Pith.

Artificial Intelligence · arXiv:2610.33667 v1 · submitted 24 Oct 2026 · 47 references

SWE-Ladder: Learning to Localize, Reproduce, and Patch via Progressive Privileged Self-Distillation

Xudong Cai, Zhengxi Lu, Yuchen Yan, Jian Shao, Jun Xiao, Yueting Zhuang, Yongliang Shen

The paper says

“Resolving a repository-level issue typically requires a software-engineering (SWE) agent to localize the faulty code, reproduce the failure, and patch the code.”

more of the abstract

Reinforcement learning trains such agents from a single test result on the final patch, while on-policy self-distillation (OPSD) adds token-level supervision from a teacher with privileged evidence, exposing the full reference patch at every step. Neither provides stage-specific guidance for localization or reproduction. We propose SWE-Ladder, which turns the verified repair of each training instance into stage-specific evidence: the changed files and lines for localization, the fail-to-pass tests for reproduction, and the diff for patching. A rule-based router assigns each step of an end-to-end rollout to a stage based on its tool call, a frozen teacher sees only the matching evidence, and the stages are distilled in repair order while GRPO optimizes the full trajectory. With Qwen3-4B-Instruct-2507 on SWE-bench Verified under the R2E-Gym scaffold, SWE-Ladder improves pass@1 from 9.6% to 16.4%, compared with 10.4% for GRPO and 12.4% for OPSD with the full patch. Under the OpenHands scaffold, it improves pass@1 from 13.0% to 17.6%. Reversing the stage order under R2E-Gym reduces pass@1 to 11.4%.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 27 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: major revision.

A cleanly motivated method with honest limitations; the statistical foundation is thinner than the headline numbers suggest, but the paper is worth engaging with. This line is wrong

3 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 47 references say.

The paper says, at §4.2, Table 2; Limitations (sentence b167)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.95):

“The reported training curves and router comparison each fol- low one checkpoint lineage without independent random-seed repeats or confidence intervals, so the observed optimization trends and downstream differences should be interpreted as configuration- specific evidence.”

Pith says: Objection 1 of 3 · major

All numerical comparisons (SWE-Ladder vs. GRPO, GRPO+OPSD, SDAR) are based on a single training run per configuration. The paper itself states in the Limitations that 'the reported training curves and router comparison each follow one checkpoint lineage without independent random-seed repeats or confidence intervals, so the observed optimization trends and downstream differences should be interpreted as configuration-specific evidence.'

more of objection 1

On a binary pass@1 metric with 500 tasks, the standard error near 15% is about 1.5 percentage points; the headline gaps (e.g., SWE-Ladder 16.4% vs. GRPO+OPSD 12.4%) are 4 pp, which could be within the noise of a single seed. Without variance estimates or repeated runs, the central claim that SWE-Ladder reliably outperforms the baselines is not statistically supported. This is a load-bearing issue because the paper's primary message is quantitative improvement.

The paper says, at §4.3, Figure 5; §4.4, Figure 8

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):

“The reported training curves and router comparison each fol- low one checkpoint lineage without independent random-seed repeats or confidence intervals, so the observed optimization trends and downstream differences should be interpreted as configuration- specific evidence.”

Pith says: Objection 2 of 3 · major

The training-dynamics diagnostics and the neural-router comparison also rest on single runs. The observed stability of SWE-Ladder's entropy and gradient norm (versus GRPO's decline) might be seed-dependent. The claim that the rule-based router gives better downstream results than the neural router (Figure 8) likewise lacks error bars. The paper's own caveat about configuration-specific evidence applies here as well. Adding at least three random seeds and reporting means and standard deviations would substantially strengthen the empirical basis.

The paper says, at §4.5, Table 5

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.66):

“The reported training curves and router comparison each fol- low one checkpoint lineage without independent random-seed repeats or confidence intervals, so the observed optimization trends and downstream differences should be interpreted as configuration- specific evidence.”

Pith says: Objection 3 of 3 · major

The privilege–target alignment ablation (Tables 5) uses a reduced 30-step budget and reports small differences (0.8 and 1.4 pp). With a single run, these differences are indistinguishable from noise. The conclusion that 'privilege should match the target stage' is plausible but not reliably established by these data. The paper should either provide confidence intervals or temper the claim.

6 minor objections

The paper says, at Figure 2 and Appendix A.1

Pith says: minor objection 1 of 6

The 'Verification' metric (post-edit re-run of a reproduction script) is a custom behavioral diagnostic, not a standard benchmark metric. The main text (Figure 2 caption and Section 1) should explicitly reference Appendix A.1 for the definition, as it is central to the paper's claim that SWE-Ladder improves verification while others do not.

The paper says, at Table 2

Pith says: minor objection 2 of 6

The SDAR baseline drops sharply (4.6% on R2E-Gym, 6.8% on OpenHands) relative to the base model. This may reflect hyperparameter sensitivity or implementation mismatch. The paper should comment on whether SDAR's hyperparameters were tuned and whether this drop is expected, as it affects the fairness of the comparison.

show 4 more

The paper says, at §3.5, Appendix A.7

Pith says: minor objection 3 of 6

The progressive schedule uses 30, 20, 20 iterations for LOC, REP, PATCH respectively. The total of 70 iterations is the same as the baselines, but the allocation is asymmetric. Could the performance of SWE-Ladder be partly due to more iterations on the early stage (LOC)? The authors should discuss or ablate the effect of phase budget distribution.

The paper says, at Table 2, column heads

Pith says: minor objection 4 of 6

The column 'env_done(%)' should be explicitly defined (e.g., percentage of trajectories that terminate normally vs. hitting a resource limit). This is mentioned in Appendix A.5 but should be noted in the table caption or a footnote.

The paper says, at Appendix A.2

Pith says: minor objection 5 of 6

The rule-based router is described in detail here, but many readers may miss it. Consider summarizing the key backtracking rules (e.g., three consecutive exploration actions revert to LOC) in the main text, as the router is a core component of the method.

The paper says, at General

Pith says: minor objection 6 of 6

The paper states that all results are 'configuration-specific evidence' and that no generalizations are claimed beyond the tested settings. This is commendably honest, but it also highlights that the conclusions are tentative. The abstract and introduction should reflect this caveat more prominently, for instance by noting that the results are based on a single seed and should be confirmed with repeated runs.

What would settle it

Repeat the full SWE-Ladder training with five independent random seeds (different initializations of the base model, different rollout sampling) for SWE-Ladder, GRPO, and GRPO+OPSD on R2E-Gym. Report mean pass@1 and 95% confidence intervals (or standard deviation) across seeds. If the 95% CI for SWE-Ladder does not overlap with the CI for GRPO+OPSD, the claim is statistically robust. If the intervals overlap substantially, the claim weakens.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 27 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic partly agrees: says it holds up only with conditions 28 s · 19,708 tokens read, 2,710 written Central numerical claims rest on a single training run per configuration with no confidence intervals; the paper itself warns that differences are 'configuration-specific evidence', so the reported improvements may not be statistically reliable.
  3. Editor wrote the verdict above 19 s · 20,637 tokens read, 1,658 written
  4. Referee major revision: 3 major objections, 6 minor 18 s · 20,141 tokens read, 2,089 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper proposes a sensible idea—decomposing privileged teacher evidence by repair stage (localization, reproduction, patching) and introducing the stages progressively. The execution is careful: same base model, same scaffolds, multiple baselines, informative ablations, and training-dynamics diagnostics. The authors are unusually forthright about their limitations (single model, single benchmark, single seed per configuration). I'd rather have a paper like this than one that oversells.

The core novelty is the stage-specific evidence and the ordering schedule. The ablations show that removing any stage, reversing the order, or training all stages at once drops pass@1, supporting the design. The stage-wise behavioral diagnostics (localization, reproduction, verification) provide a clearer picture of what each method actually learns. The paper also ships code and data, which is real evidence.

The main soft spot is statistical: the headline gains (16.4% vs 12.4% for GRPO+OPSD, etc.) come from one training run per configuration. The authors acknowledge this (b167: 'configuration-specific evidence'), but acknowledgment does not change the fact that on a 500-task binary metric with ~15% solve rate, the standard error is about 1.5 pp; a 4 pp gap may or may not replicate. The same issue applies to the order and alignment ablations. Without seed repeats, the reader cannot distinguish a robust effect from random fluctuation. The other limitations (single backbone, single benchmark, reliance on verified repair metadata) are real but more typical for a first paper of this kind.

Who is this for? Anyone working on post-training for SWE agents or on using privileged information in multi-turn RL. The paper deserves a serious referee. I would accept it for peer review, and I'd ask for at least 2-3 independent seeds (or a statistical test like bootstrap) to confirm that the reported gaps are systematic. The authors already have the infrastructure; running a few extra seeds would substantially strengthen the claims.

The strongest objection

The paper's headline pass@1 gains (9.6%→16.4% on R2E-Gym, 13.0%→17.6% on OpenHands) are each reported from exactly one training run. No independent seed repeats, confidence intervals, or significance tests are provided. The authors explicitly state in the Limitations (sentence b167) that 'the observed optimization trends and downstream differences should be interpreted as configuration-specific evidence.' This means the observed gaps (e.g., 16.4% vs 12.4% for GRPO+OPSD) could be within the noise of a single seed. On a binary pass@1 metric with 500 tasks, the standard error of a proportion near 15% is roughly 1.5 percentage points; a 4-point gap is not necessarily significant without multiple runs. The stage order ablations in Table 4 (11.4%, 12.6%, 16.4%) lack similar error bars. The ablation on privilege–target alignment (Table 5) uses a reduced 30-step budget and shows small differences (0.8 and 1.4 pp) that could also be noise. Without seed-level variance, the reader cannot assess whether SWE-Ladder is reliably better than alternatives or whether the ablations truly reveal sensitivity to the design choices. The paper acknowledges this limitation but does not address it, so the central claim lacks sufficient statistical grounding.

The reader identified the weakest assumption as the clean separation of evidence, which is supported by ablations. However, the more fundamental empirical concern is the single-seed nature of all experimental results, which the reader noted but still accepted. The paper's own limitations text cautions against overinterpretation. Given that the central claim is quantitative and comparative, the lack of statistical evidence is a load-bearing issue that, in my view, makes acceptance premature. I recommend CONDITIONAL, pending seed-repeat validation. This does not invalidate the method's novelty or thoroughness, but it raises the correctness risk from 'medium' to 'high' until variance is quantified. I agree with the reader on the method's novelty and the careful ablations, but disagree on the verdict threshold given the statistical weakness.

The referee's report

If the reported improvements are statistically reliable, the work makes a genuine contribution. The stage-wise diagnostics provide insight into what different training signals teach, and the progressive distillation schedule is a principled way to decompose a long-horizon task without constraining the agent's actions. The ablations are thorough and the method is clearly described. The results on two scaffolds strengthen the claim of generality, and the open-source code supports reproducibility.

Pith's reading of the paper (a machine's summary, not the authors')

The paper aims to establish that a software-engineering agent's training benefits when the privileged teacher sees only the evidence relevant to the current repair stage (localization, reproduction, or patch generation), and when the stages are introduced in the order they are used during repair. This matters because existing methods either give a single reward per trajectory (GRPO) or show the full reference patch at every step (OPSD), neither of which provides targeted supervision for intermediate skills. The proposed method, SWE-Ladder, improves pass@1 on SWE-bench Verified from 9.6% to 16.4% under the R2E-Gym scaffold and from 13.0% to 17.6% under the OpenHands scaffold, outperforming both GRPO and full-patch OPSD. Ablations confirm that removing any stage, reversing the order, or training all stages at once reduces performance, and that a rule-based router with 92% stage accuracy outperforms a more precise neural router.

What carries the argument. The method's load-bearing components are (1) a deterministic rule-based router that assigns each tool-call step to one of three repair stages (LOC, REP, PATCH) based on its function and path, using an ordered state machine with backtracking; (2) three evidence slices extracted from the verified repair metadata—changed file paths and old-side hunk lines for localization, fail-to-pass test identifiers and code for reproduction, and the reference diff for patching; (3) a progressive training schedule that activates one stage at a time (30 iterations for LOC, then 20 for REP, then 20 for PATCH), conditioning a frozen teacher on only that stage's evidence and applying the On-Policy Self-Distillation (OPSD) loss only to tokens assigned to the active stage; (4) a joint objective combining GRPO over the full trajectory with the stage-masked OPSD term.

Core claim. The paper discovers that giving the teacher stage-specific privileged evidence (changed files for localization, fail-to-pass tests for reproduction, the diff for patching) and progressively introducing these stages in the natural repair order (LOC → REP → PATCH) yields a more capable and stable agent than either outcome-only reinforcement learning or full-patch on-policy self-distillation. On SWE-bench Verified, SWE-Ladder is the only method that improves localization, reproduction, verification, and pass@1 simultaneously. The paper further finds that stage order is critical: reversing the order drops pass@1 from 16.4% to 11.4%, and training all stages concurrently reaches only 12.4%. A rule-based stage router, despite lower per-step accuracy than a neural router, produces more temporally consistent labels and leads to better downstream performance.

The premise it rests on. The central assumption is that the verified repair metadata can be reliably parsed into three independent evidence views (changed files for localization, fail-to-pass tests for reproduction, diff for patching) and that the rule-based router assigns each tool-call step to the correct stage without leaking cross-stage information or fragmenting trajectories.

If the paper is right (Pith's inferences, not the authors')

The stage-aligned supervision principle can be applied to other multi-step agent tasks where intermediate behaviors are poorly captured by final rewards, such as scientific experiment design or robotic task sequencing.

The progressive schedule suggests a natural curriculum for complex workflows: train prerequisite skills (localization) before downstream skills (reproduction, patch generation), which could inform training for other decomposable agents.

SWE-Ladder achieves 98.6% trajectory completion versus 36.0% for GRPO, indicating that stage-specific distillation may prevent the agent from drifting or terminating prematurely, a robustness benefit that extends beyond raw solve rate.

The rule-based router's advantage over a neural router implies that temporal consistency in stage labeling is more important than per-step accuracy for training, a finding that may guide the design of labeling systems for agent supervision.

Because no external data or model is used beyond the verified repair metadata and a frozen base-model teacher, the approach is reproducible without costly human annotation or proprietary APIs.

The principle of stage-matched privileged evidence could generalize to other domains such as scientific reasoning, where a teacher should see only the hypothesis at the hypothesis-formation stage and only the experiment design at the experiment-design stage.

The observation that full-patch supervision lowers verification suggests a 'lazy teacher' effect: providing too much privileged information can discourage the student from learning intermediate validation steps, a risk that likely applies to any multi-step reasoning task.

The neural router's failure despite higher per-step agreement implies that training signal consistency over time is more critical than per-step annotation accuracy, which could influence how we design automatic labeling systems for agent supervision.

A direct test: if the rule-based router were replaced by an oracle with perfect temporal consistency (e.g., a human annotator), performance might improve further, but the current gap (16.4% vs. 12.0% for neural) suggests the labeling-quality trade-off merits deeper study.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Resolving a repository-level issue typically requires a software-engineering (SWE) agent to localize the faulty code, reproduce the failure, and patch the code. Reinforcement learning trains such agents from a single test result on the final patch, while on-policy self-distillation (OPSD) adds token-level supervision from a teacher with privileged evidence, exposing the full reference patch at every step. Neither provides stage-specific guidance for localization or reproduction. We propose SWE-Ladder, which turns the verified repair of each training instance into stage-specific evidence: the changed files and lines for localization, the fail-to-pass tests for reproduction, and the diff for patching. A rule-based router assigns each step of an end-to-end rollout to a stage based on its tool call, a frozen teacher sees only the matching evidence, and the stages are distilled in repair order while GRPO optimizes the full trajectory. With Qwen3-4B-Instruct-2507 on SWE-bench Verified under the R2E-Gym scaffold, SWE-Ladder improves pass@1 from 9.6% to 16.4%, compared with 10.4% for GRPO and 12.4% for OPSD with the full patch. Under the OpenHands scaffold, it improves pass@1 from 13.0% to 17.6%. Reversing the stage order under R2E-Gym reduces pass@1 to 11.4%.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-27

The authors' main claim, as the classifier picked it out 0.97

“With Qwen3-4B-Instruct-2507 on SWE-bench Verified under the R2E-Gym scaffold, SWE-Ladder improves pass@1 from 9.6% to 16.4%, compared with 10.4% for GRPO and 12.4% for OPSD with the full patch.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.34 · human participants 0.02 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.04

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Resolving a repository-level issue typically requires a software-engineering (SWE) agent to localize the faulty code, reproduce the failure, and patch the code. background 1.00 0.05 0.02 asserted 0.97
Reinforcement learning trains such agents from a single test result on the final patch, while on-policy self-distillation (OPSD) adds token-level supervision from a teacher with privileged evidence, exposing the full reference patch at every step. background 0.98 0.06 0.03 asserted 1.00
Neither provides stage-specific guidance for localization or reproduction. gap 1.00 0.07 0.04 asserted 1.00
We propose SWE-Ladder, which turns the verified repair of each training instance into stage-specific evidence: the changed files and lines for localization, the fail-to-pass tests for reproduction, and the diff for patching. contribution 0.91 0.90 0.20 asserted 0.99
A rule-based router assigns each step of an end-to-end rollout to a stage based on its tool call, a frozen teacher sees only the matching evidence, and the stages are distilled in repair order while GRPO optimizes the full trajectory. method 1.00 0.88 0.06 asserted 1.00
With Qwen3-4B-Instruct-2507 on SWE-bench Verified under the R2E-Gym scaffold, SWE-Ladder improves pass@1 from 9.6% to 16.4%, compared with 10.4% for GRPO and 12.4% for OPSD with the full patch. result 1.00 0.96 0.47 asserted 1.00
Under the OpenHands scaffold, it improves pass@1 from 13.0% to 17.6%. result 1.00 0.94 0.10 asserted 1.00
Reversing the stage order under R2E-Gym reduces pass@1 to 11.4%. result 0.99 0.93 0.06 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-27: 1 request, 6,478 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (11)
Figure 1 of the paper
Figure 1. Supervision signals compared in this pa￾per. GRPO assigns one reward to the whole trajectory; GRPO+OPSD shows the teacher the full reference patch at every step; SWE-LADDER gives the teacher stage￾specific evidence and introduces the stages one at a time. denser supervision (Zhao et al., 2026; Hübotter et al., 2026). The same model acts as a teacher that receives privileged evidence withheld from the student and scores every token of the student’s trajectory (Lu and Thinking Machines Lab, 2025; Agarwal et al., 2024). In SWE, the available evi￾dence is the verified repair of each…
Figure 2 of the paper
Figure 2. Stage-wise diagnostics on SWE-bench Ver￾ified. Bars show changes from the base model at the final checkpoint; labels mark the changes discussed in Section 1. Base values appear under each metric. Defi￾nitions are in Appendix A.1. different skills. Localization needs to know where the fault is, reproduction needs to know how the failure shows up, and patching needs to know what to change. During a localization step, a teacher that sees the full patch receives not only localiza￾tion cues but also patch-generation details, and may therefore favor editing over searching. A similar…
Figure 3 of the paper
Figure 3. Stage-routed privileged distillation for SWE. A rule-based router labels each step of an end-to-end rollout as LOC, REP, or PATCH. In each phase, a frozen teacher receives one stage’s evidence and distills only that stage’s tokens, while GRPO optimizes the full trajectory. human-validated repository issues, and report pass@1(Jimenez et al., 2024; Chowdhury et al., 2024). An issue is counted as solved only when the submitted patch passes all fail-to-pass and pass-to￾pass unit tests. The primary RL comparisons use the same issues, R2E-Gym (Jain et al., 2025) agent scaffold, sandbox,…
Figure 5 of the paper
Figure 5. Training-dynamics diagnostics. Entropy, gradient norm, response length, and shaped training reward of GRPO, GRPO+OPSD, SDAR, and SWE￾LADDER over 70 iterations. Shaded regions mark the LOC, REP, and PATCH phases of SWE-LADDER. 4.3 Training Dynamics To investigate training stability, we compare SWE￾LADDER with GRPO, GRPO+OPSD, and SDAR
Figure 6 of the paper
Figure 6. Action counts per turn across checkpoints. For the base model, LOC30, REP50, and PATCH70, each panel stacks the number of trajectories taking a search/read/view, file-edit, test/reproduction, or other action at each turn on SWE-bench Verified (n = 500). Privilege should match the target stage. We compare matched and swapped LOC/PATCH ev￾idence under a 30-iteration budget (
Figure 7 of the paper
Figure 7. Capability changes across checkpoints. Cells show changes (%) from Base for localization (Loc.), reproduction (Rep.), verification (Verify.), and pass@1 at Base, LOC30, REP50, and PATCH70. Darker blue indicates a larger signed change. Target PI → target Solve (%) Avg. steps env_done (%) LOC LOC→LOC 14.4 21.0 99.0 PATCH→LOC 13.6 ↓−0.8 15.9 ↓−5.1 98.8 ↓−0.2 PATCH PATCH→PATCH 13.2 34.0 82.8 LOC→PATCH 11.8 ↓−1.4 20.2 ↓−13.8 99.0 ↑+16.2
Figure 8 of the paper
Figure 8. reports the supplied neural-router eval￾uation points alongside the complete rule-routed SWE-LADDER schedule. Despite its lower label￾agreement score, the rule-based router gives the stronger downstream training result in this compar￾ison, so annotation accuracy alone is not a sufficient proxy for training utility. A.4 Evaluation Protocols For the OpenHands rows in
Figure 9 of the paper
Figure 9. Localization privileged information. Candidate files, old-side hunk ranges, and source fragments identify where the issue should be investigated without revealing the replacement patch
Figure 10 of the paper
Figure 10. Reproduction privileged information. Fail-to-pass tests, reference test snippets, and candidate test changes describe the failing behavior and provide a starting point for reproducing it
Figure 11 of the paper
Figure 11. Patch privileged information. The verified repair diff specifies the intended source-code correction available to the patch-stage teacher

Showing 10 of the paper's 11 figures. Figures are the authors', reproduced from the paper.

References (47)
  1. [1] Swe-bench: Can language models resolve real-world github issues? , author=. International Conference on Learning Representations , volume= on Pith
  2. [2] Chowdhury, Neil and Aung, James and Chan, Jun Shern and Jaffe, Oliver and Sherburn, Dane and Starace, Giulio and Mays, Evan and Dias, Rachel and Aljubeh, Marwan and Glaese, Mia and Jimenez, Carlos E. and Yang, John and Ho, Leyton and Patwardhan, Tejal and Liu, Kevin and Madry,... on Pith
  3. [3] Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and others , year =. doi:10.48550/arXiv.2505.09388 , url =. 2505.09388 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2505.09388
  4. [4] Swe-rebench: An automated pipeline for task collection and decontaminated evaluation of software engineering agents , author=. Advances in Neural Information Processing Systems , volume= on Pith
  5. [5] Openhands: An open platform for ai software developers as generalist agents , author=. International Conference on Learning Representations , volume= on Pith
  6. [6] Swe-master: Unleashing the potential of software engineering agents via post-training , author=. arXiv preprint arXiv:2602.03411 , year= on PitharXiv
  7. [7] Swe-agent: Agent-computer interfaces enable automated software engineering , author=. Advances in Neural Information Processing Systems , volume= on Pith
  8. [8] Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement , author=. International Conference on Learning Representations , volume= on Pith
  9. [9] Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year= on PithPith's reviewarXiv
  10. [10] Dapo: An open-source llm reinforcement learning system at scale , author=. Advances in Neural Information Processing Systems , volume= on Pith
  11. [11] Understanding r1-zero-like training: A critical perspective , author=. arXiv preprint arXiv:2503.20783 , year= on PithPith's reviewarXiv
  12. [12] Gtpo: Stabilizing group relative policy optimization via gradient and entropy control , author=. Transactions of the Association for Computational Linguistics , volume=. 2026 , publisher= on Pith
  13. [13] Swe-rl: Advancing llm reasoning via reinforcement learning on open software evolution , author=. Advances in Neural Information Processing Systems , volume= on Pith
  14. [14] Luo, Michael and Jain, Naman and Singh, Jaskirat and Tan, Sijun and Patel, Ameen and Wu, Qingyang and Ariyak, Alpay and Cai, Colin and Venkat, Tarun and Zhu, Shang and Athiwaratkun, Ben and Roongta, Manan and Zhang, Ce and Li, Li Erran and Popa, Raluca Ada and Sen, Koushik and... on Pith
  15. [15] Training long-context, multi-turn software engineering agents with reinforcement learning , author=. arXiv preprint arXiv:2508.03501 , year= on PitharXiv
  16. [16] Training software engineering agents and verifiers with swe-gym , author=. arXiv preprint arXiv:2412.21139 , year= on PithPith's reviewarXiv
  17. [17] R2e-gym: Procedural environments and hybrid verifiers for scaling open-weights swe agents , author=. arXiv preprint arXiv:2504.07164 , year= on PithPith's reviewarXiv
  18. [18] Swe-smith: Scaling data for software engineering agents , author=. Advances in Neural Information Processing Systems , volume= on Pith
  19. [19] Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year= on PithPith's reviewarXiv
  20. [20] Self-distilled reasoner: On-policy self-distillation for large language models , author=. arXiv preprint arXiv:2601.18734 , year= on PithPith's reviewarXiv
show 27 more
  1. [21] Swe-universe: Scale real-world verifiable environments to millions , author=. arXiv preprint arXiv:2602.02361 , year= on PithPith's reviewarXiv
  2. [22] Bugpilot: Complex bug generation for efficient learning of swe skills , author=. arXiv preprint arXiv:2510.19898 , year= on PitharXiv
  3. [23] Kimi-dev: Agentless training as skill prior for swe-agents , author=. International Conference on Learning Representations , volume= on Pith
  4. [24] Swe-swiss: A multi-task fine-tuning and rl recipe for high-performance issue resolution , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages= on Pith
  5. [25] Agentless: Demystifying llm-based software engineering agents , author=. arXiv preprint arXiv:2407.01489 , year= on PithPith's reviewarXiv
  6. [26] Swe-fixer: Training open-source llms for effective and efficient github issue resolution , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages= on Pith
  7. [27] Sorft: Issue resolving with subtask-oriented reinforced fine-tuning , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages= on Pith
  8. [28] Swt-bench: Testing and validating real-world bug-fixes with code agents , author=. Advances in Neural Information Processing Systems , volume= on Pith
  9. [29] Assertflip: Reproducing bugs via inversion of llm-generated passing tests , author=. Proceedings of the 2026 IEEE/ACM 48th International Conference on Software Engineering , pages= on Pith
  10. [30] Autocoderover: Autonomous program improvement, 2024 , author=. URL https://arxiv. org/abs/2404.05427 , year= on PithPith's reviewarXiv
  11. [31] On-Policy Distillation , author =. Thinking Machines Lab: Connectionism , year =. doi:10.64434/tml.20251026 , url = on Pithdoi:10.64434/tml.20251026
  12. [32] On-policy distillation of language models: Learning from self-generated mistakes , author=. International Conference on Learning Representations , volume= on Pith
  13. [33] Teacher-guided policy optimization for on-policy reasoning distillation under large policy divergence , author=. arXiv preprint arXiv:2605.13230 , year= on PithPith's reviewarXiv
  14. [34] Self-distilled rlvr , author=. arXiv preprint arXiv:2604.03128 , year= on PithPith's reviewarXiv
  15. [35] Rethinking On-Policy Self-Distillation for Thinking Models , author=. arXiv preprint arXiv:2607.05184 , year= on PithPith's reviewarXiv
  16. [36] Hdpo: Hybrid distillation policy optimization via privileged self-distillation , author=. arXiv preprint arXiv:2603.23871 , year= on PitharXiv
  17. [37] Self-distilled agentic reinforcement learning , author=. arXiv preprint arXiv:2605.15155 , year= on PithPith's reviewarXiv
  18. [38] StepOPSD: Step-Aware Online Preference Distillation for Agent Reinforcement Learning , author=. arXiv preprint arXiv:2605.27140 , year= on PithPith's reviewarXiv
  19. [39] Sod: Step-wise on-policy distillation for small language model agents , author=. arXiv preprint arXiv:2605.07725 , year= on PithPith's reviewarXiv
  20. [40] Tcod: Exploring temporal curriculum in on-policy distillation for multi-turn autonomous agents , author=. arXiv preprint arXiv:2604.24005 , year= on PithPith's reviewarXiv
  21. [41] Reinforcement learning via self-distillation , author=. arXiv preprint arXiv:2601.20802 , year= on PithPith's reviewarXiv
  22. [42] A survey of on-policy distillation for large language models , author=. arXiv preprint arXiv:2604.00626 , year= on PithPith's reviewarXiv
  23. [43] Beyond Absolute Imitation: Anchored Residual Guidance for Privileged On-Policy Distillation , author=. arXiv preprint arXiv:2606.10385 , year= on PithPith's reviewarXiv
  24. [44] EviSD: Evidence-Conditioned Self-Distillation for Search-Augmented Agents , author=. arXiv preprint arXiv:2608.01359 , year= on PithPith's reviewarXiv
  25. [45] Ladder: Self-improving llms through recursive problem decomposition , author=. arXiv preprint arXiv:2503.00735 , year= on PithPith's reviewarXiv
  26. [46] Adaptive Teacher Exposure for Self-Distillation in LLM Reasoning , author=. arXiv preprint arXiv:2605.11458 , year= on PithPith's reviewarXiv
  27. [47] 2026 , month = aug, howpublished = on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment