# Aviary: training language agents on challenging scientific tasks

**Authors:** Siddharth Narayanan, James D. Braza, Ryan-Rhys Griffiths, Manu Ponnapati, Albert Bou, Jon Laurent, Ori Kabeli, Geemi Wellawatte, Sam Cox, Samuel G. Rodriques, Andrew D. White  
**arXiv:** [2412.21154](https://arxiv.org/abs/2412.21154)  
**Date:** 2024-12-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.21154v1/

## Key points

- Language agents are a promising solution for automating intellectual tasks in science due to their ability to interact with tools using natural language or code.
- Aviary has been introduced as an extensible gymnasium for language agents, formalizing agents as policies that solve language-grounded partially observable Markov decision processes.
- Aviary focuses on three challenging scientific tasks: manipulating DNA constructs for molecular cloning, answering research questions by accessing scientific literature, and engineering protein stability.
- Language agents supported by open-source LLMs within Aviary can exceed both frontier LLM agents and human experts on multiple tasks at significantly lower inference costs.
- Aviary has proven to be a valuable resource for developing language agents capable of tackling complex scientific tasks efficiently, achieving impressive performance levels surpassing human-level task performance while maintaining cost-effectiveness.
- Collaborative efforts at FutureHouse supported by Eric and Wendy Schmidt, along with compute resources from the National AI Research Resource Pilot with support from NVIDIA, have been instrumental in driving progress in this area.
- The open-source nature of both Aviary and the LDP frameworks ensures accessibility for implementing environments and language agents across various domains.

## Summary

In the realm of solving complex real-world tasks in the field of science, a series of actions and observations are required. This process involves multiple cycles of analysis, tool utilization, and experimentation. Language agents have emerged as a promising solution for automating intellectual tasks in science due to their ability to interact with tools using natural language or code. However, the flexibility of these agents presents conceptual and practical challenges for software implementations. To address these challenges,  has been introduced as an extensible gymnasium for language agents. Agents are formalized as policies that solve language-grounded partially observable Markov decision processes known as . Within Aviary, five environments have been implemented, with a focus on three challenging scientific tasks: manipulating DNA constructs for molecular cloning, answering research questions by accessing scientific literature, and engineering protein stability. These environments were carefully selected for their emphasis on multi-step reasoning and relevance to contemporary biology research. Through online training and scaling inference-time compute capabilities, it has been demonstrated that language agents supported by open-source LLMs can not only match but exceed both frontier LLM agents and human experts on multiple tasks at significantly lower inference costs - up to 100 times lower in some cases. Furthermore, Aviary has proven to be a valuable resource for developing language agents capable of tackling complex scientific tasks efficiently. The introduction of the  has provided a formal structure for describing agent tasks and showcasing them as stochastic computation graphs. By leveraging behavior cloning, expert iteration, and inference-time sampling techniques with trained Llama-3.1-8B EI agents within Aviary's environments, impressive performance levels have been achieved that surpass human-level task performance while maintaining cost-effectiveness. The collaborative efforts at FutureHouse supported by Eric and Wendy Schmidt have played a crucial role in this advancement. Utilizing compute resources from the National AI Research Resource Pilot with support from NVIDIA has also been instrumental in driving progress in this area. The open-source nature of both Aviary and the LDP frameworks ensures accessibility for implementing environments and language agents across various domains. Overall, this work signifies a significant step towards high-throughput automation of meaningful scientific tasks within biology using efficient computational methods.