SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training

If you like our project, please give us a star ⭐ on GitHub for the latest update.

✨ News

• [24 Feb 2026] We have open-sourced the training and inference code, as well as all associated models for SWE-Master. Start training and evaluating your own code agents!
• [4 Feb 2026] ⚡️⚡️ SWE-Master: We introduce SWE-Master, a fully reproducible post-training framework for Qwen2.5-Coder-32B that integrates agentic SFT and RL to achieve 61.4% (Pass@1) and 70.8% (TTS@8) resolve rates on SWE-bench Verified. Meanwhile, the framework incorporates IDE-level capabilities during inference via LSP-driven tool.
• [4 Feb 2026] ⚡️⚡️ SWE-World: We introduce SWE-World, a fully Docker-free framework that replaces physical execution environments with learned surrogates. It lifts Qwen2.5-Coder-32B from 6.2% to 55.0% on SWE-bench Verified via agentic SFT and RL, and further attains 68.2% through test-time scaling (TTS@8).

💡 Overview

SWE-Master is an end-to-end, open-source post-training pipeline for software engineering (SWE) agents. It covers trajectory synthesis & curation, long-horizon supervised fine-tuning (SFT), reinforcement learning with execution feedback (RLVR), and test-time scaling (TTS) via simulated verification and ranking. Furthermore, the framework supports advanced tool-use capabilities, including LSP-integrated tools (based on the Language Server Protocol), and incorporates a summary-based context manager for efficient state tracking.

🤖 Models

You can find all the trained models Here, including:

• SWE-Master-4B-SFT
• SWE-Master-4B-RL
• SWE-Master-32B-SFT
• SWE-Master-32B-RL
• SWE-Master-32B-RL-LSP

✨ Key Insights

1. Fully Post-Training Pipeline WE-Master provides a transparent, end-to-end framework spanning trajectory synthesis, long-horizon SFT, and RL with execution feedback, demystifying how high-performance SWE capabilities emerge from open-source models.
2. Effective Test-Time Scaling (TTS) via Simulated Feedback: Beyond simple repetition, we leverage LLM-based environment feedback for simulated verification and ranking. This strategy narrows the gap between Pass@1 and Pass@k, reaching a 70.8% resolve rate.
3. IDE-Level Semantic Navigation with LSP-Driven Tools: By integrating Language Server Protocol (LSP) capabilities, SWE-Master replaces fragile text searches with deterministic tools (e.g., go-to-definition), significantly reducing search friction and enhancing interaction efficiency in complex codebases.

✨ Training Recipe for SWE-Master

Data Curation and SFT

We integrate multiple open-source SWE datasets packaged with Docker environments and generate rollouts with teacher models, then apply filtering to produce SFT-ready trajectories. The statistical metrics are shown below:

Additionally, we perform $N$ rollouts for each issue and utilize the average resolve rate as a proxy for task difficulty. As illustrated in the figure below, the resulting distribution exhibits a bimodal pattern, where the majority of issues are either consistently solved (trivial) or consistently failed (intractable). Consequently, we exclude these polar extremes from the candidate pool, selectively retaining only those issues that yield a mixture of successful and failed trajectories.

Reinforcement Learning

We adopt Reinforcement Learning from Verifiable Rewards (RLVR) as our foundational framework, utilizing the Group Relative Policy Optimization (GRPO) algorithm to directly optimize SWE-Master-SFT. Drawing on established RL methodologies, we incorporate several empirical optimizations to enhance training stability and performance. The final objective function is formulated as follows (refer to the technical report for comprehensive details):

$$ \mathcal{J}(\theta) = \mathbb{E}{q \sim P(Q), {o_i}{i=1}^G \sim \pi_{\text{old}}} \left[ \frac{1}{G} \sum_{i=1}^G \frac{1}{L_{\text{max}}} \sum_{t=1}^{|o_i|} \min \left( \rho_{i,t}(\theta) \hat{A}i, \text{clip}\left(\rho{i,t}(\theta), \beta_{\text{low}}, \beta_{\text{high}}\right) \hat{A}_i \right) \right] $$

where $P(Q)$ denotes the distribution of input prompts, and the clipping bounds are defined as $\beta_{\text{low}} = 1-\varepsilon_{\text{low}}$ and $\beta_{\text{high}} = 1+\varepsilon_{\text{high}}$. The term $L_{\text{max}}$ represents a fixed constant for length normalization. The probability ratio $\rho_{i,t}(\theta)$ and the group-relative advantage $\hat{A}_i$ are given by:

$$ \rho_{i,t}(\theta) = \frac{\pi_\theta(o_{i,t}|q, o_{i,<t})}{\pi_{\text{old}}(o_{i,t}|q, o_{i,<t})}, \quad \hat{A}i = R_i - \frac{1}{G-1}\sum{j=1, j\neq i}^G R_j $$

The training dynamics of RL are shown in the figure below.

Test-Time Scaling

We implemente two primary paradigms for TTS: sequential scaling, which involves increasing the allowance of interaction turns; and parallel scaling, which generates multiple trajectories and corresponding patches, followed by employing a specific verifier to select the optimal candidate for submission. For parallel scaling, we use SWE-World as the verifier to choose the optimal patch.

✨ IDE-Level Code Capacity

Overview of LSP Tools

The Language Server Protocol (LSP) is a standardized standard that decouples development tools from language-specific intelligence, enabling IDE-grade semantic insights such as cross-file reference tracking and symbol resolution. Inspired by human-IDE workflows, we implemented lsp_tool, which adapts the LSP specification into a function-calling format optimized for LLMs. This tool encapsulates repository navigation, dependency analysis, and code understanding into a unified API. By utilizing a specialized parser to handle the underlying JSON-RPC complexity, lsp_tool empowers agents with deterministic, semantics-aware navigation, bridging the gap between raw protocol data and model-driven code exploration.

Case Study

To explicitly demonstrate how lsp_tool empowers the model to transcend the limitations of lexical search, we present a comparative case study on pydata_xarray-6812 from SWE-bench Verified. The task requires fixing a subtle bug where .swap_dims() incorrectly modifies the original dataset object in-place due to improper reference handling. The figure below visualizes the contrasting trajectories between the current frontier framework, OpenHands, and our enhanced framework integrated with lsp_tool, and the detailed analysis is provided in the paper:

🏃 Quick Start

Data Preparation

Data Download

bash ./data_preparation/download_swe_datasets.sh

Data Preprocessing (Optional)

Add a prior difficulty grading to each issue. These grades are generated using Minimax-M2 and GLM-4.6.

# Add prior difficulty scores
python ./data_preparation/difficulty_score_add.py

# Filter for high-quality issues by removing those with consistently all-pass or all-fail rollouts based on difficulty scores
python ./data_preparation/difficulty_score_filter.py

Offline Unit Test Generation (Optional)

Generate unit tests for swe-gym, swebench-verified, and swe-rebench to facilitate direct loading during evaluation. This process generates two types of caches:

1. JSON files for unit tests mapped to each instance_id within the specified base_dir.
2. A dataset containing the make_test_spec field.

During inference, the system first checks if the dataset contains the make_test_spec field, then checks for the cached directory in the workspace. If neither is found, it reverts to the default loading method.

Note: If base_dir is specified, you must update the corresponding base_dir in ./R2E-Gym/src/r2egym/agenthub/runtime/docker.py.

python ./data_preparation/make_test_spec.py \
    --base_dir /your/custom/cache/path \
    --data_file_path /your/path/to/data.json

For a completed conversion demo, refer to: ./data_examples/inference_data/ood_data The corresponding unit tests cache is located at: ./data_examples/inference_data/ood_data_unit_tests_cache

Inference & Evaluation

Environment Setup

# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env

# Activate venv
uv venv
source .venv/bin/activate
uv sync && uv pip install -e .

# Install SWE dataset-related packages for inference
uv pip install ./swebench_fork_swegym-2.0.13-py3-none-any.whl

uv pip install ./swebench_fork_swerebench-4.0.3-py3-none-any.whl

uv pip install ./swesmith-0.0.7-py3-none-any.whl

Inference

• We have built our project upon the R2E-Gym framework and conducted further development. For parameter details, see Parameter Description

Standard Inference

# Batch generation
bash ./R2E-Gym/rollout_foundation_model/run_deepseek_v3.sh

# Debugging (e.g., verifying Docker connectivity and interaction workflows)
bash ./R2E-Gym-opensource/run.py

Inference with LSP Tool

# Standard inference with LSP-Tool integration
bash ./R2E-Gym/rollout_foundation_model/lsp/run_deepseek_v3_lsp.sh

# Few-shot configuration: used to prevent the policy model from over-relying on traditional tools during training (LSP-Tool is disabled here)
bash ./R2E-Gym/rollout_foundation_model/lsp/run_minimax_lsp_few_shot.sh

# Model-specific few-shot configurations are located here:
ls ./R2E-Gym/rollout_foundation_model/lsp/swe_LSP_few_shot

# Quality filtering for trajectories using LSP-Tool (Rule-based and LLM-based)
bash ./R2E-Gym/rollout_foundation_model/lsp/sft_data_filter/batch_run_llm_judge.sh

Inference with Context Manager

bash ./R2E-Gym/rollout_foundation_model/memory/run_deepseek_v3_memory.sh

# Convert a single trajectory using context manager into multiple training trajectories
python ./R2E-Gym/rollout_foundation_model/memory/sft_process_memory_data_01.py

Examples and Visualization

A sample trajectory file generated after inference is available at:
./data_examples/r2e-gym-inference-traj/glm46_0_used_swe_rebench_1_demo.jsonl

To facilitate a more intuitive and efficient analysis of the generated trajectories, you can use the following scripts to convert them into HTML format for visualization:

# Generate English visualization
python ./R2E-Gym/app/json2html_for_view-en.py [jsonl-file-path] -o [html-file-path]

# Generate Chinese visualization
python ./R2E-Gym/app/json2html_for_view-zh.py [jsonl-file-path] -o [html-file-path]

For the demo data provided, the corresponding visualized HTML files are located at:

• English version: ./data_examples/r2e-gym-inference-traj/glm46_0_used_swe_rebench_1_demo-en.html
• Chinese version: ./data_examples/r2e-gym-inference-traj/glm46_0_used_swe_rebench_1_demo-zh.html

Reproducing SWE-Master Results

Use the following scripts to reproduce the evaluation results for SWE-Master:

# Launch vLLM server
bash ./R2E-Gym/rollout_trained_model/vllm_server_launch.sh

# Run evaluation
bash ./R2E-Gym/rollout_trained_model/qwen25_coder_15k_lsp_with_gerenal_04.sh

For Test-Time Scaling (TTS), we utilize SWE-World as a verifier to perform candidate patch selection. Detailed methodology is available in the original paper, and the implementation can be found in the SWE-World repository, with the TTS pipeline further documented in README.md.

SFT Training

[!IMPORTANT] We utilize OpenRLHF as our framework for Supervised Fine-Tuning (SFT). It is an excellent open-source repository.

[!IMPORTANT] Training was conducted on an internal platform with packages pre-installed in the environment image. The "Environment Setup" provided here aims to replicate our training environment as closely as possible. If you encounter issues, please open an issue, and we will investigate and resolve it promptly.

Environment Setup

The packages required for the training environment are listed below and can be downloaded as needed:

cat ./SFT_ENV.txt

You can refer to the environment configuration provided by OpenRLHF for installation. If you are using Ray, ensure that the environment is configured uniformly across all nodes:

cd ./OpenRLHF_SFT
uv venv --python 3.11
source .venv/bin/activate
git clone https://github.com/OpenRLHF/OpenRLHF.git
cd OpenRLHF
pip install -e .

Data Processing (Optional)

You can directly process the trajectories generated via inference using the R2E-Gym framework by running the following scripts. This will convert them into an SFT-compatible format and apply necessary filtering:

python ./OpenRLHF_SFT/SFT_data_pre_process/r2e_to_openrlhf_format/0_covert_r2e_format_to_sft_foramt.py

python ./OpenRLHF_SFT/SFT_data_pre_process/r2e_to_openrlhf_format/1_init_format_filter.py

Similarly, you can sequentially run the scripts in the directory below to obtain Best-of-N (BoN) statistics. Note that the prior BoN-based difficulty grading introduced in earlier section was generated using the code within this directory: ./OpenRLHF_SFT/SFT_data_pre_process/bon_filter

An example of SFT data is provided here for formatting reference: ./data_examples/sft_data

Training

# Pre-tokenize multi-turn interaction trajectories and compute loss masks (Optional, but recommended to avoid training bottlenecks)
# Directly embeds function descriptions into the system prompt
python ./OpenRLHF_SFT/scripts_swe_master/sft_data_pre_tokenize.py 

# Train the model to support tool_calls
python ./OpenRLHF_SFT/scripts_swe_master/sft_data_pre_tokenize_toolcall.py 

# Optional: This step is required if you pre-tokenized the multi-turn data in the previous steps
mv ./OpenRLHF_SFT/scripts_swe_master/sft_dataset.py ./OpenRLHF_SFT/datasets/ 

# Launch training
bash ./OpenRLHF_SFT/scripts_swe_master/qwen_25_coder_32B_new_remove_01_not_dedep.sh

RL Training

[!IMPORTANT] Our development is built upon the open-source frameworks DeepSWE and rLLM, both of which are highly recommended repositories.

[!IMPORTANT] Training was conducted on an internal platform with packages pre-installed in the environment image. The "Environment Setup" provided here aims to replicate our training environment as closely as possible. If you encounter issues, please open an issue, and we will investigate and resolve it promptly.

Environment Setup

The packages required for the training environment are listed below and can be downloaded as needed:

cat ./RL_ENV.txt

You can refer to the environment configurations provided by DeepSWE and rLLM for installation. If you are using Ray, ensure that the environment is configured uniformly across all nodes:

cd ./DeepSWE_RL/rllm
uv venv --python 3.11
source .venv/bin/activate
uv pip install -e .
git clone https://github.com/verl-project/verl.git # We use version v0.5
uv pip install -e ./verl

Additionally, you must update the local paths for VeRL and R2E-Gym (corresponding to the training and inference frameworks, respectively) in the following file:

• Update the paths corresponding to the two sys.path.insert functions within train_agent in: ./DeepSWE_RL/rllm/rllm/trainer/verl/train_agent_ppo.py

Furthermore, to support training across different SWE datasets, you must ensure the relevant dataset libraries are installed when initializing Ray via the run_ppo_agent function in ./DeepSWE_RL/rllm/rllm/trainer/verl/train_agent_ppo.py (Alternatively, these can be pre-installed during the initial environment setup).

Data Processing (Optional)

If your training server operates in an offline environment (no internet access), you must pre-process the unit tests for the specific SWE datasets and then standardize their formats:

# Standardize the format
python ./data_preparation/make_test_spec_for_rl.py

An example of processed data ready for RL training can be found here for formatting reference: ./data_examples/rl_data/ood_data

Training

bash ./DeepSWE_RL/rllm/examples/swe/swe_rl_debug_hls_try_arsenal_long_step_and_long_time.sh

Acknowledgements

• Our work builds upon several outstanding open-source repositories, including OpenRLHF, R2E-Gym, VeRL, DeepSWE and rLLM. We are sincerely grateful to the authors for making their high-quality training code available to the community.
• This work was conducted at the BOSS Zhipin Nanbeige Lab. We would like to express our gratitude to the company for providing the necessary computing resources and professional guidance.

📄 Citation

Please kindly cite our report if they are helpful for your research.

@misc{song2026swemasterunleashingpotentialsoftware,
      title={SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training}, 
      author={Huatong Song and Lisheng Huang and Shuang Sun and Jinhao Jiang and Ran Le and Daixuan Cheng and Guoxin Chen and Yiwen Hu and Zongchao Chen and Wayne Xin Zhao and Yang Song and Tao Zhang and Ji-Rong Wen},
      year={2026},
      eprint={2602.03411},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2602.03411}, 
}

📄 License

This project is released under the MIT License.

📞 Contact

For any questions or feedback, please reach out to us at songhuatong123@ruc.edu.cn.