A 27B agent just beat Claude Opus 4.8 and GPT-5.5 on held-out research replication.

Replica turns paper replication into a scalable RL task space. Replicating a paper forces the same hypothesis-driven exploration as open research, and it surfaces details the original authors left underspecified.

The reward signal comes from an auto-generated rubric judge that runs low-noise and agrees with human assessment of replication quality.

Faraday, the resulting 27B agent, calls coding agents as tools. Rollout analysis shows it takes a more scientifically principled approach rather than gaming the rubric.

The authors argue this points toward long-horizon scientific capability trained into weights, without requiring complex harnesses.

Paper: https://arxiv.org/abs/2608.13331

Track more trending AI papers in our academy: https://academy.dair.ai/