Interpretability competition launched to test logit lenses on Thinking Machines' Inkling

benno_krojer · x · 2026-11-21

The author is hosting a competition around Thinking Machines' Inkling model. The model's quirks — looping and no single clean residual stream — raise the question of whether existing verbalization lenses (logit lens, the author's latentlens) still work. Goals: develop new lenses, find the best readout locations, and surface novel discoveries. Registration is open.

Original post →

More from Research

Research channel →