Santiago Aranguri

Log­its as a new mon­i­tor for eval­u­a­tion awareness

Pre­dict­ing Rare LLM Failures with 30× Fewer Rollouts

Ver­bal­ized Eval Aware­ness In­flates Mea­sured Safety

Probe-Based Data At­tri­bu­tion: Sur­fac­ing and Miti­gat­ing Un­de­sir­able Be­hav­iors in LLM Post-Training

Re­pro­duc­ing steer­ing against eval­u­a­tion aware­ness in a large open-weight model

SAE on ac­ti­va­tion differences

Tied Cross­coders: Ex­plain­ing Chat Be­hav­ior from Base Model