acceptodds
Under review as a conference paper at ICLR 2027

SWE-Loop: Trace-Guided Self-Improvement for Coding Agents

Abstract

Coding agents are commonly improved by collecting additional trajectories or tuning a fixed task distribution, leaving open whether their own execution evidence can dynamically steer what they learn next. We present SWE-Loop, a trace-guided closed-loop framework for self-improving coding agents. At each iteration, SWE-Loop generates repository-level software-engineering tasks, obtains rollouts from a teacher model and the current student model, diagnoses semantically attributable failures, and updates the next task distribution to target recurring capability deficits. Supported by a multi-agent research team that validates tasks, curates trajectories, and controls the experimental process, SWE-Loop curates execution-verified trajectories as training data while analyzing teacher and student traces to identify shared blind spots and student-specific regressions. The framework disentangles semantic/model failures from system-level failures, including setup, provider, timeout, verifier, and infrastructure errors, and evaluates each iteration on a mainstream software-engineering benchmark and group-disjoint evaluation task families using resolution rate, patch coverage, regression rate, and compute cost. Across a three-iteration evolution loop, we demonstrate that trace-guided task synthesis yields generalizable improvements beyond non-adaptive baselines that train on self-generated tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
coding agents, trace-guided learning, self-improvement, task synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_tR8zOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper is based on the observation that failure modes in self-improvement tasks can be used for future task selection and proposes a teacher-student distillation method

Critical strengths

The observation that the authors make is that current self-improvement methods in coding tasks do not use failure types to curate next tasks for self-improvement, and it is a potential research gap, especially considering that evaluations are the costly part of the coding agents.

Critical weaknesses

I think the paper misrepresents the contribution as a self-improvement paper, whereas the entire method requires a stronger teacher model to supervise and propose a concurrent trajectory. Hence, the claim of the paper about improving self-improvement cannot be justified in this way. For example, the performance is by definition, limited of the fixed teacher model.

The paper is really hard to follow in depth. For example, the main figure is not helpful in understanding the main loop of the paper and feels like a set of buzzwords. It would be helpful to describe the process under the figure as well and justify the words in the boxes of the plot.

The comparisons in the main table are unfair because this is the only method that has access to a bigger model, which it can distill. AbsoluteZero and other baseline methods do not have bigger teacher models.

Questions for authors

I think the method itself could be useful for some distillation cases. I'd like to suggest that the authors reframe the paper from a self-improvement paper (which it is not) to a distillation paper and propose a situation in which the authors' method could be used.

I would also be interested in the performance of the model if the teacher model uses the same foundational model as the student model.

What is the relation of the method the authors propose to the on-policy distillation methods[1]

[1] - On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes. Agarwal et al.2024

Reviewer_TsF4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces SWE-Loop, a closed-loop training framework for coding agents. It generates repository-level tasks, collects paired teacher and student trajectories, attributes failures to specific capabilities, and uses those diagnoses to guide the next round of synthetic task generation and supervised fine-tuning.

The authors evaluate three iterations of training a Qwen3.5-9B student and report improvements across SWE-bench and Terminal-Bench benchmarks.

Critical strengths
  • The overall problem has become quite trending and relevant
  • The authors pay attention to failure attribution, distinguishing actual model failures from infrastructure, verifier, and task-construction errors.
  • The methodology is relatively clear, and the discussion of contamination, task validity, paired evaluation, and retention demonstrates awareness of common pitfalls in coding-agent research.
  • The authors evaluate on real and difficult benchmarks
Critical weaknesses
  • Table 2 presents what initially appears to be a conventional component ablation. However, the caption explicitly admits that these are historical score anchors and draft annotations.
  • I do not understand why such a table is presented as experimental evidence in a conference submission.
  • The authors introduce a four-pillar failure taxonomy but provide virtually no quantitative evidence about the quality of this attribution process.
  • Later in Section 5.3, the authors describe what the analysis intends to report rather than presenting its findings.
  • Novelty across Socratic-SWE, SWE-Gym and a range of recent work is very incremental, and rather feels forced, than genuinely motivated
  • Overall, there are many instances where the writing itself promises/planned for a lot of things (L94), but the paper fails to present matched results. It appears that the work was somewhat in progress and not complete at the time of submission.
Questions for authors
  • What do the authors mean by “the latter uses a different agent harness.” Am I right in understanding that the harness is not comparable across settings? If so, what is the point of presenting Table 1 results?
  • I couldn’t find any AI use statement in the paper. According to my experience working with agents as well as their style of writing, the paper seems heavily AI assisted, could you provide more details regarding AI use?
Reviewer_Vx7nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SWE-Loop proposes an adaptive training framework in which paired teacher/student execution traces identify capability gaps and guide the next distribution of repository-level repair tasks. Task-validity and infrastructure gates precede attribution to four software-engineering pillars, and verified successful or recovery trajectories supply SFT data with retention replay. The implementation uses a Qwen3.5-9B student and GLM-5.2 teacher. The paper reports three Guided iterations across five external benchmarks; iteration 3 reaches 51.20% on SWE-bench Verified and 42.67% on Lite.

Critical strengths

A well-motivated approach to adaptive training. The paper presents a compelling motivation for adaptive training: as a model’s capabilities improve, its training data should evolve to target its remaining weaknesses. SWE-Loop follows this intuition by using execution traces to identify capability gaps and guide subsequent task generation.

Critical weaknesses
  1. Limited baselines and ablations to support the effectiveness of trace-guided training.

The paper does not include a matched static-data SFT baseline, making it difficult to determine whether trace-guided task generation provides benefits over training on a fixed task distribution. Additionally, Table 2 does not present controlled one-factor ablations, as the reported results come from earlier Guided iterations and separate pilot training recipes. Without these comparisons, it is difficult to isolate the contribution of the proposed components or establish whether adaptive task generation is responsible for the observed improvements.

  1. Insufficient details about the training data.

Although the paper describes its data-generation pipeline, it does not clearly specify the actual training data used or the amount of SFT data in each iteration. This makes it difficult to compare the results fairly with prior work or determine whether the performance gains come from the proposed method or differences in training data. More detailed reporting of dataset composition, size, and changes across iterations would improve reproducibility and strengthen the empirical evaluation.

Questions for authors
  1. Static-data SFT comparison: Have the Fixed, Random, and Outcome-only baselines described in §4.3 been evaluated? If so, could the authors provide their per-iteration results under matched training and data-generation budgets? In particular, does Guided consistently outperform static-data SFT?
  2. Ablation clarification: Are there any controlled one-factor ablations beyond the historical comparisons in Table 2? Which component-level improvements are supported by controlled experiments rather than comparisons across different training iterations or pilot settings?
  3. Training-data details: Could the authors clarify which repositories and tasks were used for training, along with the amount of SFT data in each iteration? How does the training data compare in size and composition with that used by prior methods?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.