Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding

Boyin Tan, Haoning Deng, Junyuan Zhang, Junjielong Xu, Pinjia He, Youcheng Sun

TL;DR

The paper studies how maintainers choose among multiple candidate proposals for software issues and introduces SWE-Manager, an 8B reasoning model trained with reinforcement learning to both select the best proposal and synthesize a golden proposal. A manual study on 2,852 GitHub issues reveals four main rationales—Risk and Safety, Fix Depth, Maintainability, and Others—informing model design. SWE-Manager combines supervised fine-tuning with DAPO RL to maximize correct selection and high-quality consolidation, achieving 53.21% accuracy and 57.75% earn rate on SWE-Lancer Manager, outperforming GPT-5. The P2A framework tests end-to-end impact, showing SWE-Manager-8B matching GPT-5 in end-to-end patch success on SWE-Lancer IC and offering unique improvements via golden-proposal guidance. Overall, the work demonstrates proposal consolidation as a practical mechanism to improve both decision quality and downstream implementation in real-world software engineering workflows.

Abstract

Large language model (LLM) research in software engineering has largely focused on tasks such as code generation and bug repair. In practice, teams often draft multiple candidate proposals for fixing an issue and then deliberate on one golden proposal for implementation. This selection requires not only assessing the issue's scope, impact, and urgency, but also a clear understanding of each proposal's strengths and weaknesses. A good selection could make issue resolution more reliable while reducing regression and operational risk, whereas a poor choice can increase risk and even cause unpredictable failures. We first conduct a manual study of real-world issues to characterize the rationales maintainers use when selecting among competing proposals. Motivated by these findings, we introduce SWE-Manager, a joint selection and synthesis approach that selects the best proposal and synthesizes a golden proposal. SWE-Manager is an 8B model trained via reinforcement learning (RL) to compare proposals, justify its choice, and synthesize a golden proposal for implementation. We view proposal selection as a reasoning task, mirroring how technical managers review competing proposals by weighing issue context and each proposal's solution without executing code or running tests. On the SWE-Lancer Manager benchmark, SWE-Manager achieves 53.21 selection accuracy and 57.75 earn rate, earning 152,750 dollars and outperforming strong baselines including GPT-5. To further evaluate the effectiveness of SWE-Manager in real-world issue resolution, we design the P2A framework, which simulates a real-world workflow where multiple proposals are drafted, reviewed, and a golden proposal is selected for implementation ...

SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding

TL;DR

The paper studies how maintainers choose among multiple candidate proposals for software issues and introduces SWE-Manager, an 8B reasoning model trained with reinforcement learning to both select the best proposal and synthesize a golden proposal. A manual study on 2,852 GitHub issues reveals four main rationales—Risk and Safety, Fix Depth, Maintainability, and Others—informing model design. SWE-Manager combines supervised fine-tuning with DAPO RL to maximize correct selection and high-quality consolidation, achieving 53.21% accuracy and 57.75% earn rate on SWE-Lancer Manager, outperforming GPT-5. The P2A framework tests end-to-end impact, showing SWE-Manager-8B matching GPT-5 in end-to-end patch success on SWE-Lancer IC and offering unique improvements via golden-proposal guidance. Overall, the work demonstrates proposal consolidation as a practical mechanism to improve both decision quality and downstream implementation in real-world software engineering workflows.

Abstract

Large language model (LLM) research in software engineering has largely focused on tasks such as code generation and bug repair. In practice, teams often draft multiple candidate proposals for fixing an issue and then deliberate on one golden proposal for implementation. This selection requires not only assessing the issue's scope, impact, and urgency, but also a clear understanding of each proposal's strengths and weaknesses. A good selection could make issue resolution more reliable while reducing regression and operational risk, whereas a poor choice can increase risk and even cause unpredictable failures. We first conduct a manual study of real-world issues to characterize the rationales maintainers use when selecting among competing proposals. Motivated by these findings, we introduce SWE-Manager, a joint selection and synthesis approach that selects the best proposal and synthesizes a golden proposal. SWE-Manager is an 8B model trained via reinforcement learning (RL) to compare proposals, justify its choice, and synthesize a golden proposal for implementation. We view proposal selection as a reasoning task, mirroring how technical managers review competing proposals by weighing issue context and each proposal's solution without executing code or running tests. On the SWE-Lancer Manager benchmark, SWE-Manager achieves 53.21 selection accuracy and 57.75 earn rate, earning 152,750 dollars and outperforming strong baselines including GPT-5. To further evaluate the effectiveness of SWE-Manager in real-world issue resolution, we design the P2A framework, which simulates a real-world workflow where multiple proposals are drafted, reviewed, and a golden proposal is selected for implementation ...
Paper Structure (39 sections, 2 equations, 7 figures, 4 tables)

This paper contains 39 sections, 2 equations, 7 figures, 4 tables.

Table of Contents

  1. INTRODUCTION
  2. MANUAL STUDY
  3. Study Design
  4. Proposal Selection Reasons
  5. METHOD
  6. SWE-Manager
  7. Task Formulation
  8. Data Preparation.
  9. SFT
  10. DAPO Optimization.
  11. Training Settings
  12. P2A
  13. EXPERIMENT
  14. Benchmarks
  15. Baselines
  16. ...and 24 more sections

Figures (7)

  • Figure 1: Manager task overview. A real GitHub issue is paired with multiple independently written candidate proposals. The model plays the role of a technical manager: it reviews the candidates, rejects less suitable options, and selects the proposal that would be implemented. Performance is evaluated by comparing the selected proposal to the proposal that best matches the approach ultimately merged into the repository; in reward-based settings, a correct match earns the associated payout for that issue.
  • Figure 2: SWE-Manager methodology overview. We formulate proposal selection as taking an issue and its candidate proposals as input, and producing a reasoning trace, a selected proposal ID, a selection justification, and a synthesized golden proposal as output. Then, we curate GitHub issues paired with multiple candidate proposals and a maintainer-selected best-proposal ID. Conditioned on the issue, proposals, and the selected ID, we prompt GPT-5 to generate supervision targets including the reasoning trace, justification, and golden proposal. Starting from Qwen3-8B, we first apply supervised fine-tuning on these targets to teach the model how to reason about proposal selection. We then optimize it with DAPO using weighted rewards for reasoning quality, selection correctness, justification, and golden-proposal synthesis, yielding a model that both selects the best proposal and produces an actionable consolidated plan.
  • Figure 3: P2A comprises three roles: Proposal Agent, Technical Manager, and Implementation Agent. The Proposal Agent samples multiple candidate proposals; each proposal includes an issue restatement, a root-cause analysis, and a proposed solution. The Technical Manager compares proposals and decides on the final proposal for implementation. SWE-Manager plays a Technical Manager role in P2A. Finally, the Implementation Agent follows the golden proposal to produce a code patch that resolves the issue.
  • Figure 4: Left: counts of correct selections and errors grouped by the number of candidate proposals per issue. Right: counts of correct selections and errors grouped by the task reward money. The figure highlights how performance varies with candidate-set size and reward level. As the number of candidate proposals grows, the performance of each method drops. SWE-Manager achieves the strongest performance across most bins.
  • Figure 5: Overlap of solved SWE-Lancer IC instances. Panels (a) and (b) show that P2A and mini-SWE-agent succeed on notably different instance subsets, indicating a distinct coverage profile induced by the P2A workflow, which proves that the golden proposal is effective to guide the Implementation Agent to solve the issue. Panel (c) defines Union as instances solved by any of the top-3 proposal-source backbones, and highlights P2A-only successes that are not recovered by any single source model, supporting the value of golden-proposal consolidation. Panel (d) shows that changing the model of Technical Manager mainly determines which issues are solved, helping explain the earn-rate differences under reward weighting.
  • ...and 2 more figures