Marathoner: Ultra-Long-Horizon Autonomous Intelligence

Marathoner. Ultra-Long-Horizon Autonomous Intelligence. By Zhang Ruiyang and six others, from the University of Macau, Peking University, and one other institution. Published on the twenty-eighth of September. No code repository is linked to it.

People can work persistently toward a long-term goal. Given a hard task, a person may keep at it for months or even years until a specific objective is met. Marathoner sets out to give an autonomous agentic model a version of that: the ability to execute over an ultra-long horizon. The capability has to be instilled into a base model, and that requires tasks hard and long enough to train on.

The authors' answer is a comprehensive post-training pipeline. It has three stages: synthesizing ultra-long-horizon tasks, rejection sampling fine-tuning, and reinforcement learning. A new reward strategy is added on top. Each stage feeds the next. The synthesized tasks are what the teacher's trajectories are generated on, and the model fine-tuned on those trajectories is the starting point for reinforcement learning.

Task synthesis draws on major release pull requests from diverse GitHub repositories, specifically ones that add more than a thousand lines of new code. These are the main source of challenging task-level data. On top of that, Multi-Task Chaining strings several generated tasks together into a single, more challenging one, which lets the synthesis reach difficulty at the frontier level.

For rejection sampling fine-tuning, a strong teacher model is combined with diverse harnesses to generate trajectories on the synthesized tasks. The base model is then fine-tuned on the trajectories that survive rejection sampling. Only the runs that pass the filter become training examples, so the base model learns from the teacher's successful attempts rather than from all of them.

Reinforcement learning starts from that cold-started model. During rollouts, it performs real-world execution through harnesses, each run in its own independent sandbox, which the authors credit with building genuine ultra-long-horizon execution capability. The reward includes a new strategy, Later Stage Bonus Reward, which explicitly encourages the model to make meaningful maneuvers during the later stages of execution.

Evaluation spans five benchmarks containing ultra-long-horizon tasks. Across them, Marathoner achieves consistent and substantial improvements over its base model, and even surpasses a strong proprietary model. Further analysis shows that it can work consistently for more than ten hours, and make more than a thousand tool calls, on highly challenging tasks. Those two figures are the point of the exercise: a session measured in hours, and an agent that is still acting at the end of it.

The name fits the design. Training data comes from pull requests of a thousand lines and more. Tasks are chained into longer ones. Rollouts are executed for real in sandboxes. And the reward pays specifically for meaningful work late in a run. Each piece pushes on endurance, on what the model is still doing hours in, rather than on any single step.