T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks

Hugging Face Daily Papers Papers
reinforcement-learning mixture-of-experts terminal-tasks long-horizon agent

Summary

T1 is a 122B Mixture-of-Experts model trained with reinforcement learning for long-horizon terminal tasks, achieving state-of-the-art results on benchmarks like Terminal-Bench 2.1 and surpassing models such as GPT-5.4 and GLM-5.1.

Agent usage is shifting toward long-horizon tasks such as coding and scientific discovery, among which terminal tasks are especially important. We introduce T1, a Mixture-of-Experts model of 122B total trained with reinforcement learning, operating a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task's own verifier. We provide a comprehensive recipe: First, an aggressively warm-started to stabilize actor-critic training, with a dense process reward scoring trajectories by the absolute number of passing verifiers. Second, stable optimization through TITO construction, training on the exact sampled token identifiers with drift repair at turn boundaries, and rollout routing replay, recording the sampler's per-token expert choices at every MoE layer and replaying them during training. Third, fully out-of-distribution training corpus: isolated seeds and synthesized tasks disjoint from Terminal-Bench 2.1 ensures gains reflect genuine capability transfer over benchmark overfitting. Together, TITO and R3 cut the training-to-inference log-probability difference from 0.021 to 0.013, with exactly aligned zero token drift in the loss region. On Terminal-Bench 2.1, our post-train pipeline raises initial base model from 43.8% to T1 with 64.0% resolved. On Long-Horizon Terminal Bench, T1 reaches 27.9% and surpasses GPT-5.4 and GLM-5.1.
Original Article
View Cached Full Text

Cached at: 09/11/26, 02:14 AM

Paper page - T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks

Source: https://huggingface.co/papers/2609.11042 Published on Sep 10

·

Submitted byhttps://huggingface.co/TberiusJunyao

Junyaoon Sep 10

Abstract

T1 is a 122B Mixture-of-Experts model trained with reinforcement learning to execute long-horizon terminal tasks in a cloud sandbox, achieving state-of-the-art results through stable actor-critic optimization and out-of-distribution training.

Agent usage is shifting toward long-horizon tasks such as coding and scientific discovery, among which terminal tasks are especially important. We introduce T1, aMixture-of-Expertsmodel of 122B total trained withreinforcement learning, operating a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task’s own verifier. We provide a comprehensive recipe: First, an aggressively warm-started to stabilizeactor-critictraining, with a denseprocess rewardscoring trajectories by the absolute number of passing verifiers. Second, stable optimization throughTITOconstruction, training on the exact sampled token identifiers with drift repair at turn boundaries, androllout routing replay, recording the sampler’s per-token expert choices at every MoE layer and replaying them during training. Third, fully out-of-distribution training corpus: isolated seeds and synthesized tasks disjoint fromTerminal-Bench2.1 ensures gains reflect genuine capability transfer over benchmark overfitting. Together,TITOand R3 cut the training-to-inference log-probability difference from 0.021 to 0.013, with exactly aligned zero token drift in the loss region. OnTerminal-Bench2.1, our post-train pipeline raises initial base model from 43.8% to T1 with 64.0% resolved. On Long-Horizon Terminal Bench, T1 reaches 27.9% and surpasses GPT-5.4 and GLM-5.1.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2609\.11042

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper1

#### TberiusJunyao/T1-122B-A10B Image-Text-to-Text• 125B• Updatedabout 1 hour ago • 26 • 1

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.11042 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.11042 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Tmax: A simple recipe for terminal agents

Hugging Face Daily Papers

Tmax introduces a simplified RL training recipe for terminal agents, achieving state-of-the-art performance with a 9B parameter model using a novel data generation taxonomy and an expanded open-source dataset.