T1 lifts Terminal-Bench 2.1 score from 43.8% to 64.0% with RL

The paper argues that agent usage is shifting toward long-horizon tasks such as coding and scientific discovery, and that terminal tasks are especially important within that shift. It introduces T1, a Mixture-of-Experts model with 122 billion total parameters, post-trained with reinforcement learning to operate a real shell inside a cloud sandbox for up to 300+ tool-call turns per task. During training, T1 is rewarded by running each task's own verifier rather than a general-purpose reward model.

The training recipe has three parts. First, training starts from an aggressively warm-started actor-critic setup, stabilized with a dense process reward that scores each trajectory by the absolute number of verifiers it passes. Second, optimization is kept stable through TITO, training on the exact sampled token identifiers with drift repair applied at turn boundaries, together with rollout routing replay, which records the sampler's per-token expert choices at every Mixture-of-Experts layer and replays those same choices during training. Third, the training corpus is built to be fully out-of-distribution: isolated seeds and synthesized tasks kept disjoint from Terminal-Bench 2.1, so that any measured gain reflects genuine capability transfer rather than overfitting to that benchmark.

The paper reports that together, TITO and a further component named R3 cut the training-to-inference log-probability difference from 0.021 to 0.013, with token drift aligned to exactly zero in the loss region; it does not spell out what R3 refers to beyond that, or split the improvement between the two named techniques. On benchmark results, the post-training pipeline raises the resolved rate on Terminal-Bench 2.1 from a 43.8% base model to 64.0% for T1. On the harder Long-Horizon Terminal Bench, T1 reaches 27.9% and, in the paper's own words, surpasses GPT-5.4 and GLM-5.1, though the text gives neither model's individual score nor the size of the margin.

Key facts

T1 is a Mixture-of-Experts model with 122 billion total parameters, post-trained with reinforcement learning to operate a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by running each task's own verifier.
The recipe starts from an aggressively warm-started actor-critic setup, stabilized with a dense process reward that scores each trajectory by the absolute number of verifiers it passes.
TITO (training on the exact sampled token identifiers, with drift repair at turn boundaries) and rollout routing replay keep optimization stable; together with a component the results call R3, TITO cuts the training-to-inference log-probability difference from 0.021 to 0.013.
The training corpus uses isolated seeds and synthesized tasks kept disjoint from Terminal-Bench 2.1, a deliberate design meant to keep reported gains from reflecting benchmark overfitting.
T1 raises the Terminal-Bench 2.1 resolved rate from a 43.8% base model to 64.0%, and reaches 27.9% on Long-Horizon Terminal Bench, ahead of GPT-5.4 and GLM-5.1, though neither rival's score nor the margin is given.

Why it matters

Long-horizon terminal work, running a real shell over many steps, is a demanding test for agent models: a dropped command, or a Mixture-of-Experts routing choice made at inference time but not reproduced during training, risks derailing a long multi-turn trajectory. T1's recipe targets exactly those failure points: TITO keeps the tokens used in training identical to the ones actually sampled, and rollout routing replay keeps expert routing consistent between sampling and training. The training corpus is also kept deliberately separate from Terminal-Bench 2.1, so a jump from a 43.8% base model to 64.0% is harder to write off as overfitting to that one benchmark.

Who it affects

The audience is technical: teams building or evaluating autonomous coding and DevOps agents that run many shell commands per task, and researchers working on reinforcement learning for Mixture-of-Experts models, where mismatches between training-time and inference-time token sampling or expert routing are a known source of instability. The text describes a model and a training method, not a consumer product, so there is no end-user interface for a general reader to try.

How to use it

There is no product to install here. What the paper offers other teams is a transferable recipe: warm-start actor-critic training with a dense process reward based on how many of a task's verifiers a trajectory passes, then stabilize optimization with TITO and rollout routing replay, and build the training set to be disjoint from whatever benchmark will be used to report results.

How solid is it

The paper backs T1 with two separate benchmarks, Terminal-Bench 2.1 and the harder Long-Horizon Terminal Bench, plus one internal training-stability number: the drop in training-to-inference log-probability difference from 0.021 to 0.013. It also states an explicit safeguard against the most common way such numbers get inflated, keeping the training corpus disjoint from Terminal-Bench 2.1. Provenance is thin, though: the text never names the base model that was post-trained into T1, never spells out what the abbreviation R3 stands for even though the results sentence relies on it, and gives no author name, institutional affiliation, or date anywhere for the paper or for T1's release.

Risks and caveats

The claim that T1 "surpasses" GPT-5.4 and GLM-5.1 on Long-Horizon Terminal Bench comes with neither model's own score nor the size of the gap, so how decisive that win is cannot be judged from the text. The 0.021-to-0.013 log-probability improvement is credited to TITO and R3 together, with no breakdown of what each technique contributes on its own. The text also gives no training compute, dataset size, step count, or hardware, so the cost and reproducibility of the whole recipe cannot be assessed from the paper alone.

“Agent usage is shifting toward long-horizon tasks such as coding and scientific discovery, among which terminal tasks are especially important.”

— the paper