BALROG

eval
University College London

Benchmarking Agentic LLM/VLM Reasoning On Games (ICLR 2025): rolls NetHack, MiniHack, Crafter, and TextWorld into the standard long-horizon LLM-agent benchmark, with external adoption including NVIDIA NIM benchmarking. UCL DARK (Paglieri, Rocktäschel).

Paper (arXiv)GitHub

Paper

Venue ICLR 2025
arXiv HTML

Evaluation Details

Domains 3
Scoring agentic task success
Domains: games, agents, reasoning
evalagentsreasoning