Researchers have developed JaxAHT, a new open-source library built with JAX, designed to accelerate and standardize the research process for ad hoc teamwork (AHT) in artificial intelligence. This library aims to overcome the computational costs and lack of standardized benchmarks that have previously hindered AHT progress. JaxAHT offers a unified framework for generating teammates, training ego agents, and evaluating them against unseen partners, achieving significant speedups compared to PyTorch implementations. The library also includes a suite of evaluation teammates across various domains like Level-Based Foraging, Overcooked, and Hanabi, and was used to benchmark different AHT methods, revealing that no single algorithm consistently outperforms others. AI
IMPACT This library could significantly speed up AI research in multi-agent coordination and teamwork.
RANK_REASON The cluster describes a new open-source library and benchmark suite for AI research, published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →