Reproduce “Tmax: A simple recipe for terminal agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hamish Ivison, Junjie Oscar Yin, Rulin Shao, Teng Xiao, Nathan Lambert, Hannaneh Hajishirzi
TMax creates 13,565 executable terminal environments across nine design axes, audits and repairs their verifiers, and trains agents with outcome-based reinforcement learning. The resulting open TMax-9B solves 33.0% of Terminal-Bench 2.0 tasks, versus 21.0% for its Qwen3.5-9B base model. This benchmark result does not show that every design choice is necessary or that the tasks cover real-world coding environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.