acceptodds
Under review as a conference paper at ICLR 2027

SCoPR: Scaling State-Evolution Horizons for Long-Horizon Agents

Abstract

Coding LLMs are evolving from single-turn code generators into software-engineering agents that operate over extended horizons in real repositories. As agent-authored PR trajectories such as OpenClaw grow rapidly, a central question emerges: can software-engineering trajectories generated by coding agents, after rejection sampling, be used to train stronger long-horizon agents, and can scaling trajectory horizon yield measurable capability gains? We introduce SCoPR to generate candidate rollouts with a chain-of-PR pipeline and select high-quality chains by measuring how each intermediate state in the functional evolution aligns with the golden patch. To enrich the state-evolution process, we introduce two auxiliary pipeline improvements: world-model execution feedback, which mitigates rollout interruption caused by unavailable or unstable code execution, and diff-preserving handoff, which prevents state information loss during workspace transitions. We find that rejection-sampled agent-authored trajectories provide effective supervision for state-evolution-oriented training data. Moreover, under controlled data budgets, increasing the effective horizon yields additional capability gains. These results suggest that agent-authored code data, when combined with human-authored software data, can simultaneously expand the available training resource and improve downstream performance. More importantly, the trajectory horizon extends the scaling axis of long-horizon agents beyond raw token quantity toward cumulative state evolution and persistent goal pursuit, providing a more effective path for unlocking the intrinsic potential of long-horizon agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
coding agent, long horizon agent, llm training

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yRBXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper builds training trajectories from chains of related pull requests. It preserves earlier code changes across stages, uses simulated execution feedback when real execution is difficult, and filters patches against reference changes. It studies whether longer chains and OpenClaw-derived data improve downstream agent performance.

Critical strengths
  1. Preserving changes across PR stages is a reasonable way to train agents for tasks that depend on earlier edits.

  2. The evaluation covers both coding and tool-use tasks, giving a broader view beyond a single benchmark.

Critical weaknesses
  1. The comparisons mainly cover Qwen3-30B-A3B and Step-3.5-Flash. Frontier reference baselines such as Claude and GLM are missing from the main comparison, and needs to justify the choice of OpenClaw as the sole agent-authored source.

  2. Only 959 rollouts are retained, and the horizon study does not control tokens or compute. This is limited evidence for scaling and needs comparison with TOUCAN, APIGen-MT, and Simia.

  3. The novelty is unclear. The paper should explain how its rejection sampling differs from the existing strategy and framework.

Questions for authors

See Weaknesses.

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

SCoPR is a training-data construction pipeline for coding agents that operate across dependent pull-request stages. Building on the chain-of-PR formulation, it preserves transition diffs during workspace handoffs, uses GLM-4.7 to provide provisional execution feedback when real execution is unavailable or unstable, and filters candidate patches against reference patches at intermediate stages. The pipeline processes 148,655 PRs from 17 repositories, constructs 5,193 queries, and retains 959 rollout samples. The paper evaluates Qwen3-30B-A3B and Step-3.5-Flash across six benchmarks, studies the contribution of an OpenClaw-derived source collection, and varies PR-stage horizon while fixing the training-sample count at 722. The full mixture obtains the highest reported averages in the presented comparisons, while the fixed-sample horizon study shows an association between longer chains and higher average performance.

Critical strengths
  1. A useful data-design problem grounded in software evolution. Training agents on dependent repository changes is well motivated: later development stages may need to extend, diagnose, or repair code produced earlier. Preserving transition information and generating trajectories across such stages is a practically relevant direction for coding-agent training.

  2. Encouraging downstream results with some appropriately qualified interpretations. On Qwen, SCoPR improves SWE-bench Verified from 24.2% to 27.2% and Toolathlon from 3.7% to 5.6%, and it exceeds daVinci-Agency on six displayed metrics while tying the remaining one. The full source mixture also has a higher reported average than the configuration without OpenClaw for both evaluated model families. The paper usefully acknowledges that OpenClaw-only training does not improve the average and that simulated feedback is not evidence of executable correctness.

Critical weaknesses
  1. The experiments do not isolate the proposed horizon effect or clearly establish the contribution beyond the closest prior work. Table 5 fixes the number of training samples at 722, but not token exposure, compute, or repository/task composition. Longer chains can therefore provide more supervised content without demonstrating a benefit specifically from cross-stage dependence. Section 5.5 explicitly acknowledges this limitation, whereas the conclusion states that the gains hold “under matched token budgets” (lines 484–485). This contradiction should be resolved. Moreover, daVinci-Agency already includes PR chains, candidate-patch propagation, reference-guided evaluation, and a chain-extension study, while simulated execution feedback has been studied in SWE-World. SCoPR may offer a useful integration, but the lack of separate ablations of diff preservation and simulated feedback leaves its additional contribution insufficiently established.

  2. Missing protocol details and limited independent validation make the comparative results difficult to assess. The paper does not sufficiently specify training hyperparameters, context truncation and loss masking, or sample/token budgets for the data baselines and source-mixture variants. Consequently, the effects of data construction cannot be separated from training exposure, and the OpenClaw comparison remains confounded by source content and possibly mixture size. Table 2 reports judge acceptance rather than independently verified correctness; Table 3 shows a 9.1-percentage-point gap between simulated and real feedback but does not calibrate the filtering procedure. Appendix E also allows the evaluator to give gold-patch-derived repair instructions, making it important to clarify their use in retained training trajectories and their exclusion from benchmark-time evaluation. Finally, the modest gains over the strongest data baseline are reported without uncertainty estimates. These are limitations of the available evidence, not proof that the method fails or that test leakage occurred.

Questions for authors
  1. Horizon controls: What are the actual training-token counts, update counts, length distributions, and truncation rates for each row of Table 5? Are the rows derived from the same underlying chains? Please reconcile Section 5.5 with the conclusion's statement about matched token budgets. If a matched-token comparison or a comparison that trains on the same PR stages independently already exists, please report it.

  2. Fairness and attribution: Please provide the sample counts, token budgets, and training settings for the baselines and source configurations, and explain the relationship between the 959 retained rollouts and the 722-sample horizon datasets. Are component ablations available that separately remove transition-diff preservation or simulated feedback while holding the source queries and generation budget fixed?

  3. Acceptance and evaluation protocol: What is the denominator and success criterion in Table 2, and how do attempts and retries interact? Is Table 3 scored by real execution, and is gold-patch-guided feedback excluded from all benchmark-time attempts? How are evaluator comments used in retained training samples, and what independent execution checks or judge-calibration results are available?

  4. Reliability and data separation: Could existing task-level results provide paired confidence intervals for the principal comparisons, with the number of training seeds and evaluation repeats stated explicitly? Please also describe the issue/PR/patch/commit decontamination protocol between source histories and evaluation tasks.

Reviewer_DdpDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SCoPR is a data pipeline that builds long-horizon software-engineering SFT trajectories from chains of dependent pull requests. Building on daVinci-Agency’s PR-chain formulation and reference-guided selection, SCoPR introduces diff-preserving handoffs and world-model execution feedback. The first is diff-preserving handoff, which carries each stage’s candidate diff, repair trace and base hash into the next stage. The second is world-model execution feedback, where GLM-4.7 predicts execution results when real execution is unavailable. The third is reference-guided rejection sampling, where an LLM judge scores each stage’s patch against the gold PR patch. From 148,655 PRs in 17 repositories (5,193 queries), the pipeline produces 959 rollout samples. The authors fine-tune Qwen3-30B-A3B-Thinking and Step-3.5-Flash and observed moderate improvement over baseline.

Critical strengths

The fixed-sample horizon study shows a monotone trend over five chain lengths, with SWE-bench Verified rising from 23.6% to 28.6%, which gives an actionable data-design direction under a fixed sample budget.

Whether fast-growing agent-authored code can serve as training supervision is a timely question, studied here on two model families. The negative standalone result for OpenClaw-only training is itself informative.

Critical weaknesses

Limited and noisy Improvements. Over daVinci-Agency, the average improves by only 0.010: about half comes from DS-1000 (+3.4), SWE-bench Verified gains just 1.0 point (5 of 500 tasks), and no run-to-run variability or uncertainty estimates are reported.

Relatively incremental contribution over previous work. Lack of ablation studies on SCoPR’s additions over daVinci-Agency

Horizon confounded with token count. Fixing the sample count does not fix the training signal, because a 5-PR sample contains about five PR-level tasks and far more tokens than a 1-PR sample. The paper acknowledges this (ll. 466–470), yet the Conclusion claims gains “under matched token budgets” (ll. 484–485).

Unequal data versus the main baseline. The four overlapping metrics in the daVinci-Agency row exactly match those reported by Jiang et al. (2026), whose study used a 239-sample dataset. The corresponding training-data budget in SCoPR is not specified. Even so, SCoPR without OpenClaw only ties daVinci-Agency (0.263 vs. 0.264).

Questions for authors

Are the judge’s line-level fix comments, which are derived from the reference patch (Appendix E), fed back within retained trajectories, and can the repair trace r_t or the stage-t base commit carry reference-patch information into later stages? What fraction of retained samples succeeded only after such feedback?

What decontamination was done against SWE-bench Verified, given that sympy (8,760 PRs, Table 6) is one of its source repositories, and against DS-1000, given the numpy and scipy sources?

How are 722 samples per horizon level obtained from a 959-sample collection, and what are the sample, token and step counts for each row of Tables 1, 4 and 5? Which model and acceptance threshold does the evaluator use?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.