Replica: 27B Agent Beats Claude Opus 4.8 and GPT-5.5 on Research Replication

A 27B agent using the Replica framework (RL from auto-generated rubrics) outperformed much larger models on held-out research replication tasks. This challenges the assumption that only massive models can do deep scientific reasoning and points to a new paradigm for training capable agents without brute-force scaling. The breakthrough has major implications for applied AI efficiency and the future of agentic systems.