Marathoner: Ultra-Long-Horizon Autonomous Intelligence

A post-training recipe from Ant Group and the University of Macau for coding agents that keep working on one task for hours. Tasks come from major-release pull requests: 100,000 PRs mined from 10,000 GitHub repositories, split 2:3:5 into Easy, Medium and Hard tasks by the amount of new code (100–200 lines, 200–1,000, and over 1,000). Each becomes a Harbor-format task: the repository checked out just before the PR, an instruction taken from the PR's opening comment, the release notes or an LLM reading of the diff, and a verifier made of the PR's new tests plus the existing suite. Multi-Task Chaining merges five such tasks, with all their repositories, into one harder "Frontier" task. For rejection-sampling fine-tuning, Kimi K3 ran 53,000 tasks in three harnesses (Claude Code, Codex and OpenClaw) so the student would not overfit to one; the 40,820 trajectories that passed every test trained Qwen3.5-9B by full-parameter SFT at 256K sequence length. RL then ran on 8,000 further tasks (5,000 direct, 3,000 chained) in Ant's AReaL integrated with Harbor, on 32 H800 GPUs, with a random harness per task and a Later Stage Bonus Reward: Qwen-3.8-Max summarizes each trajectory into phases and flags exceptionally valuable ones, and a flag in the second half of the run adds 0.5 to the binary test reward.

Run in Claude Code, Marathoner-9B scores 26.4 on FrontierSWE, 34.7 on NL2Repo, 8.2 on SWE-Marathon, 57.2 on Terminal-Bench 2.0 and 77.5 on SWE-bench Verified, against 10.2, 17.9, 0, 27.3 and 43.8 for its Qwen3.5-9B base, and above Qwen3.6-27B (21.6, 32.6, 2.6, 56.4, 73.3) on all five. It edges past Gemini-3.1-Pro in Gemini CLI on FrontierSWE (25.9) and SWE-Marathon (5.9) but remains far below the leaders (GPT-6-Astra in Codex: 89.1 and 48.7). Growing the RL pool from 1,000 to 8,000 tasks raised FrontierSWE and Terminal-Bench 2.0 scores steadily, and a case trajectory on a Hard task ran 11.8 hours with 1,273 tool calls. The first author, Ruiyang Zhang, is at the University of Macau and Ant Group; the co-corresponding authors are Qingpei Guo (Ant Group) and Zhedong Zheng (University of Macau), and six of the seven authors list Ant Group. The paper links no weights or code, and none were on Hugging Face or GitHub at filing.