Repo

ArmoRM-SAE-Alignment

@christianrogersby Christian Rogers · University of Utah

0

Do reward models reward making stuff up? An interpretability analysis of ArmoRM vs. RLFR-style hallucination probes, with Matryoshka SAEs showing factuality-labeled features barely contribute to the RM's own correctness scores.

Built withPythonfactualityhallucination-detectioninterpretabilityllmmechanistic-interpretabilityreward-model

See what other members are building and what each project is looking for.