T1, a 122B MoE Terminal Agent, Goes From 43.8% to 64.0% on Terminal-Bench 2.1 With RL on a Real Shell

arXiv·high signalXBlueskyLinkedInCopy link

T1 (arXiv 2609.11042) trains with RL for 300+ tool-call turns per task in a cloud sandbox, rewarded by each task's own verifier and trained on a corpus disjoint from the benchmark. Two stability tricks carry most of the recipe. TITO trains on the exact sampled token IDs, and rollout routing replay records per-token MoE expert choices. Together they cut train-inference log-prob drift from 0.021 to 0.013. The arXiv HTML reports T1 at 64.0% versus Claude Opus 4.6 at 63.8% and GPT-5.4 at 54.8%. On Long-Horizon Terminal Bench it scores 27.9%.