acceptodds
Under review as a conference paper at ICLR 2027

Marathoner: Ultra-Long-Horizon Autonomous Intelligence

Abstract

Humans naturally possess the ability to work persistently toward long-term goals. Given a challenging task, humans can continuously work for months or even years to accomplish a specific objective. Following this spirit, strong proprietary models such as Fable 5 and GPT-6-Astra have placed increasing emphasis on developing such capability, and these models can now continuously work for days to tackle challenging problems. In this paper, we propose Marathoner, an autonomous agentic model possessing the ability of ultra-long-horizon execution. Specifically, we propose a comprehensive post-training pipeline to instill this critical capability into base model, including Ultra-Long-Horizon Task Synthesis, rejection sampling finetuning, and reinforcement learning. For Ultra-Long-Horizon Task Synthesis, we leverage major release PRs containing 1000+ lines of new code from diverse GitHub repositories as the primary source for synthesizing challenging task-level data. Additionally, we introduce Multi-Task Chaining, which chains multiple generated tasks into a single more challenging task, enabling the synthesis of tasks with frontier-level difficulty. For rejection sampling finetuning, we combine strong teacher model with diverse harnesses to generate trajectories on our synthesized tasks and conduct supervised finetuning on base model with rejection sampled trajectories. For reinforcement learning, cold-started model performs real-world execution through harnesses in independent sandboxes during rollout process, effectively facilitating the acquisition of genuine ultra-long-horizon execution capability. We further propose a novel reward strategy, Later Stage Bonus Reward, which explicitly encourages model to perform meaningful maneuvers during later stages of execution. Through extensive evaluation on 5 benchmarks containing ultra-long-horizon tasks, Marathoner achieves consistent and substantial performance improvements over base model and even surpasses performance of strong proprietary model. Further analysis shows that Marathoner can consistently work for 10+ hours and conduct 1000+ tool calls on highly challenging tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Ultra-Long-Horizon, Coding Agent

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cHguOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Marathoner, a method designed to enhance the long-duration execution capabilities of open-source code agents through post-training. The authors construct verifiable tasks from GitHub pull requests and extend task duration by chaining multiple tasks together. They then employ a combination of rejection sampling fine-tuning, reinforcement learning, and "late-stage reward boosting" to encourage the model to continue making meaningful progress during extended execution. Experimental results demonstrate that Marathoner-9B significantly outperforms its base model across five benchmarks involving code and terminal operations, capable of running continuously for hours and performing hundreds of tool calls.

Critical strengths

The scope of work and the research direction are commendable. This paper presents large-scale synthesis tasks, corresponding executable and verifiable training environments, and a complete post-training pipeline integrating RFT and RL. Marathoner-9B demonstrates significant improvements over the base model across 5 benchmarks.

Critical weaknesses
  1. The two key methodological innovations lack compelling evidence. "Multi-Task Chaining" essentially concatenates independent tasks rather than addressing tasks with genuine long-term dependencies; the "Later Stage Bonus Reward" relies on the subjective judgment of an LLM without verification of human alignment or reward reliability.

  2. The primary evaluation metric remains the final task success rate; execution time, step count, and tool call frequency are presented merely as descriptive statistics for Marathoner, without corresponding metrics for other baselines. A longer execution process could signify effective, sustained progress, or it could imply lower efficiency—such as numerous repetitive or futile attempts. Consequently, the current experiments cannot clearly distinguish whether performance gains stem from enhanced ultra-long-horizon capabilities, general improvements in programming proficiency, increased test-time computation, or even a model bias toward continuing execution rather than terminating early.

Questions for authors
  1. How did the authors ensure that the 100k training PRs did not overlap with the repositories, issues, or ground-truth solutions found in the five evaluation benchmarks? Was any deduplication performed?

  2. How many trajectories actually exceeded 10 hours or 1,000 tool calls? What proportion of these trajectories ultimately resulted in task completion? Can the authors report the distributions of execution time, execution steps, and tool call counts separately for the base model, the RFT model, and the final model, while distinguishing between successful and failed trajectories? The currently reported average execution times—3.56 hours on FrontierSWE and 0.58 hours on Terminal Bench 2.0—still fall short of the core claim regarding "sustained execution for over 10 hours."

  3. The "Multi-Task Chaining" strategy involves randomly chaining multiple unrelated tasks; while this certainly increases the total workload, it does not necessarily require sustained reasoning or long-range planning from the model. The model might simply be completing a series of independent sub-tasks one after another. Consequently, performance gains resulting from multi-task chaining cannot directly demonstrate superior long-horizon credit assignment, memory, or planning capabilities.

  4. In "Environment Scaling" experiments, were the number of training steps and the total rollout computational cost kept constant? If not, how can it be proven that the performance improvements stem primarily from environmental diversity rather than increased training computation?

  5. Can the authors report the total token count, number of context overflows, and number of compression events for the case study involving "11.8 h, 872 steps, and 1,273 tool calls."

5 comments
AuthorsOpenReview ↗

Response to Reviewer cHgu (Part 1/5)

Comment(1): "Multi-Task Chaining" essentially concatenates independent tasks rather than addressing tasks with genuine long-term dependencies

Response(1): Thanks for your valuable suggestion. We address this question from two perspectives.

(1) Advanced Multi-Task Chaining. Following your suggestion, we further develop an advanced Multi-Task Chaining strategy to synthesize tasks whose constituent subtasks exhibit explicit long-term logical dependencies. Specifically, given a selected GitHub repository, we first identify its major-release PRs and construct a PR list containing the titles and descriptions of these PRs. We then leverage an LLM to select groups of five major-release PRs that exhibit logical dependencies with one another. Such dependencies may involve building more complex functionalities upon capabilities introduced by previous PRs, refactoring the implementation of functionalities introduced by earlier PRs, or combining capabilities implemented in multiple previous PRs to introduce new functionalities. This selection process is based on the extracted PR titles and descriptions. For each GitHub repository, multiple groups of five interdependent major-release PRs can potentially be identified. We then synthesize challenging tasks based on these selected PR groups. Specifically, we first construct an individual task corresponding to each PR and subsequently chain these tasks together using our original Multi-Task Chaining strategy. This advanced strategy produces Frontier tasks whose constituent subtasks exhibit explicit logical dependencies, making them particularly challenging because the agentic model must maintain long-term plans, preserve execution context and memory, and coordinate multiple dependent implementation stages to successfully complete the overall task.

In practice, we construct 6,000 such tasks from 2,000 GitHub repositories. We incorporate 3,000 tasks into the RFT training stage and another 3,000 into the RL training stage, while retaining the entire original training task pool. All other training settings and hyperparameters remain unchanged. We refer to this new strategy as Advanced Multi-Task Chaining. The experimental results are presented below.

ModelsFrontierSWETerminal Bench 2.0
Marathoner-9B25.457.2
Marathoner-9B w. Advanced Multi-Task Chaining27.658.4

We observe that this advanced Multi-Task Chaining strategy leads to further improvements in ultra-long-horizon execution capabilities. Tasks synthesized through Advanced Multi-Task Chaining contain subtasks with explicit logical dependencies, requiring the agentic model to perform more sophisticated long-term planning and maintain execution context and memory to successfully complete the overall task. During RFT, these tasks encourage the teacher model to demonstrate sophisticated long-term planning and execution behaviors, which can then be transferred to the student model through high-quality supervision. During RL, the agentic model further practices and refines these capabilities by solving complex tasks with long-term dependencies in real-world sandbox environments, leading to more effective acquisition of ultra-long-horizon execution capabilities. We will incorporate these additional discussions into Appendix C.

(2) During the design of our method, we also considered chaining tasks across mutually related GitHub repositories. Our initial idea was to manually design rules to identify repositories with complementary functionalities, such as front-end, back-end, and database repositories. We could then construct complex tasks by chaining tasks from these different repositories, resulting in highly challenging tasks with long-term dependencies across multiple codebases. For example, an agent might need to modify the database schema, update the corresponding back-end protocols, and finally adapt the front-end interface accordingly. However, this approach requires extensive hand-crafted rules to identify relationships between repositories and construct meaningful cross-repository dependencies, making it relatively inflexible and difficult to scale. Therefore, we ultimately adopted our current Multi-Task Chaining design, which is more scalable and suitable for large-scale task synthesis. Moreover, it generates highly diverse tasks across different repositories and domains, facilitating the development of more generalizable and robust ultra-long-horizon execution capabilities.

AuthorsOpenReview ↗

Response to Reviewer cHgu (Part 2/5)

Comment(2): the "Later Stage Bonus Reward" relies on the subjective judgment of an LLM without verification of human alignment or reward reliability.

Response(2): Thanks for your insightful question. Yes, the Later Stage Bonus Reward is currently assigned based on LLM-based subjective judgments. Following your suggestion, we conducted an additional human evaluation to assess the reliability of these judgments and their agreement with human assessments. Specifically, we recruited five human experts with STEM backgrounds to evaluate the reward assignments. We randomly sampled 100 trajectories from Marathoner's RL training process and assigned 20 trajectories to each expert. For each trajectory, the expert was instructed to carefully examine the execution process, summarize it into several major stages with concise descriptions, and assign a binary flag (0 or 1) to each stage to indicate whether it contained highly valuable operations. If at least one stage in the second half of the execution received a flag of 1, the trajectory was assigned a Later Stage Bonus Reward of 0.5; otherwise, the bonus reward was 0. We then compared the reward assignments made by human experts with those generated by the LLM. The results are presented below. Specifically, we report the percentage of trajectories for which human experts and the LLM agree on the reward assignment, i.e., both assign a bonus reward of either 0.5 or 0.

Human Expert #1Human Expert #2Human Expert #3Human Expert #4Human Expert #5Average
Calibration Result90%95%85%100%95%93%

We observe that LLM-based judgments of the Later Stage Bonus Reward exhibit high agreement with human expert assessments. This demonstrates the reliability of using LLMs to assign the Later Stage Bonus Reward and provide effective supervision for encouraging the agentic model to perform meaningful and valuable operations during later stages of execution. Notably, we adopt Qwen3.7-Max as the LLM judge for the Later Stage Bonus Reward during RL training. Qwen3.7-Max is itself a strong agentic coding model with substantial ultra-long-horizon execution capabilities, making it well-suited to identifying valuable operations and meaningful progress within long-horizon execution trajectories. Its strong coding and long-horizon reasoning capabilities may account for the high reliability of its judgments and the strong agreement with human expert assessments. We will incorporate these additional analyses into Appendix C.

Comment(3): The primary evaluation metric remains the final task success rate; execution time, step count, and tool call frequency are presented merely as descriptive statistics for Marathoner, without corresponding metrics for other baselines.

Response(3): Thanks for your insightful comment. Following your suggestion, we provide additional execution-horizon metrics for the baselines to facilitate a more comprehensive comparison, including average execution time, average number of steps, and average number of tool calls. The results are presented below:

ModelsHarnessFrontierSWETerminal Bench 2.0
MetricAvg. Execution TimeAvg. StepsAvg. Tool CallsAvg. Execution TimeAvg. StepsAvg. Tool Calls
GPT-6-AstraCodex12.84h895.31637.31.18h149.3493.7
Claude Fable 5.1Claude Code10.27h723.61438.21.27h163.8583.3
Kimi K3Claude Code9.07h638.41294.60.76h137.2403.8
Marathoner-9BClaude Code3.56h426.2648.30.58h104.8239.4

We observe that Marathoner still falls behind flagship proprietary models, such as GPT-6-Astra and Claude Fable 5.1, in terms of average execution time, number of execution steps, and tool-call count. These frontier models demonstrate stronger capabilities in sustaining ultra-long-horizon execution and persistent task solving, which contribute to their higher task completion rates. These results suggest that Marathoner does not simply rely on excessive tool calls or unnecessarily prolonged execution to solve tasks. Instead, there remains substantial room to further strengthen its ultra-long-horizon execution capabilities, enabling it to sustain more sophisticated and persistent problem-solving processes and achieve better performance on challenging tasks.

AuthorsOpenReview ↗

Response to Reviewer cHgu (Part 3/5)

Comment(4): How did the authors ensure that the 100k training PRs did not overlap with the repositories, issues, or ground-truth solutions found in the five evaluation benchmarks? Was any deduplication performed?

Response(4): Thanks for your valuable question. We address this concern from two perspectives. (1) Following your suggestion, we conducted a thorough contamination analysis of our synthesized training tasks against all evaluation benchmarks. For FrontierSWE, NL2Repo, SWE-Marathon, and Terminal-Bench 2.0, which are manually constructed and curated by human experts, we primarily examine repository-level overlap between the repositories involved in benchmark tasks and those used for our training task synthesis. For SWE-Bench Verified, which is constructed based on GitHub issues, we additionally examine issue-level overlap.

Specifically, for FrontierSWE, NL2Repo, SWE-Marathon, and Terminal-Bench 2.0, we first identify all GitHub repositories associated with the benchmark tasks, including repositories directly used or mentioned as references. We then compare these repositories against the 10,000 GitHub repositories used in our training task synthesis. We find that only 17 out of the 10,000 training repositories overlap with repositories associated with these benchmarks. Furthermore, through manual inspection of all overlapping cases, we find that the task instructions and usage scenarios involving these repositories are substantially different between the benchmark tasks and our synthesized training tasks, indicating that the actual task-level duplication is very limited.

For SWE-Bench Verified, we first extract all PRs used in our training data that are associated with GitHub issues and identify their corresponding issues and repositories. We then compare both repository-level and issue-level overlap with those associated with SWE-Bench Verified. We find that only two tasks in our training data have PR-associated issues that overlap with issues in SWE-Bench Verified. Notably, the overlapping SWE-Bench Verified issues generally involve relatively simple problems and require only a small number of code changes, whereas our synthesized training tasks require PRs involving at least 100 lines of newly added code. Therefore, despite the issue-level overlap, the actual task requirements and implementation complexity are substantially different.

(2) To further eliminate potential contamination concerns, we perform complete deduplication of our training task pool, removing all tasks synthesized from repositories overlapping with the evaluation benchmarks, as well as tasks associated with overlapping GitHub issues. We then conduct both RFT and RL training using this deduplicated training dataset while keeping all other training settings unchanged. The experimental results are presented below:

ModelsHarnessFrontierSWENL2RepoSWE-MarathonTerminal Bench 2.0SWE-Bench Verified
Marathoner-9BClaude Code25.434.78.257.277.5
Marathoner-9B w. Data DeduplicationClaude Code25.634.68.557.177.4

We observe no substantial performance degradation in Marathoner after deduplicating the training data. This suggests that the effectiveness of our post-training pipeline does not rely on potential training–evaluation data overlap. Instead, our pipeline develops genuine and robust ultra-long-horizon execution capabilities that generalize to challenging benchmark tasks, including those associated with repositories or issues that potentially overlap with the original training data.

Reviewer_4LngOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Marathoner-9B, a Qwen3.5-9B post-trained model for ultra-long-horizon agentic coding tasks. The corresponding end-to-end pipeline covers (1) long-horizon task synthesis (2) sft finetuning using rejection-sampled trajectories by Kimi-K3 (3) RL training using GRPO driven by the fixed ratio of final result correctness and a new "later stage bonus". On five benchmarks, Marathoner-9B shows performance improvements at both sft and RL phases. However, the details of task construction and reward design are insufficiently justified and the experimental designs are also needed reconsiderations and enhancements.

Critical strengths
  1. Important problem with a scalable, verifiable task-construction recipe. The author open-sources their recipes for multi-hour autonomous execution with each component detail, esp. how they synthesize challenging tasks, generate rollout trajectories under different harness frameworks, etc.. This can serves as a useful reference for long-horizon task synthesis and agentic post-training in other domains.
Critical weaknesses
  1. Method
    • Task construction: Task difficulty is not calibrated to the target QWen3.5-9B model, and the claimed curriculum is not realized. (1) The current difficulty evaluation is based on code line for major release PR mining, which may not reflect what is actually hard for Qwen3.5-9B. (2) The rejection sampling uses the stronger Kimi-K3 under harness frameworks, so retained tasks reflect the teacher's capability rather than the student's. This makes the claim in §2.1 that the tiers enable the agent to "progressively extend" its capability unconvincing.
  2. Experiments
    • No variance reported: No standard deviation of performance statistics are reported in Table 1 and 3, and no corresponding reflects are drawn in figure 3. Suggestion: report mean ± std (or 95% CI) over multiple evaluation runs (ideally multiple training seeds for the ablations), and add error bars to Figure 3.
    • Ablation design: "Marathoner (vanilla)" is undefined (which harness? which data?), components are only added one at a time, and the per-component gains (+1.4 to +3.5) are not shown to be beyond noise. A leave-one-out ablation with variance is needed.
    • The claimed "ultra-long-horizon" capability is overclaimed: The "10+ hours / 1000+ tool calls" claim rests on a single case study (Appendix F). Average execution is 3.56 h on FrontierSWE and 0.58 h on TB2 (Table 2), and SWE-Bench Verified is not a long-horizon benchmark. To convince the claim, it's better to report related statistics, such as duration/tool-call distributions per benchmark, success as a function of horizon, and chained tasks whose subtasks depend on each other, etc..
Questions for authors
  1. Task construction
    • Q1: What is Qwen3.5-9B's pass rate on each difficulty tier, before and after RFT?
    • Q2: How are the 40,820 retained RFT trajectories distributed across the Easy/Medium/Hard/Frontier tiers?
    • Q3: Is there any curriculum schedule over difficulty tiers, or is the task mixture fixed throughout training?
  2. Experiments
    • Q4: Are the results in Tables 1, 3, 4, and 6 from a single run? What is the variance across runs?
    • Q5: How is "Marathoner (vanilla)" in Table 3 configured: which harness(es), and which training data?
    • Q6: Beyond the case study in Appendix F, how often does Marathoner actually run for 10+ hours or make 1000+ tool calls on each benchmark?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 4Lng (Part 1/5)

Comment(1): The current difficulty evaluation is based on code line for major release PR mining, which may not reflect what is actually hard for Qwen3.5-9B.

Response(1): Thanks for your valuable comment. Yes, our current task difficulty classification is based on the number of code lines in the corresponding PR diff. Following your suggestion, we conducted a further quantitative analysis to evaluate the performance of Qwen3.5-9B across our synthesized Easy, Medium, Hard, and Frontier tasks. Specifically, we randomly sampled 1,000 tasks from each difficulty level and evaluated Qwen3.5-9B with Claude Code on these tasks. The evaluation results are presented below:

ModelHarnessEasyMediumHardFrontier
Qwen3.5-9BClaude Code29.4 ± 0.613.5 ± 0.82.8 ± 0.50

We observe a clear and consistent decline in Qwen3.5-9B's performance as task difficulty increases from Easy to Frontier. This quantitatively demonstrates that our difficulty classification based on PR diff size effectively captures the actual difficulty of tasks for the model. By mixing tasks across different difficulty levels, our training pipeline enables progressive learning, allowing the model to gradually extend its capabilities from short-horizon execution to increasingly challenging ultra-long-horizon task solving.

We further provide an auxiliary ablation study to investigate the benefits of training with mixed-difficulty tasks. Specifically, we conduct additional RL training runs using only Easy, Medium, or Hard tasks, respectively, while keeping the total number of training tasks identical to that of our original mixed-difficulty RL training dataset. The results are presented below:

RL Training DataFrontierSWETerminal Bench 2.0
Easy21.5 ± 0.752.7 ± 0.8
Medium22.6 ± 0.554.9 ± 0.6
Hard20.8 ± 0.951.7 ± 0.7
Mixed25.2 ± 0.657.4 ± 0.5

We observe that training with mixed-difficulty tasks yields superior performance compared to training with tasks from any single difficulty level. Mixing tasks of different difficulties facilitates a more effective learning process, allowing the model to first develop proficiency in relatively simple problem-solving and progressively strengthen its capabilities for more challenging tasks. In contrast, training exclusively on Easy tasks limits the model's ability to handle complex tasks, while training exclusively on Hard tasks provides insufficient learning signals due to the relatively low frequency of successful task completion, resulting in a slower and less effective learning process. We will incorporate these additional ablation results and analyses into Appendix C.

Comment(2): The rejection sampling uses the stronger Kimi-K3 under harness frameworks, so retained tasks reflect the teacher's capability rather than the student's.

Response(2): Thanks for your insightful comments. Following your suggestion, we conducted a statistical analysis to examine how Qwen3.5-9B performs on tasks that Kimi-K3 successfully solves and those on which Kimi-K3 fails. Specifically, we randomly sampled 1,000 tasks from the RFT training task pool, evaluated Qwen3.5-9B with Claude Code on these tasks, and compared its performance with that of Kimi-K3 using the same harness. We then analyzed Qwen3.5-9B's performance separately across the two groups of tasks. The results are presented below:

Teacher Correct & Student CorrentTeacher Fail & Student Fail
6.2%98.3%

We observe that Qwen3.5-9B performs poorly on tasks that Kimi-K3 successfully solves. This demonstrates that successful trajectories generated by Kimi-K3 can provide effective supervision and learning signals for Qwen3.5-9B during the RFT stage, thereby improving its ultra-long-horizon execution capabilities. Moreover, Qwen3.5-9B also fails on the majority of tasks where Kimi-K3 fails, suggesting that its task-solving capabilities largely remain within the capability range covered by Kimi-K3. We sincerely Thanks again for this valuable question, which has helped us strengthen our paper by providing a quantitative analysis and deeper insights into the underlying mechanism of the RFT process.

AuthorsOpenReview ↗

Response to Reviewer 4Lng (Part 2/5)

Comment(3): No standard deviation of performance statistics are reported in Table 1 and 3, and no corresponding reflects are drawn in figure 3.

Response(3): Thanks for your valuable suggestion. Following your advice, we repeated all evaluations in Table 1 five times and calculated the corresponding mean and standard deviation. We will update Table 1 accordingly to report the results in a mean ± standard deviation format:

ModelsHarnessFrontierSWENL2RepoSWE-MarathonTerminal Bench 2.0SWE-Bench Verified
Proprietary Model with Harness
GLM-5.2Claude Code67.5 ± 0.668.1 ± 0.813.8 ± 0.572.3 ± 0.982.7 ± 0.4
Qwen-3.7-MaxClaude Code48.0 ± 0.741.8 ± 0.510.1 ± 0.858.4 ± 0.679.0 ± 0.9
Kimi K3Claude Code53.9 ± 0.548.7 ± 0.916.5 ± 0.668.5 ± 0.784.0 ± 0.8
GPT-6-AstraCodex89.3 ± 0.478.0 ± 0.648.9 ± 0.787.2 ± 0.590.8 ± 0.9
Claude Fable 5.1Claude Code86.0 ± 0.873.4 ± 0.542.6 ± 0.986.1 ± 0.693.6 ± 0.4
Gemini-3.1-ProGemini CLI35.1 ± 0.742.6 ± 0.86.1 ± 0.582.6 ± 0.988.5 ± 0.6
Open Source Model with Harness
Qwen3.5-4BClaude Code5.9 ± 0.53.0 ± 0.707.3 ± 0.837.4 ± 0.6
Qwen3.5-9BClaude Code6.0 ± 0.84.9 ± 0.509.1 ± 0.744.0 ± 0.9
Qwen3.5-35B-A3BClaude Code8.5 ± 0.67.0 ± 0.9012.4 ± 0.847.5 ± 0.7
Qwen3.6-27BClaude Code21.4 ± 0.932.8 ± 0.62.8 ± 0.756.2 ± 0.575.1 ± 0.8
Qwen3.6-35B-A3BClaude Code19.4 ± 0.530.5 ± 0.81.5 ± 0.648.5 ± 0.972.6 ± 0.7
Qwen3.8-27BClaude Code26.2 ± 0.740.5 ± 0.55.9 ± 0.871.0 ± 0.678.6 ± 0.9
Ultra-Long-Horizon Agentic Model
Marathoner-9BClaude Code25.6 ± 0.834.5 ± 0.68.4 ± 0.957.0 ± 0.577.7 ± 0.7

For Table 3, we conducted RL training three times for each ablation setting using different random seeds and report the corresponding mean and standard deviation. We also restructured the ablation study into a leave-one-out design. The results are presented below:

SettingFrontierSWETerminal Bench 2.0
Marathoner25.6 ± 0.657.2 ± 0.9
Marathoner wo. Multi-Task Chaining23.1 ± 0.554.7 ± 0.8
Marathoner wo. Diverse-Harness Trajectory Generation22.7 ± 0.853.1 ± 1.2
Marathoner wo. Later Stage Bonus Reward22.3 ± 0.453.6 ± 0.7

For Figure 3, we also repeated RL training three times for each scaling setting using different random seeds. The corresponding results are shown in xxx. We will update Tables 1 and 3 and Figure 3 to include the mean and standard deviation across multiple runs, providing a more comprehensive assessment of the stability and consistency of our results.

Comment(4): *"Marathoner (vanilla)" is undefined (which harness? which data?), components are only added one at a time, and the per-component gains (+1.4 to +3.5) are not shown to be beyond noise. A leave-one-out ablation with variance is needed. *

Response(4): Thanks for your valuable suggestion. We address this question from two perspectives: the experimental settings of Marathoner (Vanilla) and the leave-one-out ablation results with variance. (1) Marathoner (Vanilla) is a simplified version of Marathoner, trained without Multi-Task Chaining, Diverse-Harness Trajectory Generation, or Later Stage Bonus Reward. Specifically, it is trained exclusively on Easy, Medium, and Hard tasks during both the RFT and RL stages, without incorporating Frontier tasks introduced through Multi-Task Chaining. During RFT, the base model is trained on trajectories generated by Kimi-K3 using Claude Code as the harness. During RL, the model is trained solely with the task correctness reward. All evaluation results are obtained using Claude Code as the harness. (2) Following your suggestion, we conduct additional ablation studies using a leave-one-out design, which provides a more systematic and informative assessment of the individual contributions of different components. Specifically, we perform RL training three times for each ablation setting using different random seeds and report the corresponding variance. All ablation settings are evaluated using Claude Code as the harness. The results are presented below:

SettingFrontierSWETerminal Bench 2.0
Marathoner25.6 ± 0.657.2 ± 0.9
Marathoner wo. Multi-Task Chaining23.1 ± 0.554.7 ± 0.8
Marathoner wo. Diverse-Harness Trajectory Generation22.7 ± 0.853.1 ± 1.2
Marathoner wo. Later Stage Bonus Reward22.3 ± 0.453.6 ± 0.7

We observe that Marathoner consistently achieves superior performance compared to the ablation settings, each of which removes one of our proposed components. These results demonstrate the effectiveness of Multi-Task Chaining, Diverse-Harness Trajectory Generation, and Later Stage Bonus Reward in enhancing ultra-long-horizon execution capabilities. We will replace Table 3 with this updated leave-one-out ablation table.

AuthorsOpenReview ↗

Response to Reviewer 4Lng (Part 3/5)

Comment(5): To convince the claim, it's better to report related statistics, such as duration/tool-call distributions per benchmark, success as a function of horizon, and chained tasks whose subtasks depend on each other, etc..

Response(5): Thanks for your constructive advice. We address this concern from three perspectives. (1) Following your suggestion, we conduct a statistical analysis of Marathoner-9B's execution trajectories on FrontierSWE and Terminal Bench 2.0, examining the distributions of execution duration and tool-call counts. We also include the corresponding distributions of the base Qwen3.5-9B for comparison. The results are available at https://anonymous.4open.science/r/Marathoner_Rebuttal-C7F1/Distribution_Execution_Time_Tool_Call.pdf. On FrontierSWE, we observe long-tailed distributions for both execution duration and tool-call counts. In particular, Marathoner-9B has multiple instances where it executes continuously for over 10 hours or performs more than 1,000 tool calls. This demonstrates that ultra-long-horizon execution is not merely an occasional phenomenon, but a genuine problem-solving capability that Marathoner can effectively exercise on challenging tasks. Compared with the base Qwen3.5-9B, Marathoner exhibits a substantial shift toward longer execution durations and higher tool-call counts, demonstrating the effectiveness of our proposed post-training pipeline in developing ultra-long-horizon execution capabilities. On Terminal Bench 2.0, we also observe several instances where Marathoner executes for more than two hours. The overall execution duration is considerably shorter than that on FrontierSWE, primarily because Terminal Bench 2.0 contains relatively less challenging tasks that generally do not require such extensive execution. (2) We further analyze the relationship between execution horizons and task-solving success on FrontierSWE and Terminal Bench 2.0. The results are available at https://anonymous.4open.science/r/Marathoner_Rebuttal-C7F1/Success_vs_Horizon.pdf. Specifically, we present the cumulative success rates across different execution horizons for both benchmarks. We observe that successful trajectories involving ultra-long execution account for a substantial proportion of Marathoner's overall successful executions. Compared with Qwen3.5-9B, Marathoner also exhibits a clear increase in the execution duration of successful trajectories. These results demonstrate that Marathoner can effectively sustain execution over extended periods and eventually solve challenging tasks successfully. Moreover, Marathoner achieves substantially higher overall success rates than Qwen3.5-9B on both benchmarks. (3) We do not intend to overclaim the ultra-long-horizon capabilities of Marathoner. Through our case studies and practical deployments, we have observed multiple instances in which Marathoner sustains execution for several hours and eventually solves challenging tasks successfully. Our original statement in the Abstract is: "Further analysis shows that Marathoner can consistently work for 10+ hours and conduct 1000+ tool calls on highly challenging tasks."

Comment(6): What is Qwen3.5-9B's pass rate on each difficulty tier, before and after RFT?

Response(6): Thanks for your valuable question. Following your suggestion, we conducted evaluations of Qwen3.5-9B before and after the RFT stage using Claude Code as the harness. Specifically, we randomly sampled 1,000 tasks from each of the four difficulty levels (Easy, Medium, Hard, and Frontier) in our synthesized task pool and evaluated the model on these tasks before and after RFT to quantify the improvements brought by RFT across different difficulty levels. Since our full task pool contains 100,000 tasks, evaluating all tasks would incur substantial computational costs and require considerable time. The evaluation results are presented below:

ModelHarnessEasyMediumHardFrontier
Qwen3.5-9BClaude Code29.4 ± 0.613.5 ± 0.82.8 ± 0.50
Qwen3.5-9B-RFTClaude Code61.0 ± 0.742.9 ± 0.616.1 ± 0.94.4 ± 0.5

We make two key observations from these results. (1) Qwen3.5-9B exhibits a clear and gradual decline in performance as task difficulty increases from Easy to Frontier. This quantitatively demonstrates that our task difficulty classification, based on the number of code lines in PR diffs, effectively captures the actual difficulty of tasks for Qwen3.5-9B. Consequently, mixing tasks across different difficulty levels can facilitate progressive learning, enabling the model to gradually extend its ultra-long-horizon execution capabilities. (2) After RFT, the 9B-scale model achieves substantial performance improvements across all four difficulty levels. This demonstrates the effectiveness of our RFT process in instilling foundational ultra-long-horizon task-solving capabilities into the base model.

Reviewer_fXWDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to train a 9B powerful ultra-long-horizon agentic model with a very simple receipt. The training process is very simple only using SFT and GRPO algorithms. The performance of the trained model is great compared with Gemini-3.1-Pro.

Critical strengths

This paper studies an very important problem: how to train a small model with ultra-long-horizon agentic ability. This paper proposes solutions to collect training data, and develops a very simple receipt to train 9B models with SFT and RL. The performance of

Critical weaknesses
  1. Is it possible to release GitHub repositories you collected?
  2. Is it possible to train Marathoner-4B, currently, you only trained one 9B model which is not convincing?
  3. Could you split the results into two stages: Qwen3.5-9B-Base + SFT, Qwen3.5-9B-Base + SFT + RL? I mean, at table 1, could you show two lines to make reader understand the impact of SFT and RL seperately?
  4. I am curious about whether claude code are open-source or not :) If not, why you can use the claude code as harness?
Questions for authors

Please check critical weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer fXWD (Part 1/2)

Comment (1): "Is it possible to release GitHub repositories you collected?"

Response (1): Thanks for your valuable suggestions and thorough review. Yes, we are happy to release our collected GitHub repositories. We have made the list publicly available at https://anonymous.4open.science/r/Marathoner_Rebuttal-C7F1/repository.txt. During the repository collection process, we intentionally prioritized diversity across different domains and types of codebases. These repositories serve as the initial sources for our Ultra-Long-Horizon Task Synthesis pipeline. A diverse repository pool provides broader coverage and richer source materials for task synthesis, facilitating the development of more robust and generalizable ultra-long-horizon execution capabilities. Specifically, our collected repositories span a wide range of domains and characteristics, including front-end and back-end development, low-level programming, mature and large-scale codebases, repositories from GitHub Trending, and LLM-related projects, among others.

Comment (2): "Is it possible to train Marathoner-4B"

Response (2): Thanks for your insightful comments. Yes, this is indeed possible. In fact, our initial exploration started with a 4B-scale model. After observing promising results at this scale, we subsequently extended our experiments to 9B-scale models. We present the experimental results for the 4B-scale model below. Specifically, we use Qwen3.5-4B as the base model and adopt the same post-training pipeline, training data, and experimental settings as those used for the 9B-scale model.

ModelHarnessFrontierSWENL2RepoSWE-MarathonTerminal Bench 2.0SWE-Bench Verified
Qwen3.5-4BClaude Code5.73.207.137.6
Marathoner-4B-RFTClaude Code8.610.30.424.847.3
Marathoner-4BClaude Code17.426.11.243.968.2

We observe that Marathoner-4B achieves clear and consistent performance improvements over Qwen3.5-4B across all benchmarks. These results demonstrate the effectiveness of our proposed post-training pipeline in instilling ultra-long-horizon execution capabilities into vanilla base models. Notably, we use the same training data and experimental settings for Marathoner-4B and Marathoner-9B, without any model-specific tuning, further demonstrating the robustness and scalability of our proposed pipeline. Moreover, we also deploy Marathoner-4B on a local server and evaluate it on several real-world coding tasks, where we observe promising ultra-long-horizon execution capabilities in practice. We have incorporated these additional results into Section 3.2 to provide a more comprehensive evaluation in the main paper.

Comment (3): "Could you split the results into two stages: Qwen3.5-9B-Base + SFT, Qwen3.5-9B-Base + SFT + RL?"

Response (3): Thanks for your thoughtful suggestion. Yes, we can do that. We will update Table 1 accordingly. Specifically, Marathoner-9B-RFT denotes Qwen3.5-9B-Base post-trained with SFT, while Marathoner-9B denotes Qwen3.5-9B-Base post-trained with both SFT and RL.

ModelsHarnessFrontierSWENL2RepoSWE-MarathonTerminal Bench 2.0SWE-Bench Verified
Proprietary Model with Harness
GLM-5.2Claude Code71.347.213.682.587.5
Qwen-3.7-MaxClaude Code58.244.610.365.879.2
Kimi K3Claude Code79.848.935.986.486.2
GPT-6-AstraCodex89.178.248.787.490.7
Claude Fable 5.1Claude Code86.273.242.885.991.8
Gemini-3.1-ProGemini CLI25.942.85.960.380.1
Open Source Model with Harness
Qwen3.5-4BClaude Code5.73.207.137.6
Qwen3.5-9BClaude Code10.217.9027.343.8
Qwen3.5-35B-A3BClaude Code9.315.1022.642.3
Qwen3.6-27BClaude Code21.632.62.656.473.3
Qwen3.6-35B-A3BClaude Code19.230.71.348.371.8
Ultra-Long-Horizon Agentic Model
Marathoner-9B-RFTClaude Code18.223.82.949.758.4
Marathoner-9BClaude Code26.434.78.257.277.5

We observe that applying RFT to Qwen3.5-9B already yields clear performance gains across all benchmarks. Building upon these initial improvements, RL further generalizes the foundational ultra-long-horizon execution capabilities acquired through RFT by allowing the agent to practice and refine its abilities in real-world sandbox environments, leading to further performance gains. Overall, across all our evaluated benchmarks, RL contributes more substantial improvements in ultra-long-horizon execution capabilities than RFT.

AuthorsOpenReview ↗

Response to Reviewer fXWD (Part 2/2)

Comment (4): "I am curious about whether claude code are open-source or not :) If not, why you can use the claude code as harness?"

Response (4): Thanks for your valuable question. Claude Code has been open-sourced by Anthropic (https://github.com/anthropics/claude-code). Specifically, we can use Claude Code as an agent harness because its backend model can be replaced with a locally served model. In the standard usage scenario, Claude Code is launched directly with Anthropic's official models. However, its backend model can be replaced by overriding the relevant configurations, including the API endpoint and model name. In practice, we first deploy our trained model checkpoint on a local server using vLLM and obtain the corresponding API endpoint. We then use LiteLLM to set up a proxy that translates the API requests into a format compatible with Claude Code. Finally, we override Claude Code's default model endpoint to point to our locally served model, enabling us to perform operations such as model evaluation and trajectory generation. The same procedure applies to other agent harnesses, such as Codex and OpenClaw. We will incorporate these implementation details into Appendix B to clarify the process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.