Adrià Garriga-alonso

Craft­ing Poly­se­man­tic Trans­former Bench­marks with Known Circuits

Pac­ing Out­side the Box: RNNs Learn to Plan in Sokoban

Com­pact Proofs of Model Perfor­mance via Mechanis­tic Interpretability

Catas­trophic Good­hart in RL with KL penalty

An eval­u­a­tion of cir­cuit eval­u­a­tion metrics

Ophiol­ogy (or, how the Mamba ar­chi­tec­ture works)

Does liter­acy re­move your abil­ity to be a bard as good as Homer?

Thomas Kwa’s re­search journal

On Fre­quen­tism and Bayesian Dogma

A com­par­i­son of causal scrub­bing, causal ab­strac­tions, and re­lated methods

Causal scrub­bing: re­sults on in­duc­tion heads

Causal scrub­bing: re­sults on a paren bal­ance checker

Causal scrub­bing: Appendix

Causal Scrub­bing: a method for rigor­ously test­ing in­ter­pretabil­ity hy­pothe­ses [Red­wood Re­search]

The No Free Lunch the­o­rems and their Razor

Löb’s the­o­rem sim­ply shows that Peano ar­ith­metic can­not prove its own soundness