Inherent Labs Introduces Faraday 27B AI Scientist Agent for Research Replication
Inherent Labs introduced Faraday, a 27-billion parameter AI scientist agent trained using long-horizon reinforcement learning on the new Replica task suite. By leveraging GPT-5.5 Codex as an execution tool, Faraday outperforms larger models like Claude Opus 4.8 in reproducing scientific research figures and experiments.
Impact: Medium
Why it matters
Engineers building autonomous agents can adopt Faraday's architectural pattern of pairing a smaller fine-tuned directing agent with larger foundation models as tools.
TL;DR
- 01Faraday demonstrates that a 27B parameter agent can direct a much larger LLM (GPT-5.5 Codex) as an execution tool.
- 02Training agents on underspecified research replication tasks builds generalizable open-ended scientific problem-solving skills.
- 03Multi-sample aggregation and turn-level credit assignment stabilize RL training on noisy LLM-based reward signals.
Key facts
- Model Parameters
- 27B
- Benchmark Tasks
- 310 tasks across 100 research papers
- Tool Model
- GPT-5.5 Codex
Long-Horizon Reinforcement Learning on Replica
Inherent Labs developed Faraday using Replica, a synthetic RL environment containing 310 replication tasks derived from 100 machine learning and scientific research papers. Each task challenges the agent to reproduce published research figures under constrained compute and time budgets without access to original plotting code. Instead of relying on hand-coded evolutionary loops, Faraday utilizes long-horizon reinforcement learning to internalize scientific hypothesis testing.
Hierarchical Agent Orchestration with Large Coding Tools
Rather than operating standalone, the 27B Faraday model acts as a high-level directing agent that orchestrates GPT-5.5 Codex to execute code and experiments. During training, Faraday learned to manage GPT-5.4-mini before seamlessly generalizing to direct GPT-5.5 Codex at test time. To stabilize RL reward signals across multi-step research trajectories, the authors implemented auto-generated per-task rubrics, multi-sample aggregation, and turn-level credit assignment.
✓ When to use
- Designing hierarchical multi-agent architectures where a small controller directs large LLM tools.
- Setting up long-horizon RL rewards for complex coding or scientific workflows.
✕ When NOT to use
- Building low-latency or real-time user-facing chatbots.
- Simple single-turn code generation tasks where a single LLM prompt is sufficient.
What to do today
- Evaluate hierarchical agent designs using a fine-tuned small controller model over direct long-prompting of large LLMs.
- Implement per-task rubric evaluation and turn-level credit assignment when training agents on multi-step workflows.
Sources