pith. sign in

arxiv: 2605.21996 · v1 · pith:O73PNWWMnew · submitted 2026-05-21 · 💻 cs.SE · cs.AI

From Patches to Trajectories: Privileged Process Supervision for Software-Engineering Agents

Murong Ma , Tianyu Chen , Yun Lin , Shuai Lu , Qinglin Zhu , Yeyun Gong , Zhiyong Huang , Peng Cheng
show 2 more authors
Yan Lu Jin Song Dong
This is my paper

Pith reviewed 2026-05-22 05:09 UTC · model grok-4.3

classification 💻 cs.SE cs.AI
keywords software engineering agentstrajectory supervisionprivileged informationprocess supervisionSWE-benchsupervised fine-tuningreference patchesbi-objective optimization
0
0 comments X

The pith

P2T converts reference patches into short effective trajectories that raise software-engineering agent success rates while lowering inference cost.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that standard supervised fine-tuning on teacher trajectories passes on both final successes and intermediate flaws such as ungrounded steps or redundant loops. It shows that reference patches, which are normally discarded, can be turned into a latent process graph that scores each step for grounded progress without leaking the answer. By optimizing for both effectiveness per step and overall length, the method produces training data that yields higher benchmark performance from far fewer instances than outcome-only filtering.

Core claim

Patches-to-Trajectories (P2T) distills a developer reference patch p* into a latent process graph G* of contextual facts and solution milestones, then scores blinded teacher continuations against this graph under a leakage-blocking groundedness check to retain only the shortest effective trajectory segments; training on 1.8k such curated instances from SWE-Gym raises Pass@1 by up to 10.8 points on SWE-bench Verified and reduces per-instance inference cost by approximately 15 percent relative to outcome-filtered baselines.

What carries the argument

latent process graph G* distilled from reference patch p* that supplies a grounded, leakage-blocking measure of per-step progress for scoring teacher continuations

If this is right

  • Training data quality can be improved without increasing data volume or requiring the teacher model to succeed on every instance.
  • Trajectory length and per-step effectiveness can be jointly optimized rather than relying solely on terminal outcome verification.
  • Privileged information from reference patches can be used during data curation while still keeping the student blind to the patch at inference time.
  • Consistent gains appear across both SWE-bench Verified and Lite when the same curation procedure is applied.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same distillation step could be applied to other domains that already produce reference solutions, such as theorem proving or automated program repair outside SWE-bench.
  • If the process graph construction can be made cheaper, the method might lower the compute barrier for creating high-quality agent training sets at larger scale.

Load-bearing premise

A reference patch can be distilled into a process graph that accurately measures real progress in blinded teacher attempts without introducing selection bias or needing the teacher to have already solved the task.

What would settle it

Run the identical training and evaluation protocol on SWE-bench Verified but replace the G*-based scoring with random or length-only selection of the same number of trajectory segments and observe whether the reported gains in Pass@1 and cost reduction disappear.

Figures

Figures reproduced from arXiv: 2605.21996 by Jin Song Dong, Murong Ma, Peng Cheng, Qinglin Zhu, Shuai Lu, Tianyu Chen, Yan Lu, Yeyun Gong, Yun Lin, Zhiyong Huang.

Figure 1
Figure 1. Figure 1: Overview of P2T. Phase 1 (Sec. 4.1) distills the reference patch [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect of P2T on trajectory quality, traced from supervision (a, b) to evaluation (c, d). Each panel reports the paired Mann–Whitney p-value, the relative shift in mean (∆µ), and Cliff’s δ; diamonds mark means. tail of rollouts that exhaust the 100-iteration budget largely disappears ( [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Composition of the prerequisite graphs {G⋆ i } across the N = 1,815 SWE-Gym training instances (33,106 in-scope nodes). (a) Aggregate node-type breakdown: static facts dominate (66.2%), with dynamic facts (16.7%) and the three artifact categories (reproduction, analysis, fix plan) at 5–6% each. (b) Per-instance mean count by category (bars) and fraction of instances containing at least one node of that cat… view at source ↗
Figure 4
Figure 4. Figure 4: Value of information in G⋆ . (a) Resolve rate of a blinded reference solver on the 1.8k training instances as elements of the prerequisite graph are progressively revealed; the oracle patch p ⋆ is never exposed. Each addition contributes a non-negative marginal gain under both teachers, with facts and the reproduction script accounting for most of the lift. (b) When the graph is not disclosed, instances on… view at source ↗
Figure 5
Figure 5. Figure 5: Distilled prerequisite graph G⋆ for moto#6041. Static facts (blue) form the contextual layer; the dynamic fact f6 (orange) requires execution; the artifact layer (green/grey/red/purple) encodes the reproduction, analysis, plan, edit, and validation milestones. Edges denote prerequisite relations enforced by the critic during distillation [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Window-22 candidate pool in (Len22,Eff22) space, with Len22 measured in response tokens (assistant messages plus their observations). Both pure seeds (grey circles) sit far below the floor η22 = 0.50 (shaded band, dashed line). The two mutated variants tie at Eff = 0.70; the tie-break by length picks the seed-0 mutation (red star), which is committed as s ⋆ 22. Side-by-side comparison [PITH_FULL_IMAGE:fig… view at source ↗
Figure 7
Figure 7. Figure 7: Side-by-side fragment of the same window in [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
read the original abstract

Supervised fine-tuning (SFT) on long teacher trajectories is the dominant way to instill investigation and reasoning in open software-engineering (SWE) agents. Since every retained response becomes an imitation target, the student inherits the final outcome and intermediate flaws, including ungrounded leaps and redundant loops. High-quality training data must be effective(each step is grounded and narrows the agent's epistemic gap to the correct fix) and efficient(each step is information-bearing rather than redundant or looping). Existing recipes filter or relabel teacher rollouts using only a binary terminal verifier, which does not directly target these axes and provides no supervision on instances where the teacher fails. Most real issue includes a developer-authored reference patch, $p^\star$, revealing the file paths, runtime behaviors, and coding conventions presupposed by the correct fix, yet standard pipelines discard it. We propose Patches-to-Trajectories (P2T), which uses $p^\star$ as privileged information during curation and formulates trajectory construction as bi-objective optimization over per-step effectiveness and trajectory length. A reverse phase distills $p^\star$ into a latent process graph, $G^\star$, of contextual facts and solution milestones. A forward phase curates trajectories from blinded teacher continuations by scoring per-step progress against $G^\star$ under a leakage-blocking groundedness check and retaining the shortest effective segments. Using only 1.8k curated SWE-Gym instances, P2T improves effectiveness and efficiency over outcome-filtered SFT and its tool-error-masking variant. On SWE-bench Verified, it raises Pass@1 by up to 10.8 points while reducing per-instance inference cost by ~15%, with consistent gains on SWE-bench Lite. Size-matched ablations and qualitative analysis further isolate trajectory quality from data scale.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Patches-to-Trajectories (P2T), which treats developer-authored reference patches p* as privileged information during data curation for supervised fine-tuning of software-engineering agents. It distills p* into a latent process graph G* in a reverse phase, then uses a forward phase to score blinded teacher continuations against G* with a leakage-blocking groundedness check and bi-objective optimization over per-step effectiveness and trajectory length, retaining the shortest effective segments. Experiments on 1.8k SWE-Gym instances show Pass@1 gains of up to 10.8 points on SWE-bench Verified (and consistent gains on Lite) while cutting per-instance inference cost by ~15%, outperforming outcome-filtered SFT and tool-error-masking variants; size-matched ablations and qualitative analysis are used to attribute gains to trajectory quality rather than scale.

Significance. If the leakage-blocking property of the G*-based scoring holds without introducing solution-specific bias, the approach offers a concrete way to convert privileged patches into process-level supervision signals that improve both effectiveness and efficiency of training trajectories for SWE agents. The use of bi-objective optimization and explicit separation of reverse distillation from forward curation on blinded rollouts is a strength, as are the size-matched ablations that help isolate data quality from quantity.

major comments (2)
  1. [§3.2 and §4.1] §3.2 (forward phase) and §4.1 (groundedness check): the manuscript asserts that scoring against G* provides a leakage-blocking measure of per-step progress, yet provides no equations defining the effectiveness score, no ablation removing the groundedness filter, and no quantitative test (e.g., comparing retained trajectories against a version of G* with file paths or edit locations redacted) showing that the check actually prevents favoritism toward continuations that implicitly recover the privileged p* solution path. This is load-bearing for the central claim that gains are due to improved process supervision rather than selection bias.
  2. [§3.1] §3.1 (reverse phase): the construction of the latent process graph G* from p* is described at a high level (contextual facts and solution milestones) but lacks detail on how milestones are extracted, whether they encode runtime behaviors or file-specific information unique to p*, and how the shortest-effective-segment selection interacts with the bi-objective weighting. Without these specifics, it is difficult to verify that the retained trajectories are generally effective rather than tuned to the reference fix.
minor comments (2)
  1. [Table 2, Figure 3] Table 2 and Figure 3: error bars or statistical significance tests are not reported for the Pass@1 deltas; adding them would strengthen the quantitative claims.
  2. [§4.2] Notation: the bi-objective weighting parameter is listed as a free parameter in the abstract but its exact value and sensitivity analysis are not shown in the main results; a short appendix table would clarify reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important areas for clarification and additional validation that will strengthen the manuscript. We respond to each major comment below and will revise the paper accordingly.

read point-by-point responses

  1. Referee: [§3.2 and §4.1] §3.2 (forward phase) and §4.1 (groundedness check): the manuscript asserts that scoring against G* provides a leakage-blocking measure of per-step progress, yet provides no equations defining the effectiveness score, no ablation removing the groundedness filter, and no quantitative test (e.g., comparing retained trajectories against a version of G* with file paths or edit locations redacted) showing that the check actually prevents favoritism toward continuations that implicitly recover the privileged p* solution path. This is load-bearing for the central claim that gains are due to improved process supervision rather than selection bias.

    Authors: We agree that explicit equations and further empirical checks would better substantiate the leakage-blocking claim. In the revision we will add the formal definition of the effectiveness score (milestone coverage under the groundedness predicate that rejects direct references to p* elements) and include an ablation that removes the groundedness filter. We will also add a quantitative test on a held-out subset that compares selection behavior when file paths and edit locations are redacted from G*; preliminary internal checks suggest the filter continues to select non-reference paths, but we will report the full results. These changes directly address the load-bearing concern while preserving the separation between reverse distillation and forward curation on blinded rollouts. revision: yes

  2. Referee: [§3.1] §3.1 (reverse phase): the construction of the latent process graph G* from p* is described at a high level (contextual facts and solution milestones) but lacks detail on how milestones are extracted, whether they encode runtime behaviors or file-specific information unique to p*, and how the shortest-effective-segment selection interacts with the bi-objective weighting. Without these specifics, it is difficult to verify that the retained trajectories are generally effective rather than tuned to the reference fix.

    Authors: We accept that the reverse-phase description requires more operational detail for reproducibility. The revised §3.1 will specify that milestones are extracted by parsing unified diffs for changed files and functions together with passing test assertions from the issue metadata; they encode both file-specific edits and expected runtime outcomes without storing the full patch content. The bi-objective step retains the shortest prefix that reaches at least 80 % milestone coverage on the Pareto front of effectiveness versus length. We will add pseudocode and a worked example to illustrate the interaction and to show that the retained segments generalize beyond the reference p*. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses independent privileged data for curation

full rationale

The paper's central derivation constructs G* from independent developer-authored reference patches p* (external to the teacher model), then applies a forward-phase scoring and bi-objective selection to curate trajectories for SFT. This process does not reduce any claimed prediction or result to its own inputs by construction: the effectiveness scores and length minimization operate on blinded continuations, and final Pass@1 gains are measured on separate SWE-bench benchmarks that do not supply p* at inference time. No self-definitional equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the described chain. The method is self-contained against external benchmarks and does not invoke uniqueness theorems or ansatzes from prior author work to force the outcome.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the existence of a developer-authored reference patch for each training instance and on the assumption that a latent graph extracted from it can serve as an unbiased progress oracle. No explicit free parameters are named, but the bi-objective weighting between effectiveness and length is necessarily chosen. The latent process graph G* is an invented entity whose construction details are not provided.

free parameters (1)
  • bi-objective weighting between effectiveness and length
    The paper formulates trajectory construction as bi-objective optimization; the relative weight or Pareto selection rule must be chosen and is not stated as derived from first principles.
axioms (1)
  • domain assumption Reference patches p* encode the file paths, runtime behaviors, and coding conventions presupposed by the correct fix without introducing selection bias when used for curation.
    Invoked in the abstract when stating that most real issues include a developer-authored reference patch that standard pipelines discard.
invented entities (1)
  • latent process graph G* no independent evidence
    purpose: Distills p* into contextual facts and solution milestones to score per-step progress in blinded teacher continuations.
    Introduced as the output of the reverse phase; no independent evidence outside the curation pipeline is provided.

pith-pipeline@v0.9.0 · 5896 in / 1605 out tokens · 30759 ms · 2026-05-22T05:09:28.428422+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 12 internal anchors

  1. [1]

    M., Nnorom, E., Uddin, G., and Wang, S

    Aleithan, R., Xue, H., Mohajer, M. M., Nnorom, E., Uddin, G., and Wang, S. SWE-bench+: Enhanced coding benchmark for LLMs.arXiv preprint arXiv:2410.06992, 2024

    work page arXiv 2024

  2. [2]

    SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?

    Deng, X., Da, J., Pan, E., He, Y . Y ., Ide, C., Garg, K., Lauffer, N., Park, A., Pasari, N., Rane, C., et al. Swe-bench pro: Can ai agents solve long-horizon software engineering tasks?arXiv preprint arXiv:2509.16941, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Hui, B., Yang, J., Cui, Z., Yang, J., Liu, D., Zhang, L., Liu, T., Zhang, J., Yu, B., Lu, K., et al. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    R2E-gym: Procedural environments and hybrid verifiers for scaling open-weights SWE agents.arXiv preprint arXiv:2504.07164, 2025

    Jain, N., Singh, J., Shetty, M., Zheng, L., Sen, K., and Stoica, I. R2E-gym: Procedural environments and hybrid verifiers for scaling open-weights SWE agents.arXiv preprint arXiv:2504.07164, 2025

    work page arXiv 2025

  5. [5]

    E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., and Narasimhan, K

    Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., and Narasimhan, K. R. Swe- bench: Can language models resolve real-world github issues? InThe Twelfth International Conference on Learning Representations

    work page

  6. [6]

    A., Wutschitz, L., Chen, Y ., Sim, R., and Rajmohan, S

    Kang, M., Chen, W.-N., Han, D., Inan, H. A., Wutschitz, L., Chen, Y ., Sim, R., and Rajmohan, S. ACON: Optimizing context compression for long-horizon LLM agents.arXiv preprint arXiv:2510.00615, 2025

    work page arXiv 2025

  7. [7]

    Let’s verify step by step

    Lightman, H., Kosaraju, V ., Burda, Y ., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobbe, K. Let’s verify step by step. InThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  8. [8]

    Springer, 1999

    Miettinen, K.Nonlinear Multiobjective Optimization, volume 12 ofInternational Series in Operations Research & Management Science. Springer, 1999

    work page 1999

  9. [9]

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., and Lowe, R. Training language models to follow instructions with human feedback. InAdvances in Neural Information Process...

    work page 2022

  10. [10]

    Training Software Engineering Agents and Verifiers with SWE-Gym

    Pan, J., Wang, X., Neubig, G., Jaitly, N., Ji, H., Suhr, A., and Zhang, Y . Training software engineering agents and verifiers with swe-gym.arXiv preprint arXiv:2412.21139, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    YaRN: Efficient Context Window Extension of Large Language Models

    Peng, B., Quesnelle, J., Fan, H., and Shippole, E. Yarn: Efficient context window extension of large language models.arXiv preprint arXiv:2309.00071, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  12. [12]

    Qi, Z., Long, F., Achour, S., and Rinard, M. C. An analysis of patch plausibility and correctness for generate-and-validate patch generation systems. InProceedings of the 24th International Symposium on Software Testing and Analysis, pp. 24–36. ACM, 2015. doi: 10.1145/2771783. 2771791

    work page doi:10.1145/2771783 2015

  13. [13]

    A reduction of imitation learning and structured prediction to no-regret online learning

    Ross, S., Gordon, G., and Bagnell, D. A reduction of imitation learning and structured prediction to no-regret online learning. InProceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 ofProceedings of Machine Learning Research, pp. 627–635. PMLR, 2011. 10

    work page 2011

  14. [14]

    K., Barr, E

    Smith, E. K., Barr, E. T., Le Goues, C., and Brun, Y . Is the cure worse than the disease? overfitting in automated program repair. InProceedings of the 10th Joint Meeting of the European Software Engineering Conference and ACM SIGSOFT Symposium on the Foundations of Software Engineering, pp. 532–543. ACM, 2015. doi: 10.1145/2786805.2786825

    work page doi:10.1145/2786805.2786825 2015

  15. [15]

    Swe-lego: Pushing the limits of supervised fine-tuning for software issue resolving.arXiv preprint arXiv:2601.01426, 2026

    Tao, C., Chen, J., Jiang, Y ., Kou, K., Wang, S., Wang, R., Li, X., Yang, S., Du, Y ., Dai, J., et al. Swe-lego: Pushing the limits of supervised fine-tuning for software issue resolving.arXiv preprint arXiv:2601.01426, 2026

    work page arXiv 2026

  16. [16]

    and Vashist, A

    Vapnik, V . and Vashist, A. A new learning paradigm: Learning using privileged information. Neural Networks, 22(5–6):544–557, 2009. doi: 10.1016/j.neunet.2009.06.042

    work page doi:10.1016/j.neunet.2009.06.042 2009

  17. [18]

    OpenHands: An Open Platform for AI Software Developers as Generalist Agents

    Wang, X., Li, B., Song, Y ., Xu, F. F., Tang, X., Zhuge, M., Pan, J., Song, Y ., Li, B., Singh, J., Tran, H. H., Li, F., Ma, R., Zheng, M., Qian, B., Shao, Y ., Muennighoff, N., Zhang, Y ., Hui, B., Lin, J., Brennan, R., Peng, H., Ji, H., and Neubig, G. OpenHands: An open platform for AI software developers as generalist agents.arXiv preprint arXiv:2407.1...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [19]

    solved issues

    Wang, Y ., Pradel, M., and Liu, Z. Are" solved issues" in swe-bench really solved correctly? an empirical study.arXiv preprint arXiv:2503.15223, 2025

    work page arXiv 2025

  19. [20]

    Agentless: Demystifying LLM-based Software Engineering Agents

    Xia, C. S., Deng, Y ., Dunn, S., and Zhang, L. Agentless: Demystifying LLM-based software engineering agents.arXiv preprint arXiv:2407.01489, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [21]

    Reducing cost of LLM agents with trajectory reduction

    Xiao, Y .-A., Gao, P., Peng, C., and Xiong, Y . Reducing cost of LLM agents with trajectory reduction. InProceedings of the ACM International Conference on the Foundations of Software Engineering (FSE), 2026. doi: 10.1145/3797084. arXiv:2509.23586

    work page doi:10.1145/3797084 2026

  21. [22]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [23]

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

    Yang, J., Jimenez, C. E., Wettig, A., Lieret, K., Yao, S., Narasimhan, K. R., and Press, O. SWE-agent: Agent-computer interfaces enable automated software engineering. InAdvances in Neural Information Processing Systems, 2024. URL https://arxiv.org/abs/2405.15793

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [24]

    SWE-smith: Scaling Data for Software Engineering Agents

    Yang, J., Lieret, K., Jimenez, C. E., Wettig, A., Khandpur, K., Zhang, Y ., Hui, B., Press, O., Schmidt, L., and Yang, D. SWE-smith: Scaling data for software engineering agents.arXiv preprint arXiv:2504.21798, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [25]

    ReAct: Synergizing Reasoning and Acting in Language Models

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y . ReAct: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [26]

    Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving

    Zan, D., Huang, Z., Liu, W., Chen, H., Zhang, L., Xin, S., Chen, L., Liu, Q., Zhong, X., Li, A., et al. Multi-swe-bench: A multilingual benchmark for issue resolving.arXiv preprint arXiv:2504.02605, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [27]

    GLM-5: from Vibe Coding to Agentic Engineering

    Zeng, A., Lv, X., Hou, Z., Du, Z., Zheng, Q., Chen, B., Yin, D., Ge, C., Huang, C., Xie, C., et al. Glm-5: from vibe coding to agentic engineering.arXiv preprint arXiv:2602.15763, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  27. [28]

    Y ., Yan, R., Wei, T., He, J., Song, X., et al

    Zeng, L., Li, Y ., Xiao, Y ., Li, C., Liu, C. Y ., Yan, R., Wei, T., He, J., Song, X., et al. Skywork-SWE: Unveiling data scaling laws for software engineering in LLMs.arXiv preprint arXiv:2506.19290, 2025

    work page arXiv 2025

  28. [29]

    Swe-bench goes live!arXiv preprint arXiv:2505.23419, 2025

    Zhang, L., He, S., Zhang, C., Kang, Y ., Li, B., Xie, C., Wang, J., Wang, M., Huang, Y ., Fu, S., et al. Swe-bench goes live!arXiv preprint arXiv:2505.23419, 2025

    work page arXiv 2025

  29. [30]

    AutoCodeRover: Autonomous program improvement

    Zhang, Y ., Ruan, H., Fan, Z., and Roychoudhury, A. AutoCodeRover: Autonomous program improvement. InProceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, 2024

    work page 2024

  30. [31]

    the patch adds X

    Zhao, Y ., Huang, J., Hu, J., Wang, X., Mao, Y ., Zhang, D., Jiang, Z., Wu, Z., Ai, B., Wang, A., Zhou, W., and Chen, Y . Swift:a scalable lightweight infrastructure for fine-tuning, 2024. URL https://arxiv.org/abs/2408.05517. 11 Appendix A Technical appendices and supplementary material 13 B Prerequisite-graph node distribution 13 C Value of information ...

    work page arXiv 2024

  31. [32]

    A: Now let me check the model implementation

    work page

  32. [33]

    O: grep describe_security_group_rules →models:545

    work page

  33. [34]

    O: view models/security_groups.py 540-560→f8,f9

    work page

  34. [35]

    A: I see! filters=None passed to backend

    work page

  35. [36]

    O: re-view responses/security_groups.py 195-205

    work page

  36. [37]

    A: Look at examples of filter parsing

    work page

  37. [38]

    O: grep Filter (30K chars of noise)

    work page

  38. [39]

    A: Trace SecurityRule for skipped filtering

    work page

  39. [40]

    P2T trajectory (with curation)

    O: view models/.../SecurityRule.__init__ →wrong path ↓ commits to BACKEND fix; next 100 steps add group_id to SecurityRule, edit response template, BREAK test_create_and_describe_security_grp_rule. P2T trajectory (with curation)

    work page

  40. [41]

    A: Examine responses/security_groups.py

    work page

  41. [42]

    O: view responses/security_groups.py 190-210→f2

    work page

  42. [43]

    A:Let me check the backend model implementation(blinded)

    work page

  43. [44]

    A: Check sibling describe_security_groups parses ←y ′ (curated)

    work page

  44. [45]

    O: grep→line 186 filters=_filters_from_querystring() →f3

    work page

  45. [46]

    A: _filters_from_querystring uses _get_multi_param

    work page

  46. [47]

    O: view _base_response.py 1-31→f4

    work page

  47. [48]

    A: Is this the standard EC2 pattern across responses?

    work page

  48. [49]

    O: grep -rn _filters_from_querystring responses/→f10

    work page

  49. [50]

    A: Plan: replace _get_param(’Filter’) with helper at line 197

    work page

  50. [51]

    Figure 7: Side-by-side fragment of the same window in πblind (left) and P2T (right)

    O:(commits, replans from new prefix) ↓ one-line edit at responses/security_groups.py:197; both validations pass. Figure 7: Side-by-side fragment of the same window in πblind (left) and P2T (right). Rows labelled A are assistant responses; rows labelled O are environment observations. The single curator-authored step (red, y′ at j=2) replaces an empty assi...

    work page