Non-Language Tasks with Pretrained LLMs
Pretrained LLMs can be used for a variety of tasks, but hyperparameter turning is crucial. ICML Workshop Paper, first author.
Recent Work
Dungeons and Data: Large-Scale Nethack Dataset
Creating agents for the long horizon proc-gen environment is a challenge, but our dataset helps kickstart the processes with offline RL. NeurIPS Benchmarks and Dataset Track.
moolib: Plaform for Distributed RL
A flexible platform allowing for large scale experiments facilitates efficient RL research on complex enviroments like Nethack.
NeurIPS 2021 Nethack Challenge
Are symbolic agents currently stronger than neural agents at proc-gen, long horizon complex games like Nethack? The results of our 2021 challenge confirm.
Minecraft for Grounded Language Learning
Minecraft is an ideal environment for exploring an agent's ability to combine language understanidng and task solving. Arxiv position paper.
RL Trained Starcraft Agent
The combinatorial action space of Starcraft can be approached with Reinforcement Learning by using just a partial simulator to do local planning. Project report, first author.