Smaller AI models can beat frontier models by using them as tools

A new paper from startup Inherent introduces Faraday, a Qwen 2.7B model fine-tuned via RL to replicate missing figures from ML papers. The model's key innovation is using Codex as a tool rather than attempting the task alone. The results show Faraday outperforming much larger models like Claude and GPT-5 on

this narrow task, which is surprising but also raises questions. The researchers use a rubric-based LLM judge for training rewards, averaging judgments across multiple rollouts to reduce variance. They also implement a weighted credit assignment across agent turns, concentrating reward on tool-use actions. This approach successfully biases the model toward more

effective behavior patterns. A critical methodological concern is the asymmetric evaluation: Faraday uses Codex as a tool while baseline models get no equivalent augmentation. The optimized Codex prompt only marginally improves results, suggesting the real advantage comes from the agent orchestration rather than model capability. Human evaluations confirm Faraday produces qualitatively

more rigorous experiments with fewer shortcuts. The paper's infrastructure details reveal a complex training setup using Kubernetes, Ray, and Nvidia's NeMo RL stack with custom context parallelism for hybrid attention models. The training process was iterative and pragmatic rather than theoretically derived, switching precision and coding agents mid-training without restarting.