SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding".
Jane: Here's the gist of the paper. Large language model (LLM) research in software engineering has largely focused on tasks such as code generation and bug repair, but proposal selection remains underexplored The study introduces SWE-Manager, a joint selection-and-synthesis approach that compares candidate proposals under the issue context, selects the best proposal, explains the choice, and synthesizes a golden proposal for implementation Manual Study Insights The manual study of public GitHub issue discussions revealed three properties of maintainer proposal selection First, maintainer choices are guided by recurring high-level rationales, such as Risk and Safety (42.8%), Fix Depth (29.2%), and Maintainability (16.2%) Second, proposal selection is a context-conditioned comparison among candidate proposals where criteria emerge from issue constraints Third, proposal discussion helps formulate the final implementation proposal by providing cues for refinement through clarification and comparison SWE-Manager Design and Training The hypothesis tested is that proposal selection is learnable from large-scale maintainer-selection data SWE-Manager operationalizes this by comparing the issue and all candidate proposals jointly, and it synthesizes a golden proposal that turns the selected direction into implementation-facing guidance The model is instantiated as SWE-Manager-8B, trained from Qwen3-8B using supervised fine-tuning and reinforcement learning Training involves two stages: Supervised Fine-Tuning (SFT) to teach the manager task schema and comparative reasoning, followed by Reinforcement Learning (RL) with DAPO to optimize the decision policy Performance on Benchmarks On the SWE-Lancer Manager benchmark, SWE-Manager-8B achieves 53.21% selection accuracy and a 57.75% earned rate, outperforming GPT-5 (44.15% accuracy and 37.75% earned rate) In the P2A framework on the SWE-Lancer IC benchmark, P2A with SWE-Manager achieves a 55.6% issue-resolution rate, matching GPT-5 and improving over miniSWE-agent by 7.1% Downstream Usefulness Evaluation The P2A framework evaluates the synthesized golden proposal's usefulness by using it to guide an implementation agent to produce a patch The selected-proposal control experiment shows that the synthesized golden proposal is more effective than the selected proposal in guiding the Implementation Agent, resolving 18 more issues and increasing earned reward from 58,125 to 69,875 Ablation Study Results The ablation study indicates that proposal selection improves when training requires golden-proposal synthesis rather than only best-proposal ID prediction Requiring the model to synthesize a golden proposal encourages a more explicit evaluation of candidate strengths and weaknesses, which translates into higher selection accuracy and reward Furthermore, supervised fine-tuning provides an effective warm start by teaching the model the desired reasoning structure before applying reinforcement learning via DAPO Robustness and Limitations SWE-Manager-8B remains competitive with GPT-5 on smaller proposal pools but loses accuracy as the pool grows beyond three proposals due to distribution shift A limitation is that the current SWE-Manager operates without code access, suggesting a future direction for a tool-augmented manager Future work should rebalance training toward larger proposal pools and use curriculum strategies to handle longer contexts more robustly Related Work Context Proposal selection is framed as a decision problem where the core challenge is comparing multiple candidate solutions while jointly considering functional requirements and risks This builds on prior work that treats selection as a computational ranking problem, but SWE-Manager focuses on learning issue-specific selection and synthesis over proposal pools before implementation begins The framework integrates concepts from automated program repair by studying the managerial step of selecting among competing proposals before code modification
Paper summary: Tom: So we’ve looked at SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding, which is this joint approach comparing candidates under the issue context to select the best one and then synthesizing a golden proposal for implementation <ref:2601.22956#pg1>.
Jane: The paper suggests that this selection process contains learnable patterns based on maintainer choices like risk and safety, but they also showed that synthesizing a proposal before coding is useful for guiding the actual patch generation agent
SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding: .
Lu: What this means is that instead of just treating proposal selection as a simple ranking problem, we’re learning how to handle the complex comparison between multiple technically plausible solutions based on real engineering constraints <ref:2601.22956#pg1>.
Meng: For us in production, it means we could build systems where the AI doesn't just suggest one fix but actually provides a vetted, synthesized plan for implementation that has already passed some internal quality checks
SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding: .
Lalam: I think the authors are arguing that this capability—the selection and synthesis step—is the next big thing after basic code generation or bug repair <ref:2601.22956#pg1>.
Tom: They’re proposing a framework where the model acts like a technical manager, reviewing candidates, rejecting bad ones, and then creating that final golden proposal
SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding: .
Jane: It moves us closer to an AI that understands not just *what* code to write, but *why* it’s the right code for the specific context of a bug report
SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding: .
Lu: The authors are testing this on benchmarks like SWE-Lancer and showing that the performance gains come from requiring the model to synthesize rather than just identifying the best option alone <ref:2601.22956#pg3>.
Meng: It’s a shift from a simple suggestion engine to a decision-making system that understands workflow constraints, which is what you need for real software engineering tasks
SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding: .
Lalam: The core of this work is learning issue-specific selection and synthesis over proposal pools before any actual code modification takes place
SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding: .
Conclusion: Tom: So, we're wrapping up our look at SWE-Manager and its main conclusion about proposal selection and synthesis before implementation.
Jane: It basically boils down to this: this paper tests the idea that you can learn how to pick the best code fix from a bunch of options by training an AI on real human choices made by maintainers.
Lu: Right, they're not just looking for the single best proposal; they're building a whole system that selects it and then synthesizes a perfect plan for coding it.
Meng: So, what does this mean when you look at the authors and their results? Are we looking at some small tweak or something bigger?
Tom: The authors trained an eight billion parameter model on data from maintainer selections, using supervised fine-tuning and reinforcement learning to get that synthesis skill.
Jane: And they showed it actually beats a baseline model, GPT-five on the SWE-Lancer benchmark for selection accuracy, which is about fifty-three point two one percent.
Lu: The paper highlights that requiring the model to synthesize a golden proposal instead of just picking the best one actually makes the selection process more accurate.
Meng: I'm wondering how much effort that synthesis part adds to the overall system when you're trying to deploy it at scale? Does it slow things down too much?
Tom: Well, they found that this synthesized proposal is way better for a real implementation agent; it resolves eighteen more issues and boosts the reward score from fifty-eight thousand one hundred twenty-five up to sixty-nine thousand eight hundred seventy-five.
Jane: It seems like the authors are arguing that this selection and synthesis step is a really important part of the workflow, not just an extra feature.
Lu: They're framing it as learning issue-specific selection over proposal pools before any actual code modification starts, which is different from just treating it as a ranking problem.
Meng: So, if we take that idea—learning the context for selection—what’s the next hurdle when you try to put this into a real production environment?
Tom: The paper does flag one limitation: the model's performance starts dropping once you give it more than three proposals to choose from because of distribution shift.
Jane: That means it’s strong for simpler cases, but handling really messy, complex issues with a huge number of potential fixes is still a challenge for this current setup.
Boyin Tan, Haoning Deng, Junyuan Zhang, Junjielong Xu, Pinjia He, Youcheng Sun
Mohamed bin Zayed University of Artificial Intelligence
cs.SE
Submitted: 2026-01-30
Updated: 2026-10-31
Code: https://github.com/shuaijiumei/SWE-Manager
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 82/100
The gist: Large language model (LLM) research in software engineering has largely focused on tasks such as code generation and bug repair…
Key concepts
- Proposal Selection
- This is the core decision-making process where a system must compare several different potential solutions (proposals) for fixing a bug. The selection process considers the specific context of the issue and aims to choose the most appropriate candidate based on learned criteria like risk or maintainability.
- Golden Proposal Synthesis
- This is the final step where SWE-Manager takes the best selected proposal and refines it into a perfect, implementation-ready version. Instead of just picking one option, it synthesizes a superior guide that helps the actual coding agent write clean, correct code.
- Context-Conditioned Comparison
- The system learns to compare proposals not in isolation but based on the specific constraints and details provided by the issue itself. The criteria used for comparison—such as risk or fix depth—emerge directly from analyzing the unique requirements of that particular software problem.
- DAPO Reinforcement Learning
- This is a training technique used to optimize SWE-Manager's decision-making policy. DAPO helps the model learn the best way to select proposals by rewarding it for making good choices, ensuring it learns a robust strategy for choosing solutions.
Terminology
Summary
Large language model (LLM) research in software engineering has largely focused on tasks such as code generation and bug repair, but proposal selection remains underexplored This paper introduces SWE-Manager, a joint selection-and-synthesis approach that compares candidate proposals under the issue context, selects the best proposal, explains the choice, and synthesizes a golden proposal for implementation
Manual Study Insights
The manual study of public GitHub issue discussions revealed three properties of maintainer proposal selection First, maintainer choices are guided by recurring high-level rationales, such as Risk and Safety, Fix Depth, and Maintainability Second, proposal selection is a context-conditioned comparison among candidate proposals Third, proposal discussion helps formulate the final implementation proposal by providing refinement evidence
SWE-Manager Design and Training
SWE-Manager is designed to compare the issue and all candidate proposals jointly and synthesize a golden proposal that turns the selected direction into an implementation plan The model is instantiated as SWE-Manager-8B, trained from Qwen3-8B using supervised fine-tuning and reinforcement learning Training involves a two-stage recipe where Supervised Fine-Tuning (SFT) teaches the model the manager task schema, and subsequent Reinforcement Learning (RL) optimizes the policy using a compact weighted reward over four structured fields
Performance Evaluation
SWE-Manager-8B was evaluated on two capabilities: proposal selection and downstream usefulness of the synthesized golden proposal On the SWE-Lancer Manager benchmark, SWE-Manager-8B achieved 53.21% selection accuracy and a 57. p. 0 On the SWE-Lancer Manager benchmark, SWEManager-8B achieves 53. 21% selection accuracy and a 57. On the SWE-Lancer Manager benchmark, SWEManager-8B achieves 53. 21% selection accuracy and a 57. 75% earned rate ($99,875). To evaluate
whether the golden proposal can guide implementation, we
introduce P2A, a proposal-first, implementation-later workflow,
and evaluate it on the SWE-Lancer IC benchmark. P2A with
SWE-Manager achieves a 55. 6% issue-resolution rate, matching
the performance of P2A with GPT-5 and improving over the miniSWE-agent by 7.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Key Findings on Learnability
The ablation study showed that proposal selection improves when training requires golden-proposal synthesis rather than only best-proposal ID prediction The results suggest that index-only supervision provides limited learning signal, whereas requiring the model to synthesize a golden proposal encourages a more explicit evaluation of candidate strengths and weaknesses Supervised fine-tuning provides an effective warm start by teaching the model the desired reasoning structure before applying reinforcement learning
Downstream Utility of Synthesis
The P2A framework demonstrated that the synthesized golden proposal is more effective than just selecting the original proposal in guiding downstream implementation Specifically, using its golden proposal instead of its selected proposal resolves 18 more issues and increases earned reward from 58,125 to 69,875 This indicates that comparative analysis and synthesis of candidate proposals is a meaningful capability that improves downstream implementation
Robustness and Limitations
The robustness of SWE-Manager-8B is limited when the proposal pool grows beyond three proposals, as the input context becomes substantially longer and less represented during training A limitation is that the current SWE-Manager operates without code access, suggesting a future direction for a tool-augmented manager to selectively inspect repository context Future work should focus on rebalancing training toward larger proposal pools and using curriculum strategies to handle longer contexts
Related Work Context
Proposal selection is a recurring decision problem across the software engineering life cycle where engineers must compare multiple candidate solutions to the same software problem Prior work has addressed this complexity through rational analysis methods like ATAM and CBAM, while recent LLM-based work treats selection as a computational ranking problem SWE-Manager builds on this decision-oriented view by focusing on learning issue-specific selection and synthesis over proposal pools before implementation begins The study contrasts its focus with work on code generation and bug fixing, which primarily studies the implementation or validation of repairs
Conclusion
Learned proposal selection and synthesis can improve LLM-based issue resolution, while larger proposal pools and longer contexts remain important directions for future work The findings motivate formulating proposal selection as a learnable comparison-and-synthesis task rather than an isolated label-prediction problem SWE-Manager achieves 53.21% selection accuracy and a 57.75% earned rate, surpassing GPT-5 and supporting the claim that proposal selection policies can be learned by a small model P2A with SWE-Manager reaches a 55. p. 6 We answer this question by comparing a trained 8B model,
SWE-Manager-8B, against the pre-training backbone model
Qwen3-8B and GPT-5 in the direct selection setting, and
GPT-5 with code-access agent selection on the SWE-Lancer
Manager benchmark, using the ground-truth match rate Match
(%) and the earned rate Earned (%) as the primary metrics. This comparison directly tests whether proposal selection can
be systematically learned by a small model. On the SWE-Lancer IC benchmark,
P2A with SWE-Manager achieves a 55. 6% pass rate, matching
the performance of GPT-5 and improving over the miniSWE-agent [16] baseline without proposal guidance by 7.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
The SWE-Manager system can be deployed as a technical manager in real development workflows by comparing candidate proposals under issue context, selecting the best proposal, explaining that choice, and synthesizing a golden proposal for implementation.
-
The P2A framework powered by SWE-Manager enables an end-to-end issue resolution capability where the synthesized golden proposal guides an implementation agent to produce a patch, achieving a 55.6% pass rate on the SWE-Lancer IC benchmark, which is 7.1% better than the miniSWE-agent baseline.
-
The improved AI can learn reusable reasoning patterns for proposal selection by being trained on large-scale maintainer-selection data, moving beyond instance-specific preferences to generalize across different issue contexts and constraints.
-
The system can generate implementation guidance that improves downstream issue resolution; specifically, the selected proposal control experiment shows that
the synthesized golden proposal is more effective than the selected proposal in guiding the Implementation Agent to resolve the issue.
-
The model's performance in selection and synthesis is maximized by adopting a rich objective, as demonstrated by Fig. 4, where requiring the model to synthesize a golden proposal leads to higher selection accuracy and reward compared to predicting only the best proposal ID.
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...
Sources
- Code Llama: Open Foundation Models for Code
- StarCoder 2 and The Stack v2: The Next Generation
- SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents
- Qwen3 Technical Report
- SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
- Training language models to follow instructions with human feedback
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- DAPO: An Open-Source LLM Reinforcement Learning System at Scale
- SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuning
- Scalable Supervision for Software Agents via Patch Reasoning
- SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement
- The Stack: 3 TB of permissively licensed source code
- SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
- SWE-Fixer: Training Open-Source LLMs for Effective and Efficient GitHub Issue Resolution
Related papers
- ArkEval: Benchmarking and Evaluating Automated CodeRepair for ArkTS
- TestPrism: Rethinking Test Evaluation Beyond a Single Reference
- Mining Workflow Graphs for Black-Box Boundary Testing of Conversational LLM Agents
- Falsification-Based Verification of LLM-Generated Optimization Models: Sound Test Batteries and Their Detection Limits
- GitSkills: A Dataset of Agent Skills on GitHub
- For Coding Agents, the Repository Matters