acceptodds
Under review as a conference paper at ICLR 2027

Scaling Traces Diversity with Runtime-Decoupled World Model for Mid-training

12 views

Abstract

Frontier models increasingly incorporate agentic experience earlier—from post-training into mid-training—to build fundamentals in agentic behaviors and environment interaction. This stage teaches broad patterns of exploration, action, and recovery before specialized agentic SFT and RL, creating demand for both high-quality trajectories and diverse, distinct tasks. Existing execution-based pipelines provide high-quality traces but rely on curated, runnable environments and verifiable outcomes. In agentic coding, this concentrates supervision on resolved issues and merged pull requests, leaving open issues, development branches, and tasks with unusable images or non-reproducing gold solutions underrepresented. Our re-execution audit estimates that roughly one-third of SWE-rebench V2 tasks fall outside reliable execution-based supervision, while over 55M unresolved public issues offer an additional source of tasks without known solutions. Motivated by these gaps, we use Qwen-World-Model to simulate execution-dependent feedback while retaining exact repository state for deterministic actions. This enables trajectory collection without requiring every task to have a working image or known solution. We construct a corpus spanning 3M unique tasks and 100 B training tokens, increasing unique-task coverage by 4× at 4× the throughput of execution-based collection. In matched-budget SFT, replacing 25%, 50%, or 100% of executed traces with simulated traces yields comparable performance across four agent benchmarks, supporting their utility as training data. At matched token exposure, adding traces from new queries outperforms repeating the original collection by 4.6 pass@1 points on average, indicating that broader task coverage is more valuable than repeated rollouts at a fixed token budget. Further incorporating these diverse traces into mid-training improves pass@1 after the same SFT by +8.7, +6.2, +2.1, and +1.1 percentage points on SWE-bench Verified, Multilingual, Pro, and Terminal-Bench 2.0, respectively. These results suggest that world-model simulation can expand both the volume and breadth of agentic experience, strengthening the foundations learned during mid-training and improving downstream agent performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language models, Mid-training, Synthetic Data Generation, World Model

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_cit1OpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies scaling interactive trajectories for code-agent mid-training. To reduce the reliance of existing data generation methods on executable sandbox environments, it proposes the World Rollout framework, which constructs agent trajectories from GitHub open issues. The proposed pipeline is evaluated through mid-training experiments, real-sandbox patch verification, and agent behavior analysis.

Critical strengths
  1. This work addresses the challenge of scaling interaction trajectory generation for code-agent mid-training by constructing training data from open GitHub issues, providing a practical and scalable solution with strong practical relevance.

  2. A systematic evaluation, including mid-training experiments, real-environment validation, and agent behavior analysis, demonstrates the effectiveness of the proposed data generation framework from multiple complementary perspectives.

Critical weaknesses

1.The claimed four to sixfold token throughput improvement is based on resource modeling rather than end-to-end measurements. The abstract and main text do not clearly distinguish theoretical estimates from empirical results, leaving the actual acceleration gain unverified.

2.Simulated runtime feedback may cause distribution shift. Only final patches are validated, leaving intermediate observations unverified by real execution. Moreover, the experiments cannot distinguish gains from increased task diversity from those due to simulated trajectories.

Questions for authors

1.The standalone benefit of World Rollout-generated interaction trajectories remains unclear. Could the authors compare SFT performance using only issue text, real execution trajectories, and simulated execution trajectories under a fixed data budget?

2.The advantages of World Rollout over SWE-World remain unclear. Could the authors provide comparative experiments under matched task settings and training configurations, or a more detailed qualitative analysis?

3.Please analyze the modest performance gains on Terminal-Bench-2.0 and clarify whether they stem primarily from limited simulation fidelity or the inherent characteristics of the benchmark tasks.

4.Performance degrades when all trajectories are replaced with simulated ones in the WR100 setting. Please discuss the practical deployment risks when fully relying on simulated trajectories for training.

5.Please clarify which Terminal-Bench version (2.0 or 2.1) is used in the experiments, as the manuscript refers to both. The version number should be consistent throughout the paper and explicitly specified in the experimental setup.

4 comments
AuthorsOpenReview ↗

Response to Reviewer cit1 Part I

Thank you so much for your manual review and for the time and effort. We appreciate your recognition that World Rollout offers a practical and scalable way to build code-agent mid-training data from open GitHub issues, and that our evaluation examines the framework from complementary perspectives.

We will address your concerns and questions one by one:

W1.The claimed four to sixfold token throughput improvement is based on resource modeling rather than end-to-end measurements. The abstract and main text do not clearly distinguish theoretical estimates from empirical results, leaving the actual acceleration gain unverified.

The throughput gain was first observed in our end-to-end production data collection. Using the 80-node GB300 setup described in Appendix D, with the same budget of 320 GPUs for each method, we measured training tokens collected per day. Simulated-runtime collection produced more than 4× as many training tokens per day as execution-based collection.

Then we formalizes this observation with a throughput model in Appendix D.2. The model uses measured decode times for policy and world-model requests (13.4 s and 88.9 s per request), the average number of calls per trajectory (35.2 and 5.65, respectively), and historical wall-clock durations for execution-based rollouts (717.9 s mean, 519.3 s median). Under the same GPU budget, it estimates a 4.1× gain using the median execution time and a 5.7× gain using the mean. We will clarify this further in the paper: the speedup was indeed the actual acceleration, and the theoretical framework in Appendix D.2 provides a formal account of that empirical result.

W2.Simulated runtime feedback may cause distribution shift. Only final patches are validated, leaving intermediate observations unverified by real execution. Moreover, the experiments cannot distinguish gains from increased task diversity from those due to simulated trajectories.

Thank you for raising these points. We address them in turn.

Intermediate observation quality. We chose Qwen-AgentWorld-35B-A3B[1] to balance fidelity and collection speed. On the SWE subset of AgentWorldBench, it scores 65.6/100 on a rubric that compares intermediate simulated responses with real observations, providing direct evidence of per-turn fidelity. If intermediate fidelity is the priority, a frontier model such as Claude Opus is a plausible alternative, at the cost of slower collection. While at the end-to-end level, patches from simulated and execution-based rollouts passed real tests at nearly identical rates (22.29% vs. 22.39%). This provides indirect evidence that the simulated intermediate responses were informative enough to guide the agent toward correct fixes.

Distribution shift. We agree that simulated runtime feedback can shift the distribution of observations the agent sees during training. Empirically, however, adding simulated trajectories to execution-based data made the agent more robust on the tasks we evaluated, as shown by the behavioral analysis in Section 4.3. Training on both sources exposes the agent to a broader range of runtime responses: simulated feedback complements the observations available from execution-based trajectories.

Task diversity vs. simulated trajectories. We would like to clarify that broader task coverage is precisely the benefit World Rollout is designed to enable, and the world model is just the mechanism that makes these tasks usable. Our fixed-budget SFT replacement experiment (Section 3.2, Figure 2) directly compares simulated and execution-based trajectories without expanding the task pool. Replacing all execution-based traces with simulated ones yields comparable performance (38.8% vs. 39.3% average pass@1 and identical average pass@k). We therefore do not attribute the gains in our experiments to simulated trajectories being intrinsically better training data. Rather, simulation lets us collect trajectories for a much more diverse set of tasks that execution-based collection cannot readily cover. We interpret the gains primarily as a benefit of that added task diversity.

[1] Zuo et al. (2026). Qwen-AgentWorld: Language World Models for General Agents.

The authors have responded to the first two weaknesses. However, they did not address the five follow-up questions I raised. These important questions remain unresolved. I therefore revise my score from 3 to 2.

AuthorsOpenReview ↗

Response to Reviewer cit1 Part II

Sorry for the delayed part 2 &3 as we are revising the responses to the five questions:

Q1. The standalone benefit of World Rollout-generated interaction trajectories remains unclear. Could the authors compare SFT performance using only issue text, real execution trajectories, and simulated execution trajectories under a fixed data budget?

Thank you for this suggestion. The comparison between real execution and simulated trajectories under a fixed data budget is exactly the Original vs. WR100 setting in §3.2 (Figure 2). Both use the same base model, the same SFT budget. Original uses 100% execution-based SWE traces, and WR100 replaces all of them with World Rollout traces (pass@1 / pass@k, %):

SettingSWE-bench VerifiedSWE-bench MultilingualSWE-bench ProTerminal-Bench 2.1Avg.
Original (real execution)56.2 / 68.850.2 / 67.330.4 / 41.220.5 / 41.639.3 / 54.7
WR100 (simulated)53.4 / 66.450.1 / 68.729.2 / 41.022.5 / 42.738.8 / 54.7

Trained on simulated trajectories alone, the model reaches 38.8% average pass@1 (similar to execution-based's 39.3%) and the same average pass@k (54.7%).

Regarding issue text only: our SFT loss is applied only to assistant turns (reasoning and tool calls), while system/user messages and tool observations are masked (Appendix C.1.1). The issue text is the user message of every trajectory, so it is already present as context in both settings above. A corpus of issue text alone contains no actions or observations, so it provides no supervision for the multi-turn tool use these benchmarks evaluate; a model trained on it would not learn to solve the agentic problems.

Q2. The advantages of World Rollout over SWE-World remain unclear. Could the authors provide comparative experiments under matched task settings and training configurations, or a more detailed qualitative analysis?

Thank you for the question. World Rollout and SWE-World [2] target different training stages and data regimes, which makes a matched-setting comparison difficult to define:

  • SWE-World targets post-training on a curated set of 16.6K high-quality merged issue–PR instances (≈5.7K filtered SFT trajectories). Its transition model conditions on the ground-truth patch, and its reward model on F2P/P2P tests.
  • World Rollout targets mid-training scale: 3M+ unique tasks and 100B+ training tokens, drawn largely from open issues and development-branch states (Table 7) that have no reference fix or tests.

Matching on our task pool is not possible, since SWE-World's simulators require a gold patch and F2P/P2P tests that these tasks lack. Matching on SWE-World's curated tasks would compare two datasets that differ by orders of magnitude in volume and in curation level, and would leave out the setting we target. We therefore compare simulated traces against the strongest available reference, real execution on the same task source (§3.2 and §4.1).

For a qualitative comparison, Table 1, Table 6 (Appendix B) and Table 8 (Appendix C.2) contrast the two systems. The main differences are:

  1. Oracle-free simulation. Our world model sees only the public task, the current real diff and the interaction history; gold patches and tests are excluded by construction. This is what admits unresolved issues, open PRs, and tasks without PR test changes.
  2. Multilingual routing. A three-way capability router covers runtimes beyond Python (Go, C/C++, Java, CUDA, Julia, Elixir, OCaml), routes compound commands atomically, and explicitly refuses unsupported actions. SWE-World's released dispatch primarily detects Python and pytest, and other commands run locally.
  3. State ownership. Our simulator is state-inert: only controller-owned edits can change files, every observation records its backend and routing decision, and a hash-chained turn journal supports auditable recovery.

World Rollout builds directly on SWE-World's separation of real repository state and simulated runtime feedback (§1, §2.1) with enhanced functionalities on a much larger scale and real world setting. We will move a condensed version of this comparison into the main text.

[2] Sun et al. (2026). SWE-World: Building Software Engineering Agents in Docker-Free Environments.

AuthorsOpenReview ↗

Response to Reviewer cit1 Part III

Q3. Please analyze the modest performance gains on Terminal-Bench-2.0 and clarify whether they stem primarily from limited simulation fidelity or the inherent characteristics of the benchmark tasks.

Thank you for this question. We believe the modest Terminal-Bench gains mainly reflect the characteristics of the benchmark relative to our training data rather than limited simulation fidelity, for following reasons:

  1. Task OOD Both the World Rollout traces and the execution-based traces they are compared with are repository-level GitHub issue-resolution traces. Terminal-Bench 2.1 covers a broad range of command-line tasks (e.g., system administration, environment setup, data processing, security) and is evaluated with a different agent, Terminus-2 through Harbor (Appendix C.1.2). Gains on it measure out-of-domain transfer, which we expect to be smaller than in-domain gains on the SWE benchmarks.
  2. Terminal-Bench does not degrade as simulated data increases. If simulation fidelity were the bottleneck, replacing real traces with simulated ones should hurt Terminal-Bench. Instead, pass@1 goes 20.5 → 20.4 → 21.6 → 22.5 from Original to WR25/50/100 (Figure 2), so real and simulated SWE traces transfer to it to a similar degree. While for mid-training, pass@k improves more clearly (41.6% → 46.1%, Table 3).

Our interpretation is that the modest gains on Terminal-Bench 2.1 mainly reflect limited coverage of terminal tasks in our training data, rather than the fidelity of individual simulated responses.We can broaden the coverage of World Rollout by adding terminal-native task sources. Qwen-AgentWorld already models terminal and OS environments, so this extension could expose the agent to a wider range of interactions relevant to Terminal-Bench 2.1. We will discuss this direction in the revision.

Q4. Performance degrades when all trajectories are replaced with simulated ones in the WR100 setting. Please discuss the practical deployment risks when fully relying on simulated trajectories for training.

Thank you for this question. The average drop is moderate: WR100 reaches 38.8% pass@1 vs. 39.3% for Original, with identical average pass@k (54.7%); the largest drop is on SWE-bench Verified (56.2 → 53.4), while Terminal-Bench improves (20.5 → 22.5).

We agree that relying entirely on simulated trajectories carries a real risk: distribution shift between training feedback and deployment feedback. A world model leaves systematic patterns in the feedback it generates; Table 4b shows, for example, fewer failed calls and shorter long-tail outputs than real execution. A model trained only on simulated traces sees only this feedback distribution, while evaluation and deployment take place in real execution environments. Behaviors that depend on real feedback, such as recovering from verbose or unexpected errors, may then be under-trained, and occasional hallucinated feedback is never corrected by a real signal.

For this reason, World Rollout is designed to complement execution-based collection, for two purposes: (1) scaling data volume and task coverage on top of existing execution-based traces, and (2) providing trajectories for tasks that execution-based collection cannot reach, such as open issues, development branches, and tasks with broken images or non-reproducing gold solutions, where simulated traces are the only interaction data available. Our main results follow this usage: in mid-training, World Rollout traces are added alongside execution-based traces (Table 3, 44.4% → 46.4% average pass@1), and in the additional 200B-token setting reported in our response to Reviewer epPN, real-data exposure is kept unchanged and only simulated traces are added, reaching 46.8%. We use WR100 as a stress test of simulated-trace utility rather than as a recommended configuration, and will add this guidance and the associated risks to the Limitations section.

Q5. Please clarify which Terminal-Bench version (2.0 or 2.1) is used in the experiments.

Thank you for catching this, and sorry for the confusion. All experiments use Terminal-Bench 2.1, evaluated with the Terminus-2 agent through Harbor (Appendix C.1.2). The mentions of "Terminal-Bench 2.0" in the abstract and Introduction are typos. We will use "Terminal-Bench 2.1" consistently throughout the paper, including Table 3, and state the version explicitly in the experimental setup (§3.1).

Thank you again for your constructive feedback. We will incorporate these clarifications and fixes into the final version and acknowledge your helpful feedback. If the additional evidence addresses your concerns, we would appreciate you can keep your positive score.

Reviewer_epPNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes World Rollout, a framework for generating large-scale agentic training trajectories at low cost. It uses a world model to simulate execution-dependent feedback from tests, builds, and program execution, while keeping repository reads, searches, and edits grounded in real files. Compared with SWE-World, the main contribution is removing the need for reference fixes (gold patches), pre-existing tests, and task-specific runnable environments. This enables the use of unresolved GitHub issues and in-progress PRs to expand task diversity. Experiments show that replacing execution-based traces with simulated traces maintains comparable performance. Adding traces from new tasks improves average pass@1 by 4.6 percentage points over repeating the original data, while expanded task coverage in mid-training improves average pass@1 by 2.0 percentage points.

Critical strengths
  1. Important research direction. The paper addresses an important problem in scaling agentic training data.
  2. Strong scalability in data collection. By removing the requirements for runnable environments and reference solutions, the proposed approach can leverage a large number of real-world GitHub tasks that were previously difficult to use, substantially lowering the barriers to data collection.
  3. Promising experimental results. The paper evaluates the training utility of traces from additional tasks across multiple agent benchmarks with real execution, providing evidence for the importance of task diversity in mid-training.
Critical weaknesses
  1. Missing key ablation studies. The paper replaces 25%, 50%, and 100% of execution-based traces with simulated traces and reports comparable performance to the original setting. However, it does not compare 75% real-only data against 75% real + 25% simulated data. Therefore, it is unclear whether simulated traces provide additional training benefits or simply replace redundant execution-based data.
  2. The contributions of task diversity and the world model are not sufficiently disentangled. The paper mainly compares repeating existing trajectories with adding simulated traces from new tasks. However, the performance improvements may primarily come from exposure to a broader range of GitHub tasks rather than the simulated feedback itself. In particular, the paper lacks a control experiment using the same additional tasks without the world model, making it difficult to determine whether world-model simulation is necessary for the observed improvements.
  3. Simulated feedback may introduce hallucinations and behavioral biases. The world model may incorrectly predict successful test outcomes, potentially misleading the agent's subsequent decisions. The current filtering mechanisms do not guarantee the semantic correctness of simulated feedback. Moreover, the reported failure rate is 10.4% for real terminal observations but only 3.5% for simulated observations. Although these observations are not directly paired, the difference suggests a distribution shift between simulated and real execution feedback.
Questions for authors
  1. How consistent is world-model-generated feedback with real execution results? In particular, what is the false-positive rate where a test fails in the real environment but is predicted to pass by the world model? Could such errors cause the agent to learn incorrect debugging strategies or premature termination behaviors?
  2. How does the framework handle execution side effects? For operations that modify files or produce other persistent side effects, how is consistency maintained between simulated feedback and subsequent real repository states? If such consistency cannot be guaranteed, have the authors evaluated how these inconsistencies affect long-horizon trajectory generation and downstream agent performance?
  3. Are the benefits of task diversity independent of task quality and data distribution? The additional tasks come from different types of GitHub issues and PRs, which may differ from the original data in task difficulty, programming language, repository size, and code quality. Have the authors controlled for these factors to demonstrate that the improvements are indeed driven by broader unique-task coverage rather than shifts in the training-data distribution?
  4. Is there a better balance between execution-based and simulated traces? The paper evaluates replacement ratios of 25%, 50%, and 100%, but does not thoroughly analyze the trade-offs among performance, collection cost, and training stability. Is there an optimal mixture that substantially reduces data collection costs while minimizing the potential negative effects of inaccurate simulated feedback?
3 comments
AuthorsOpenReview ↗

Response to Reviewer epPN Part I

Thank you so much for your review and for the time and effort. We appreciate your recognition of the importance of scaling agentic training data, the scalability of World Rollout, and the promising results from broader task coverage.

We will address your concerns and questions one by one:

W1. Missing key ablation studies.

Thank you for raising this concern. Replacing redundant execution-based data is itself one way simulated traces provide additional training benefits. As discussed in our motivation, execution-based collection is constrained by the availability of runnable environments and verifiable tasks, leading to repeated exposure as training scales. Replacing some of this repetition with simulated traces from new tasks increases diversity. Simulation also substantially expands the available data volume.

Our experiments follow this logic: The SFT replacement study evaluates whether simulated traces offer training utility comparable to execution-based traces. While our mid-training study further supports the training benefits through broader task coverage, improving average pass@1 from 44.4% to 46.4% under the same token budget and identical downstream SFT. We have now added a 200B-token setting that keeps the original real-data exposure unchanged and adds only simulated traces as you suggested. This setting yields further gains, reaching 46.8% average pass@1 on 4 tasks. We will include these latest experimental results in the final version of our paper.

W2. The contributions of task diversity and the world model are not sufficiently disentangled.

Thank you for raising this point. We believe there is some misunderstanding about our contribution, and we would like to clarify it. Broader exposure to GitHub tasks is precisely the benefit we aim to enable and evaluate. Our contribution is to develop a practical, scalable pipeline that turns previously underused tasks, especially open GitHub issues and development-branch tasks, into usable training trajectories, provide empirical evidence for their training benefits, and release the complete dataset and pipeline to the community.

Specifically, these tasks may lack reference fixes, verification tests, or runnable environments, making them difficult to use through existing execution-based collection pipelines. We adopt a pretrained world model to supply runtime responses when real execution is unavailable, while keeping repository reads, searches, and edits grounded in real files. The world model serves as a collection method that enables access to these additional tasks. We do not claim that a particular world model is uniquely necessary for the training benefits; it could, in principle, be replaced by another mechanism capable of supplying suitable runtime responses. To the best of our knowledge, existing methods for generating sufficiently faithful simulated runtime responses in this setting rely on world models. We adopt Qwen-AgentWorld because it provides a strong pretrained starting point with demonstrated response fidelity. Without the world model supplying these simulated responses, we cannot incorporate tasks requiring unavailable runtime feedback into training data with the intended tool-call–response interactions.

W3. Simulated feedback may introduce hallucinations and behavioral biases.

Thank you for raising this valid concern. Expanding task coverage and data volume through simulation involves a trade-off with feedback fidelity, including hallucinations and behavioral biases. As part of this trade-off, we chose a relatively small 35B world model for collection efficiency. Replacing it with a stronger frontier model could further reduce hallucinations. Similar to LLM-as-a-judge, we expect simulated responses to become more reliable as model capabilities improve, reducing these issues over time. We will explicitly discuss hallucinations and feedback-distribution shifts in the paper’s Limitations section and leave this improve as an important future work. We appreciate your suggestion, which helps make our paper more complete.

AuthorsOpenReview ↗

Response to Reviewer epPN Part II

Q1.How consistent is world-model-generated feedback with real execution results? In particular, what is the false-positive rate where a test fails in the real environment but is predicted to pass by the world model? Could such errors cause the agent to learn incorrect debugging strategies or premature termination behaviors?

Thank you for this question. We do investigate the correctness of simulated trajectories and their consistency with execution-based counterparts in Section 4.1 and Appendix D.1. We audit 200 samples from SWE-rebench V2 by applying the generated patches of simulated trajectories in real task environments and running the official tests. Simulated and execution-based trajectories achieve similar pass rates (22.3% and 22.4%) and P2P regression rates (2.9% and 3.7%). These results provide evidence of comparable trajectory correctness under world-model-generated feedback and real execution feedback.

We assess their overall consistency through aggregate correctness rates, rather than one-to-one matching of independently sampled rollouts to calculate false positives and false negatives. For the same task, stochastic sampling by the policy model, influenced by decoding temperature, can lead to different action sequences, edits, and final patches. Even repeated rollouts in the same execution-based environment can therefore yield very different outcomes. Comparing success/failure disagreements between independently generated trajectories conflates the effects of feedback with variability introduced by the policy model and its sampling process. Such disagreements cannot directly isolate feedback errors, making them unsuitable as direct measurements of feedback false positives or false negatives.

Regarding behavioral effects, Section 4.3 reports improvements in the measured behaviors under broader task coverage. Comparing Repeat with Broad+WR, editor-failure incidence among paired successful trajectories decreases from 31.7% to 22.9%, while the share of tasks solved in all three attempts increases from 31.8% to 45.2%. These observations suggest more reliable editing and more consistent problem solving. We will discuss the remaining risk of misleading feedback in our Limitations section, as noted in W3.

Q2. How does the framework handle execution side effects? For operations that modify files or produce other persistent side effects, how is consistency maintained between simulated feedback and subsequent real repository states? If such consistency cannot be guaranteed, have the authors evaluated how these inconsistencies affect long-horizon trajectory generation and downstream agent performance?

Thank you for this question. We would like to clarify that each rollout maintains a private and light filesystem overlay that provides an authoritative repository state. Controller-owned file creations, modifications, and deletions are recorded in this overlay, including changes to original and intermediate files. Subsequent reads and searches reflect this updated state. The world model is also conditioned on the latest grounded repository context, including the current diff and changed paths. This maintains consistent file-state tracking throughout long-horizon interaction.

While for filesystem changes caused by execution, simulated runtime calls cannot create files, update generated artifacts, or modify the overlay. We intentionally maintain this boundary because simulated execution can contain potential errors. Runtime-driven file modifications were uncommon in our tests with DeepSeek and GLM policy models. We additionally conducted an audit of 2,000 sampled generated trajectories assisted by Opus 5.5, and no cases involving such execution-side-effect behavior were identified. These observations suggest a low incidence of this issue in our evaluated setting. We will include a detailed analysis of this issue and the audit results in the Discussion section. Thank you for raising this concern and helping us make the paper more complete.

AuthorsOpenReview ↗

Response to Reviewer epPN Part III

Q3. Are the benefits of task diversity independent of task quality and data distribution? The additional tasks come from different types of GitHub issues and PRs, which may differ from the original data in task difficulty, programming language, repository size, and code quality. Have the authors controlled for these factors to demonstrate that the improvements are indeed driven by broader unique-task coverage rather than shifts in the training-data distribution?

Thank you for this question. We would like to clarify two points. First, our SFT ablations primarily construct simulated traces from the same SWE-rebench V2 source used for execution-based collection, especially for tasks where execution-based trajectory collection fails. This limits cross-source differences when evaluating simulated-trace utility. The replacement study tests whether simulated traces provide comparable training utility, while the additional-query versus repetition comparison tests the benefits of expanding task coverage within this source setting.

Second, broader exposure to the training-data distribution is an intended benefit of our expanded task pool. Open issues and development-branch tasks extend training beyond the curated, execution-verifiable tasks commonly used in post-training, allowing the model to encounter previously unavailable problems and repair patterns. We therefore interpret the gains in mid-training and larger-scale experiments as the combined benefits of this broader coverage, including exposing the model to a broader distribution of training tasks. The behavioral improvements in Section 4.3 are also consistent with learning more reusable problem-solving patterns. Task diversity and distributional broadening are closely connected in this setting. Regarding task quality, the additional open-issue and development-branch tasks are generally less curated and less readily verifiable than the environment-backed tasks carefully selected for post-training. This makes a simple increase in task quality an unlikely explanation for the observed gains.

Q4.Is there a better balance between execution-based and simulated traces? The paper evaluates replacement ratios of 25%, 50%, and 100%, but does not thoroughly analyze the trade-offs among performance, collection cost, and training stability. Is there an optimal mixture that substantially reduces data collection costs while minimizing the potential negative effects of inaccurate simulated feedback?

Thank you for this suggestion. Our SFT replacement study primarily evaluates the training utility of simulated trajectories and the feasibility of our collection pipeline. The tested replacement ratios provide initial evidence that simulated traces can offer utility comparable to execution-based traces across different mixtures.

A systematic search for an optimal mixture would require a more detailed analysis of performance, collection cost, and training stability. We consider this an important direction for future work, building on the usefulness established by our current experiments. We will explicitly discuss this opportunity in the Discussion section.

Thank you again for your thoughtful feedback and suggestions. In response, we ran additional experiments, extended our analyses, and clarified our contributions. We hope these additions address your concerns and make the paper’s contributions clearer. We will incorporate the new results and clarifications from this discussion into the final version and acknowledge your helpful feedback. If the additional evidence addresses your concerns, we would appreciate your considering a higher rating to positive.

Reviewer_D5HEOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces World Rollout, a three-stage data collection framework for scaling coding-agent training data without requiring task-specific runnable environments or reference fixes. The framework consists of: (1) Task sourcing, collecting tasks from merged issue–PR pairs, open issues, and development branches; (2) Runtime-decoupled interaction, combining deterministic repository operations with simulated runtime feedback from Qwen-AgentWorld; and (3) Post-filtering, applying trajectory integrity and world-model-specific checks for filtering (Fig. 1). The authors demonstrate that the final trajectories provide comparable SFT performance to execution-based traces (Fig. 2) on four benchmarks and that expanded task coverage improves downstream performance after mid-training (Table 3).

Critical strengths
  • Scalable data collection. World Rollout expands training beyond solved, executable tasks while preserving real repository state for file ops (Fig. 1). This addresses a practical limitation of existing coding-agent training pipelines by expanding coverage / diversity.

  • Good improvement in results. Simulated traces achieve comparable overall SFT performance to execution-based traces (Fig. 2). Expanded task coverage improves average mid-training pass@1 from 44.4% to 46.4% after identical downstream SFT (Table 3).

Critical weaknesses

Limited validation of simulated trajectory correctness. To be honest, I don’t work with coding agents, so my knowledge is limited in this area. One thing I wanted to ask the authors is whether there is a way to estimate the correctness of simulated trajectories. I understand that the overall results have improved, but still would like to see an estimate of the trajectory correctness to be convinced since this work is purely empirical. Using a small amount of data to measure correctness would suffice.

Questions for authors

No additional questions.

I do not work with coding agents or midtraining LLMs. So I prefer to provide a clean review signal rather than mentioning any generic weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer D5HE

Thank you so much for your thoughtful manual review and for the time and effort. We appreciate your recognition that World Rollout addresses a practical limitation of existing coding-agent training pipelines by expanding coverage to unresolved issues and highlighting the improvement.

We will address your concern(s) one by one:

One thing I wanted to ask the authors is whether there is a way to estimate the correctness of simulated trajectories.

Thank you for your thoughtful question. We do investigate the correctness of simulated trajectories in Section 4.1, with more details in Appendix D.1. We test simulated trajectory correctness on 200 samples from SWE-rebench V2, as this subset provides valid envs, ground-truth solutions, and fail2pass tests for execution-based verification. For general unresolved issues (such as open issues), verification is harder because ground-truth solutions are unavailable.

We apply the patches produced by these trajectories to the real task environments and run the tests. Simulated and execution-based trajectories achieved very similar pass rates in their respective audit samples (22.3% and 22.4%), as well as similar P2P regression rates (2.9% and 3.7%). We take this as evidence that simulated trajectories have a similar correctness rate compared with their execution-based counterparts.

For per-turn response correctness, Qwen-AgentWorld model, the simulator underlying our framework, evaluates simulated responses against real environment outputs through AgentWorldBench [1]. The 35B-A3B model we use achieves 65.6% on SWE response fidelity, assessed across format, factuality, consistency, realism, and quality. It also achieves 60.1% accuracy in reproducing appropriate errors in SWE cases with known failures. These measurements provide a complementary reference for the correctness of the simulated feedback used in our trajectories.

We hope this addresses your concern. If you are interested, our responses to the other reviewers provide additional discussion and details that may help you better understand our work. We hope these clarifications support your maintaining a positive score for our paper.

[1] Zuo et al. (2026). Qwen-AgentWorld: Language World Models for General Agents.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.