Inherent Labs built a 27B AI scientist that beats frontier models

Giving a 27B model the reins to boss around GPT-5.5 sounds like hiring an intern to manage senior engineers. But Inherent Labs just proved that in scientific discovery, dividing the brain from the hands might actually be the winning move.

The research lab unveiled Faraday, a 27B agent tailored for scientific replication. In benchmarks requiring agents to reproduce academic paper results, this relatively small model beat both Claude Opus 4.8 and full-scale GPT-5.5.

Splitting hypothesis from hard labor

Faraday does not write code. Instead, it decouples scientific reasoning from implementation.

Rather than forcing a single frontier model to juggle hypotheses, syntax, debugging, and execution, Faraday functions purely as a principal investigator. It makes the research decisions and uses GPT-5.5 Codex as an external tool to write the code. During training, Faraday was only paired with GPT-5.4-mini, yet it generalized smoothly to commanding stronger execution agents at test time.

The benchmark itself is clever. The agent is handed an arXiv paper with all figures and plots stripped out, forced to replicate the core findings within a strict compute budget. Because scientific papers notoriously highlight successes while omitting failed attempts, reproducing results naturally forces hypothesis-driven exploration.

Teaching taste when rewards are messy

The fundamental challenge of applying reinforcement learning to open-ended science is that there are no neat, verifiable rewards. Former OpenAI researcher Lilian Weng recently highlighted this exact bottleneck: evaluation for research taste and novelty is notoriously fragile, with benchmarks like PaperBench hovering around 21% accuracy, well below human PhDs.

Faraday tackles this by generating task-specific rubrics on the fly. It combines an automated judge with multi-sample aggregation and turn-level credit assignment to suppress noise during exploration.

This mirrors Tsinghua's open-source openrsi project, where a 35B model rivaled GPT-5.6 Sol on MLE-Bench Lite. Both projects point toward the same emerging consensus: small models steering frontier tools can routinely match or outmaneuver standalone massive models in ML engineering.

Why it matters

For researchers and ML practitioners, this approach offers an immediate practical blueprint. Paper reproduction is a notorious time sink. Having a lightweight 27B model orchestrate API calls to verify someone else's paper could compress weeks of frustrating debugging into hours of autonomous runs.

For the broader industry, it shatters the assumption that only trillion-parameter monolithic giants can unlock AI for Science. Splitting architecture into specialized, taste-driven planners and generic code-execution engines will likely become the standard template for autonomous labs.

We think Inherent Labs took a smart shortcut by treating code generation as a commodity rather than an end goal. The real question mark remains the rubric-based automated judge. In open-ended domains, agents excel at finding loopholes in grading rubrics, and proving this scoring setup remains bulletproof over longer horizons will be the real test.