LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents

Researchers introduced LiteCoder-Terminal-Gen, a zero-dependency pipeline that builds executable terminal training environments from domain specifications.

They also released large LiteCoder-Terminal-SFT and LiteCoder-Terminal-RL datasets for training command-line language agents.

Models fine-tuned on these synthetic resources performed better on Terminal Bench, showing stronger terminal-task capability.

DMPO further improved results, suggesting preference optimization can boost long-horizon terminal workflows.