DeepSWE

Democratizing Reinforcement Learning for LLM Agents

DeepSWE Overview

DeepSWE-Preview is a fully open-sourced, state-of-the-art coding agent trained with only reinforcement learning (RL) to excel at software engineering (SWE) tasks. DeepSWE-Preview demonstrates strong reasoning capabilities in navigating complex codebases and viewing/editing multiple files, and it serves as a foundational model for future coding agents. The model achieves an impressive 59.0% on SWE-Bench-Verified, which is currently #1 in the open-weights category.

DeepSWE-Preview is trained on top of Qwen3-32B with thinking mode enabled. With just 200 steps of RL training, SWE-Bench-Verified score increases by ~20%.

Discover more about DeepSWE-Preview's development and capabilities in our technical blog post.

Figure 1: SWE-Bench-Verified Performance vs. Model Size for LLM Agents. Trained with only reinforcement learning (RL, no SFT), DeepSWE-Preview with test time scaling (TTS) solves 59% of problems, beating all open-source agents by a large margin. We note that DeepSWE-Preview's Pass@1 performance (42.2%, averaged over 16 runs) is one of best for open-weights coding agents.

Usage Recommendations

To get the best performance out of DeepSWE-Preview, we suggest setting:

• Temperature = 1
• Max tokens set to at least 32-64K.
• Use R2EGym's system/instance prompt and tools (file_editor.py, execution_bash.py, search.py, finish.py). See here for more details.

Training Recipe

Figure 2: Validation Score for SWE-Bench-Hard, where an agent receives positive reward if it submits the final answer and passes all tests. With just 200 steps of RL training, SWE-Bench-Verified score increases from 23→42% (+20%).

Data 🗄️

Our dataset contains 4.5K problems from a subset of R2E-Gym. To avoid data contamination during training, we filtered out problems that are derived from the same repositories as SWE-Bench-Verified , such as sympy. All problems map to individual Docker images.

Environment 🌐

Our environment wraps around R2E-Gym, an existing Gym environment for scalable curation of high-quality executable SWE environments.

State & Action. R2E-Gym defines a set of four tools as part of the action space. The output of each tool (a Python program with stdout/stderr) represents the returned state. More specifically:

• Execute Bash - Outputs both stdout and stderr of an LLM-generated bash command.
• Search - Searches and returns all occurrences of an LLM-defined query in either a directory or a single file.
• File Editor - Allows for viewing, creating, replacing strings, inserting, and undoing edits to a specific file.
• Finish/Submit - LLM has decided that it has resolved the pull request, which terminates trajectory generation.

Reward. To keep things simple, our reward function employs a sparse Outcome Reward Model (ORM):

• 1 - LLM’s generated patch passes a selected sample of tests (Pass2Pass and Fail2Pass) within a time limit. To accelerate training, our max time limit is 5 minutes, while the official SWE-Bench evaluation is 30 minutes.
• 0 - We assign no reward if the LLM’s code fails on at least one test case or times out.

RL Algorithm

We enhance the original GRPO algorithm, integrating insights from DAPO, Dr. GRPO, LOOP/RLOO, and our innovations to enable stable training and improved performance. Our final, amalgamate algorithm consists of:

• Clip High (DAPO): Increasing the upper bound of GRPO/PPO’s surrogate loss encourages exploration and stabilizes entropy.
• No KL Loss (DAPO): Eliminating KL loss prevents the LLM from being constrained to the trust region of the original SFT model.
• No Reward Standard Deviation (Dr.GRPO): Removing reward standard deviation removes difficulty bias in GRPO’s loss, ensuring hard and easy problems are better differentiated.
• Length Normalization (Dr.GRPO): Dividing surrogate loss by max context length removes length bias present in GRPO, which increases the length of incorrect responses.
• Leave One Out (Loop/RLOO): Removing one sample for advantage estimation reduces variance for policy gradient without introducing bias.
• Compact Filtering (Us): Inspired by DAPO, we mask the loss for trajectories that reach max context length, timeout during generation (20 minutes), or reach maximum steps.
• No Entropy Loss (Us): Entropy loss introduces higher instability and eventually leads to exponentially increasing entropy, which collapses training. Provided that the base model’s token-level entropy is within 0.3-1, entropy loss is not needed.

A more detailed description of the training recipe can be found in our blog post.

Evaluation

DeepSWE-Preview is evaluated via the official R2E-Gym codebase at 64k max context length and 100 max enviornment steps. DeepSWE's generated patches are then ported over to the offical SWE-bench repo to calculate final score. Below, We report Pass@1 accuracy averaged over 16 runs.