AI 代理與研究自動化
Faraday Uses a 27B Outer-Layer Agent to Orchestrate GPT-5.5, Outperforming Standalone Frontier Models on Paper-Figure Reproduction Benchmark
Inherent has unveiled Faraday, trained with long-horizon reinforcement learning, which puts a 27B model in charge of research decisions while delegating programming work to GPT-5.5. It outperforms Claude Opus 4.8 and GPT-5.5 on Replica’s held-out set, but the results should not be interpreted as showing that the 27B model alone surpasses frontier models.
Inherent recently formally introduced Faraday. Rather than replacing large coding models, it adds a “scientist” agent post-trained on research tasks as a layer above them. The team built it on Qwen3.6-27B and trained it using LoRA, a 128K context window, and a modified version of GRPO. Faraday formulates hypotheses, plans experiments, and decides what to do next, while calling Codex GPT-5.5 to write and execute the actual code.
The accompanying Replica benchmark contains 310 figure-reproduction tasks derived from 100 machine learning and AI for Science papers. The agent can see only versions of the papers with the target figures removed and must reconstruct the results within a containerd container, a one-hour limit, and a GPU allocation of one-seventh of an H200 MIG instance. Claude Opus 4.7 first generates a rubric for each task, after which multiple Codex judge samples inspect the output figures, code, and execution traces. Turn-by-turn credit assignment is used to prevent GRPO training from collapsing on long trajectories.
In comparisons using the same time and GPU budgets, with eight rollouts per task, Faraday outperformed both Claude Opus 4.8 and GPT-5.5 on 73% of in-distribution tasks and 60% of unseen AI-for-science tasks. Its average test-set scores were 6% and 8% higher, respectively. The significance of this result is that research judgment can be trained as a control policy that orchestrates more capable tools; the controller model does not necessarily need to be comparable in size to the tool model.
The limitations are equally clear. The graders, rubric generator, and coding tool all depend on closed-source frontier models, while human researchers validated only a subset of the trajectories that automated graders judged to favor Faraday. Replica also focuses on reproducing existing figures and does not yet demonstrate that the agent can reliably identify irreproducible results or conduct open-ended scientific research. Engineering teams should next watch for whether the tasks and weights are released and whether the gains persist when other coding agents are substituted.