/9 Checks Each — a Second Review Still Found Bugs — Mysterious-Desk-3492 · 2026-10-07
- GEA treats agent groups as the evolution unit, hitting 71% SWE-bench Verified with zero human help — xwang_lk · 2026-10-07