Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents

Alexander Golubev; Anton Shevtsov; Boris Yangel; Daria Litvintseva; Filipp Fisin; Ibragim Badertdinov; Karina Zainullina; Maksim Nekrashevich; Maria Trofimova; Sergei Polezhaev

REVIEW 4 major objections 4 minor 1 cited by

Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Two guided search strategies that never rewind the environment double a fine-tuned Qwen-72B software agent's average success rate on SWE-bench Verified, reaching 40.8%.

desk verdict Solid, honestly reported search methods on SWE agents, but the SOTA headline doesn't survive contact with the retry protocol. read the letter →

arxiv 2505.13652 v1 pith:CA3AQEHA submitted 2025-05-19 cs.SE cs.CL

Karina Zainullina , Alexander Golubev , Maria Trofimova , Sergei Polezhaev , Ibragim Badertdinov , Daria Litvintseva , Simon Karasik , Filipp Fisin

show 4 more authors

Sergei Skvortsov Maksim Nekrashevich Anton Shevtsov Boris Yangel

This is my paper · ORCID

classification cs.SEcs.CL

keywords test-timesearchnon-serializableenvironmentsaction-valuecritic1-steplookaheadtrajectoryselectionSWE-benchVerifiedsoftwareengineeringagentsLLM

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large language models often nail a coding task on some attempts and fail on others, and in agentic settings like Docker containers the environment cannot be rewound, so heavy search methods such as Monte-Carlo tree search are unavailable. The paper claims that two forward-only search strategies, guided by a learned action-value function, can nonetheless close much of this average-versus-best gap. On the SWE-bench Verified benchmark, combining one-step lookahead with trajectory selection doubles the success rate of a fine-tuned Qwen-72B policy, to 40.8%, which the authors report as the best result among open-weight systems at the time. The same techniques also lift a GPT-4o-based agent by a similar factor, even though the critic was never trained on GPT-4o trajectories. A sympathetic reader would take away that reliable agent performance can be bought with test-time compute alone, without needing environment checkpointing.

What carries the argument

The load-bearing object is the critic, a LLaMA-3.1-70B model repurposed to output a scalar action-value prediction for each agent turn, trained with L2 loss on temporal-difference targets that interpolate between Monte-Carlo and one-step TD estimates (the paper settles on λ=0.7). Around it sit two inference operators that only move forward through the transition function: sample-based one-step lookahead, which draws K actions from the base policy, scores them, and executes the argmax; and trajectory selection, which runs N complete trajectories and returns the one whose terminating action scores highest. Because both operators only ever advance the environment state, they avoid the serialization and branching that MCTS would require.

What would settle it

Run the identical base policy on Verified-50 with the same compute budget, but replace the critic's argmax with a random action chosen from the K candidates, and replace trajectory selection with a uniformly random trajectory from the N rollouts; if the success rate does not drop below the critic-guided rate, the reported gains come from extra sampling rather than from the learned value estimates.

Watch

Extended reading notes

Core claim

The paper's central claim is that in non-serializable environments, where intermediate states cannot be saved, copied, or reverted, guided test-time search is still possible provided the search never branches backward: at each step the agent samples K candidate actions, scores them with a learned critic estimating $Q(s,a)$, and commits forward to the best-scoring action; after N full rollouts it scores the final action of each and keeps the highest-valued trajectory. With a critic trained via TD(λ) on 80,000 bootstrapped trajectories, this combination takes a fine-tuned Qwen2.5-72B policy from a 16.2% default single-run success rate on the curated Verified-50 subset to 41.7%, and to 40.8% on the full SWE-bench Verified set with K=8 and N=15, which the paper reports as the new state-of-the-art for open-weights models. The same recipe improves a GPT-4o-based agent from 22.0% to 40.0% on Verified-50, showing transfer across policies despite the critic never having seen GPT-4o trajectories.

Load-bearing premise

The critic's predicted scores must rank candidate actions and whole trajectories by true success probability better than chance; if the critic mis-ranks, the search would be as good as random sampling while costing much more compute.

Editorial extensions

If this is right

With enough test-time compute, the open-weights Qwen-72B agent reaches 40.8% on SWE-bench Verified, the highest reported success rate among open-weight systems at the time.
Both search methods scale with compute: success rate keeps rising with more trajectory candidates N, and lookahead improves up to K=8 candidates, where the current critic hits its discriminative ceiling.
The techniques transfer to closed models: GPT-4o's success rate on Verified-50 roughly doubles from 22.0% to 40.0% when the two methods are combined, despite the critic never being trained on GPT-4o trajectories.
Critic target choice matters: TD(λ) with λ=0.7 outperforms both pure Monte-Carlo and one-step TD for lookahead, and larger critic base models (70B over 8B) guide noticeably better.
Search and the 'until submitted' retry regime are complementary: the relative gains of guided search hold both when the agent must finish on first completion and when retries are allowed, so the improvements are not just a retry artifact.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same forward-only recipe should apply to other non-serializable agentic settings, such as browser automation, device control, or running live services, where rollback is impossible; the only requirement is a critic that can score the current state plus a candidate action.
The paper notes value hacking, in which the critic's scores drift upward on unsuccessful trajectories once search follows them; this suggests iteratively retraining the critic on adversarial search-found trajectories could push the method well past 40.8%.
If the critic is available at inference time, its final-action score doubles as a self-verifier, so agent systems could use it to decide when to submit without external validation, with both safety and cost implications.
Since trajectory selection with N=15 already approaches pass@15 ceilings, combining guided search with replay-based serialization, as the authors float, might capture the remaining gap, but only in environments where replay is trustworthy.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is worth reading for its clean setup and honest reporting: two simple forward-only search operators, one-step lookahead and trajectory selection, guided by a learned critic, are evaluated with ten seeds on a 50-problem subset. The methods do appear to improve success rates, and the Verified-50 numbers with SEMs are credible and consistent. The paper also does good service by framing non-serializable environments and explaining why MCTS-style rollback is unavailable in Docker-like settings. The critic analysis is useful, and the authors are unusually candid about value hacking, weaker transfer to GPT-4o, and unequal hyperparameter tuning in the appendices. Credit where earned.

The main problem is the headline: 40.8% on SWE-bench Verified is presented as state-of-the-art for open-weights models, but Table 2 compares against baselines that almost certainly use the standard single-trajectory protocol, while the paper's own until-submitted retry loop is on by default everywhere. This is not a minor footnote; Table 1 shows the retry loop alone lifts the base policy from 16.2% to 22.8% and one-step lookahead from 26.8% to 32.4%. Retries are a generic compute multiplier available to any baseline, so the claimed SOTA cannot be attributed to guided search until the protocols are matched. Even a repeated full-set run would not fix the comparison. Relatedly, the 'doubling' claim in the abstract mixes a default-regime base rate on Verified-50 (16.2%) with an until-submitted full-set number (40.8%). Apples and oranges.

There are additional soft spots. The main hyperparameters (lambda, gamma, K, T, N) were selected on Verified-50, the same set used for the tuned results, so the gains are partly tuned on the evaluation subset. The 40.8% itself is a single full-set run with no reported full-set baseline. The search operators are classical and the paper acknowledges this, so the novelty is modest. The math is correct but textbook. I also note that the reader's original critique missed the protocol mismatch; that concern is stronger and more load-bearing than the single-run issue, though both are real.

Who should read this: anyone working on LLM-agent test-time compute or SWE-bench evaluation. The Verified-50 findings are a useful empirical data point, and the honest discussion of critic limitations is worth absorbing. I would not cite the 40.8% SOTA claim, but I would probably cite the paper for the Verified-50 results and the non-serializable framing.

Recommendation: send to peer review. A serious referee should require matched evaluation protocols, repeated full-set runs, and error bars before the SOTA claim is accepted, but the empirical core is solid enough to deserve that scrutiny.

Referee Report

4 major / 4 minor

Summary. This paper formalizes non-serializable RL environments and studies two guided test-time search strategies for LLM-based software engineering agents: sample-based 1-step lookahead (Algorithm 1) and trajectory selection (Algorithm 2), both guided by a learned critic model approximating the action-value function. Experiments use SWE-agent on SWE-bench Verified. On a 50-instance subset (Verified-50), the methods improve success rate over a fine-tuned Qwen-based policy, reaching 41.7% in the default protocol and 44.07% under the "until submitted" retry protocol when combined with N=10 trajectory candidates. On the full set, a single configuration (1-step lookahead K=8 + trajectory selection N=15) is reported at 40.8%, claimed as a new state-of-the-art for open-weights models. Transfer to GPT-4o is also presented.

Significance. If the headline 40.8% result were measured under a protocol matched to the baselines, the contribution would be useful and timely: a simple critic-guided search that avoids state serialization, with repeated-seed evaluation on Verified-50 and an honest discussion of critic limitations. The formalization of non-serializable environments and the identification of two complementary strategies are valuable. The repeated-seed Verified-50 results with SEM are a strength, as is the explicit acknowledgment of value hacking and weaker mid-trajectory discrimination on GPT-4o transfer. However, the state-of-the-art claim is not currently established because the protocol used for the 40.8% row differs from the protocol of the cited baselines, and the full-set run lacks a same-protocol baseline and measure of uncertainty. The central ideas are defensible, so the issues are fixable within the scope of a major revision.

major comments (4)

[Section 4.1, Table 1, Table 2] The "until submitted" retry protocol is stated in Section 4.1 to be the default "in all experiments, if not stated otherwise," yet Table 2 does not state the protocol for the 40.8% row. Table 1 quantifies the effect of this protocol alone: the Qwen-based policy rises from 16.2% (default) to 22.8% (until submitted), and the 1-step lookahead row rises from 26.8% to 32.4%. The Table 2 baselines (SWE-Gym 32.0, SWE-Fixer 30.2, Lingma 25.0) are reported under the standard single-trajectory SWE-bench protocol, so the claimed state-of-the-art compares an augmented retry protocol with non-retry baselines. The "doubling" claim in the abstract also mixes the Verified-50 default-regime base rate (16.2%) with the full-set until-submitted value (40.8%). Please rerun or report the full-set results and the baselines under matched protocols, or explicitly restrict the claims to a protocol-compatible comparison.
[Section 4.2, Table 2] The headline 40.8% is a single run on the full SWE-bench Verified set with no full-set baseline. The manuscript does not report the Qwen-based base policy on the full set under the same protocol, nor does it report any measure of uncertainty for the 40.8% result. Because Verified-50 is only a 50-instance estimate, the gap from 16.2% or 22.8% to 40.8% is not a controlled comparison. Please provide repeated full-set runs, or at least a full-set base-policy run in the same regime, and report the standard error or confidence interval.
[Section 4.2 vs Subsection 4.4] There is a direct internal inconsistency in the search hyperparameters. Section 4.2 states that the optimal parameters identified in Subsection 4.4 are T=0.9 and K=4, but Table 2's headline uses K=8, and Figure 8 shows that success rate saturates at K=8 rather than K=4. Furthermore, Figure 8 is obtained with a TD(0.8) critic, whereas the main results use a TD(0.7) critic. Please clarify which configuration was used for the headline result and justify why the K/T sweep applies to the critic actually used in the main experiments.
[Section 4.3 and Appendix C, Figure 11] The paper acknowledges that "1-step lookahead causes some amount of value hacking" and Figure 11 shows weaker mid-trajectory discrimination for the GPT-4o policy. Since both search operators select candidates by argmax over critic scores, the core assumption is that the critic's ranking remains valid on the search-induced distribution, not just on the base-policy trajectories used to train it. The manuscript does not provide a quantitative analysis of rank accuracy on trajectories produced by the combined search, such as the correlation between critic scores and terminal success on that distribution. Please add such an analysis or explicitly state the consequent limitation on the strength of the empirical claims.

minor comments (4)

[Table 1] The GPT-4o trajectory-selection rows (N=5) report no SEM; please state explicitly whether these are single-run results and avoid drawing statistical comparisons from them.
[Abstract and Section 4.2] The phrase "double the average success rate" is not supported by the until-submitted rows in Table 1, where the combined method gives 44.07% versus 22.8% for the base policy (a factor of 1.93). The doubling claim should be tied to the default-regime comparison or reworded.
[Section 4.3, Figures 4 and 5] The λ and γ hyperparameter sweeps are presented without error bars or a statement of the number of seeds; please specify the evaluation protocol and uncertainty for these curves.
[Section 4.1, Table 1] The SEMs for the combined methods (e.g., 44.07±0.05) are much smaller than the SEM of the base policy (22.8±1.05); please explain whether these are computed over the same 10 seeds and why the variance is so much lower for the combined method.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the guided-search results are independently evaluated against external SWE-bench Verified outcomes; disclosed protocol choices are fairness concerns, not derivation loops.

full rationale

The paper's derivation chain is self-contained and does not reduce its predictions to its inputs by construction. The guided-search operators in Algorithms 1 and 2 are defined independently of the measured outcomes: 1-step lookahead selects actions by a learned action-value estimate, and trajectory selection ranks complete trajectories by the same estimate. The critic is trained on separately collected bootstrapped trajectories with TD(lambda) targets (Section 4.1, Appendix B), not on test-set outcomes, and the headline 40.8% is a full SWE-bench Verified evaluation of a policy and critic that were not fitted to the benchmark labels. Hyperparameters such as lambda, K, T, and N are selected on the disclosed Verified-50 subset and then applied to the full set, which is standard tuning rather than a circular prediction. The until-submitted retry protocol is explicitly disclosed in Section 4.1 and Appendix E as the default, and the paper separately reports default-regime numbers in Table 1; the comparison of a retry-assisted run against external baselines in Table 2 is a legitimate evaluation-fairness concern, not a tautology. The self-citation to Badertdinov et al. (2024) is used only for training-issue collection methodology and is not load-bearing for the claimed result. The paper even acknowledges limitations of the critic, including value hacking and weaker mid-trajectory discrimination for GPT-4o in Appendix C, which further indicates the evaluation is not rigged by construction. No equation in the paper is identical to its input, no fitted parameter is renamed as a prediction, and no load-bearing uniqueness claim is imported from the authors' prior work. Therefore the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central result rests on the standard POMDP formalism, a domain assumption about Docker non-serializability, and the empirical adequacy of the learned critic. No new physical or mathematical entities are introduced. The main free parameters are search hyperparameters (λ, γ, K, T, N) and the retry cap, all selected on the evaluation subset Verified-50.

free parameters (6)

TD(λ) mixing parameter λ = 0.7
Selected via a sweep on Verified-50 (Figure 4), then used for the headline results on the same set.
discount factor γ = near 1 (exact value for final runs not stated)
Figure 5 shows SR increasing as γ approaches 1; the final critic uses TD(0.7) but the γ used in reported runs is not explicitly given.
action candidates K = 4 for Verified-50 main results; 8 for the full-set 40.8% run
Tuned on Verified-50 (Figure 8) to balance exploration and cost.
sampling temperature T = 0.9
Tuned on Verified-50 (Figure 8).
trajectory candidates N = 5 and 10 on Verified-50; 15 for the full-set run
Tuned on Verified-50; Figure 3 shows no plateau in the investigated range.
max retries in until-submitted regime = 10
Hand-chosen cap in Section 4.1; this regime is itself a form of search that boosts all methods including the baseline.

assumptions (5)

standard math POMDP formulation with sparse terminal reward r_{T-1} in {0,1}
Used to formalize the setting in Section 2.1; a standard model, not a point of dispute.
domain assumption Docker container states cannot be reliably serialized or replayed
Section 2.3 argues this from CRIU limitations and stochastic transitions; it motivates the methods and blocks MCTS. If replay were reliable, the practical contribution would shrink.
domain assumption The critic's special-token scalar output approximates Q(s,a) for the policy in question
Described in Section 4.1 and Appendix B. The method's gains depend on this approximation being adequate for both training and held-out policies.
domain assumption Verified-50 is an unbiased estimator of full SWE-bench Verified
A 50-problem random subset (Appendix D) is used to tune hyperparameters and to extrapolate baseline gains to the full set.
domain assumption The 'until submitted' retry loop is a fair comparison protocol
Section 4.1 adds up to 10 retries until 'submit'; it improves all methods and the baseline, but is itself a test-time search method, so it may affect the relative comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents." pith.science (2026). https://pith.science/paper/CA3AQEHA

@misc{pith2026250513652,
  author       = {Pith},
  title        = {Pith review of: Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CA3AQEHA}},
  note         = {Machine review of arXiv:2505.13652}
}

read the original abstract

Large language models (LLMs) have recently achieved remarkable results in complex multi-step tasks, such as mathematical reasoning and agentic software engineering. However, they often struggle to maintain consistent performance across multiple solution attempts. One effective approach to narrow the gap between average-case and best-case performance is guided test-time search, which explores multiple solution paths to identify the most promising one. Unfortunately, effective search techniques (e.g. MCTS) are often unsuitable for non-serializable RL environments, such as Docker containers, where intermediate environment states cannot be easily saved and restored. We investigate two complementary search strategies applicable to such environments: 1-step lookahead and trajectory selection, both guided by a learned action-value function estimator. On the SWE-bench Verified benchmark, a key testbed for agentic software engineering, we find these methods to double the average success rate of a fine-tuned Qwen-72B model, achieving 40.8%, the new state-of-the-art for open-weights models. Additionally, we show that these techniques are transferable to more advanced closed models, yielding similar improvements with GPT-4o.

Figures

Figures reproduced from arXiv: 2505.13652 by the authors.

**Figure 1.** The comparison of two evaluation protocols for a GPT4o-based agent: Pass@N and random sampling. The x-axis shows the number of attempts, the y-axis shows the average success rate (i.e. the percentage of correct solutions). formance often falls short of their demonstrated potential. To illustrate this challenge, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Per-instance success rate of Qwen-based policy computed over 15 runs [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** The dependency between success rate and the number of candidates N in trajectory selection. This bootstrapping process produced a total of 80,000 positive and negative trajectories. We use these trajectories to train critic models. Since our dataset contains trajectories produced by multiple policies, we incorporate policy identifiers into the critic’s system prompt to condition the predicted action-value on the p… view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: The dependency between 1-step lookahead SR and λ. 0.86 0.88 0.90 0.92 0.94 0.96 0.98 1.00 25 26 27 28 29 30 31 32 Success rate, % [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: The dependency between success rate and γ for MC. 0.86 0.88 0.90 0.92 0.94 0.96 0.98 1.00 10 12 14 16 18 20 22 Trajectory length Unsuccessful All Successful [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: The dependency between trajectory length and γ for MC. evaluation uses a critic model trained with the best parameters identified in Subsection 4.3, i.e. TD(0.7), for both 1-step lookahead and trajectory selection. For 1-step lookahead, we utilize the optimal search …

**Figure 7.** Figure 7: Critic learns to distinguish successfull and unsucessfull trajectories produced by Qwen-based policy. head and trajectory selection with N = 15, T = 0.9, and K = 8 on the full SWE-bench Verified set, where it achieves 40.8% success rate ( [PITH_FULL_IMAGE:figures/full…

**Figure 9.** Figure 9: 1-step lookahead improves success rate per instance for GPT-4o base policy, adding new solved issues [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 12.** Figure 12: The average number of retries it takes to generate a trajectory that ends with “submit” for a given fraction of the test set using the Qwen-based policy [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

**Figure 13.** Figure 13: The effects of varying training dataset size for critic on performance of Qwen-based policy + 1-step lookahead, Verified-50, default regime. for LLaMA3.1-70B. We recommend targeted hyperparameter tuning for the specific model family used to maximize performance of the…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

One Tool Is Enough: Reinforcement Learning for Repository-Level LLM Agents
cs.SE 2025-12 conditional novelty 6.0 of 10

Repository-level issue localization can be done by a single jump-to-definition tool trained with reinforcement learning, achieving strong results on SWE-bench despite using only open-weights models.

Reference graph

Works this paper leans on

12 extracted references · 2 canonical work pages · cited by 1 Pith paper

[4]

Luo, L., Lin, Z., Liu, Y ., Shu, L., Zhu, Y ., Shang, J., and Meng, L

URL https://openreview.net/forum ?id=v8L0pN6EOi. Luo, L., Lin, Z., Liu, Y ., Shu, L., Zhu, Y ., Shang, J., and Meng, L. Critique ability of large language models. CoRR, abs/2310.04815, 2023. doi: 10.48550/ARXIV.2 310.04815. URL https://doi.org/10.48550 /arXiv.2310.04815. 10 Guided Search Strategies in Non-Serializable Environments with Applications to Sof...

arXiv doi:10.48550/arxiv.2310.04815 2023
[5]

URL https: //doi.org/10.48550/arXiv.2411.00622

doi: 10.48550/ARXIV.2411.00622. URL https: //doi.org/10.48550/arXiv.2411.00622. Mudgal, S., Lee, J., Ganapathy, H., Li, Y ., Wang, T., Huang, Y ., Chen, Z., Cheng, H., Collins, M., Strohman, T., Chen, J., Beutel, A., and Beirami, A. Controlled decoding from language models. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Aus...

arXiv doi:10.48550/arxiv.2411.00622
[7]

URL https: //doi.org/10.48550/arXiv.2408.07199

doi: 10.48550/ARXIV.2408.07199. URL https: //doi.org/10.48550/arXiv.2408.07199. Schluntz, E., Biggs, S., Drain, D., Christiansen, E., Kravec, S., Rosso, F., DasSarma, N., and Chandrasekara, V . Rais- ing the bar on swe-bench verified with claude 3.5 sonnet,

arXiv doi:10.48550/arxiv.2408.07199
[8]

Setlur, A., Nagpal, C., Fisch, A., Geng, X., Eisenstein, J., Agarwal, R., Agarwal, A., Berant, J., and Kumar, A

URL https://www.anthropic.com/re search/swe-bench-sonnet. Setlur, A., Nagpal, C., Fisch, A., Geng, X., Eisenstein, J., Agarwal, R., Agarwal, A., Berant, J., and Kumar, A. Rewarding progress: Scaling automated process verifiers for LLM reasoning.CoRR, abs/2410.08146, 2024. doi: 10.48550/ARXIV.2410.08146. URL https://doi. org/10.48550/arXiv.2410.08146. Silv...

arXiv doi:10.48550/arxiv.2410.08146
[12]

URL https://doi.org/10.18653/v1/2024 .findings-naacl.55

doi: 10.18653/V1/2024.FINDINGS-NAACL.55. URL https://doi.org/10.18653/v1/2024 .findings-naacl.55. Zhang, Y ., Ruan, H., Fan, Z., and Roychoudhury, A. Au- tocoderover: Autonomous program improvement. In Christakis, M. and Pradel, M. (eds.),Proceedings of the 33rd ACM SIGSOFT International Symposium on Soft- ware Testing and Analysis, ISSTA 2024, Vienna, Au...

arXiv 2024
[265]

Pan, J., Wang, X., Neubig, G., Jaitly, N., Ji, H., Suhr, A., and Zhang, Y

URL https://doi.org/10.48550/arXiv .2412.05265. Pan, J., Wang, X., Neubig, G., Jaitly, N., Ji, H., Suhr, A., and Zhang, Y . Training software engineering agents and verifiers with swe-gym.CoRR, abs/2412.21139, 2024. doi: 10.48550/ARXIV.2412.21139. URL https: //doi.org/10.48550/arXiv.2412.21139. Putta, P., Mills, E., Garg, N., Motwani, S., Finn, C., Garg, ...

work page doi:10.48550/arxiv.2412.05265 2024
[451]

Yang, J., Jimenez, C

URL https://doi.org/10.48550/arXiv .2405.00451. 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 engi- neering. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://arxiv.org/abs/2405.15793. Yao, S., Yu, D...

arXiv doi:10.48550/arxiv.2405.00451
[489]

Xie, C., Li, B., Gao, C., Du, H., Lam, W., Zou, D., and Chen, K

URL https://doi.org/10.48550/arXiv .2407.01489. Xie, C., Li, B., Gao, C., Du, H., Lam, W., Zou, D., and Chen, K. Swe-fixer: Training open-source llms for effec- tive and efficient github issue resolution.arXiv preprint arXiv:2501.05040, 2025. Xie, Y ., Goyal, A., Zheng, W., Kan, M., Lillicrap, T. P., Kawaguchi, K., and Shieh, M. Monte carlo tree search bo...

arXiv doi:10.48550/arxiv.2407.01489

Show all 12 references

[707]

Hao, S., Gu, Y ., Ma, H., Hong, J

URL https://doi.org/10.48550/arXiv .2410.01707. Hao, S., Gu, Y ., Ma, H., Hong, J. J., Wang, Z., Wang, D. Z., and Hu, Z. Reasoning with language model is planning with world model. In Bouamor, H., Pino, J., and Bali, K. (eds.),Proceedings of the 2023 Conference on Empirical Me...

doi arXiv
[2017]

Sutton, R

URL http://arxiv.org/abs/1712.018 15. Sutton, R. S. and Barto, A. G.Reinforcement learning - an introduction. Adaptive computation and machine learning. MIT Press, 1998. ISBN 978-0-262-19398-6. URL https://www.worldcat.org/oclc/37 293240. Uesato, J., Kushman, N., Kumar, R., So...

1998 doi arXiv
[2019]

URLhttps://criu.org/. Dash, A. Understanding migration mechanisms of contain- ers using criu. V olume: 09:5, 02 2022. Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., Goyal, A., Hartshorn, A., Yang, A., Mitra...

2022 doi arXiv
[2024]

URL https: //doi.org/10.48550/arXiv.2402.05244

doi: 10.48550/ARXIV.2402.05244. URL https: //doi.org/10.48550/arXiv.2402.05244. Antoniades, A., ¨Orwall, A., Zhang, K., Xie, Y ., Goyal, A., and Wang, W. Y . Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement. CoRR, abs/2410.20285, 2024...

doi arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [4]

Luo, L., Lin, Z., Liu, Y ., Shu, L., Zhu, Y ., Shang, J., and Meng, L

URL https://openreview.net/forum ?id=v8L0pN6EOi. Luo, L., Lin, Z., Liu, Y ., Shu, L., Zhu, Y ., Shang, J., and Meng, L. Critique ability of large language models. CoRR, abs/2310.04815, 2023. doi: 10.48550/ARXIV.2 310.04815. URL https://doi.org/10.48550 /arXiv.2310.04815. 10 Guided Search Strategies in Non-Serializable Environments with Applications to Sof...

arXiv doi:10.48550/arxiv.2310.04815 2023

[2] [5]

URL https: //doi.org/10.48550/arXiv.2411.00622

doi: 10.48550/ARXIV.2411.00622. URL https: //doi.org/10.48550/arXiv.2411.00622. Mudgal, S., Lee, J., Ganapathy, H., Li, Y ., Wang, T., Huang, Y ., Chen, Z., Cheng, H., Collins, M., Strohman, T., Chen, J., Beutel, A., and Beirami, A. Controlled decoding from language models. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Aus...

arXiv doi:10.48550/arxiv.2411.00622

[3] [7]

URL https: //doi.org/10.48550/arXiv.2408.07199

doi: 10.48550/ARXIV.2408.07199. URL https: //doi.org/10.48550/arXiv.2408.07199. Schluntz, E., Biggs, S., Drain, D., Christiansen, E., Kravec, S., Rosso, F., DasSarma, N., and Chandrasekara, V . Rais- ing the bar on swe-bench verified with claude 3.5 sonnet,

arXiv doi:10.48550/arxiv.2408.07199

[4] [8]

Setlur, A., Nagpal, C., Fisch, A., Geng, X., Eisenstein, J., Agarwal, R., Agarwal, A., Berant, J., and Kumar, A

URL https://www.anthropic.com/re search/swe-bench-sonnet. Setlur, A., Nagpal, C., Fisch, A., Geng, X., Eisenstein, J., Agarwal, R., Agarwal, A., Berant, J., and Kumar, A. Rewarding progress: Scaling automated process verifiers for LLM reasoning.CoRR, abs/2410.08146, 2024. doi: 10.48550/ARXIV.2410.08146. URL https://doi. org/10.48550/arXiv.2410.08146. Silv...

arXiv doi:10.48550/arxiv.2410.08146

[5] [12]

URL https://doi.org/10.18653/v1/2024 .findings-naacl.55

doi: 10.18653/V1/2024.FINDINGS-NAACL.55. URL https://doi.org/10.18653/v1/2024 .findings-naacl.55. Zhang, Y ., Ruan, H., Fan, Z., and Roychoudhury, A. Au- tocoderover: Autonomous program improvement. In Christakis, M. and Pradel, M. (eds.),Proceedings of the 33rd ACM SIGSOFT International Symposium on Soft- ware Testing and Analysis, ISSTA 2024, Vienna, Au...

arXiv 2024

[6] [265]

Pan, J., Wang, X., Neubig, G., Jaitly, N., Ji, H., Suhr, A., and Zhang, Y

URL https://doi.org/10.48550/arXiv .2412.05265. Pan, J., Wang, X., Neubig, G., Jaitly, N., Ji, H., Suhr, A., and Zhang, Y . Training software engineering agents and verifiers with swe-gym.CoRR, abs/2412.21139, 2024. doi: 10.48550/ARXIV.2412.21139. URL https: //doi.org/10.48550/arXiv.2412.21139. Putta, P., Mills, E., Garg, N., Motwani, S., Finn, C., Garg, ...

work page doi:10.48550/arxiv.2412.05265 2024

[7] [451]

Yang, J., Jimenez, C

URL https://doi.org/10.48550/arXiv .2405.00451. 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 engi- neering. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://arxiv.org/abs/2405.15793. Yao, S., Yu, D...

arXiv doi:10.48550/arxiv.2405.00451

[8] [489]

Xie, C., Li, B., Gao, C., Du, H., Lam, W., Zou, D., and Chen, K

URL https://doi.org/10.48550/arXiv .2407.01489. Xie, C., Li, B., Gao, C., Du, H., Lam, W., Zou, D., and Chen, K. Swe-fixer: Training open-source llms for effec- tive and efficient github issue resolution.arXiv preprint arXiv:2501.05040, 2025. Xie, Y ., Goyal, A., Zheng, W., Kan, M., Lillicrap, T. P., Kawaguchi, K., and Shieh, M. Monte carlo tree search bo...

arXiv doi:10.48550/arxiv.2407.01489