Logics-SWE-Qwen3.6-27B

Official model release for One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents.

📄 Paper · 🤗 Dataset: Logics-SWE-Env-2.5K · 💻 GitHub: AgenticBigBang

📰 News

• [2026.10.30] Submitted the expanded paper as arXiv v2 and added results on Pro V2, Hard-51, Multilingual, SWE-bench Pro, and LiveCodeBench v6. We also report a 35B-A3B student trained with the same three 27B category experts.
• [2026.09.20] Released the RRE–MOPD paper on arXiv.
• [2026.09.18] Released Logics-SWE-Qwen3.6-27B under the Apache-2.0 license.

🔎 Overview

Logics-SWE-Qwen3.6-27B is a 27B-parameter model developed for repository-level software engineering agents. Starting from the Qwen3.6-27B model used in our study, it combines category-aware expert development with multi-teacher on-policy distillation into a single deployment policy.

Repository-level tasks require agents to navigate code, edit files, execute commands, inspect feedback, and iteratively repair their solutions. Our work starts from the category see-saw: aggregate progress during joint RL can conceal opposing changes across task categories. Category splitting alone does not guarantee stronger experts. We therefore develop the experts through RRE before integrating their learned behaviors into one model.

Stage 1: Refresh–Repair–Expand expert development

SWE Labeler organizes training tasks into three repository-domain categories:

• A: service, data, and security.
• B: user-facing applications.
• C: systems, tooling, and runtimes.

All experts start from the same base model. Refresh–Repair–Expand (RRE) alternates executable-reward RL with supervised replay of each expert's own verified successful trajectories. Mastery refreshes track how task success rates change; repair consolidates successful behavior, and expansion revisits additional tasks to identify the next training frontier. Agentic-miniRL provides the shared long-horizon RL recipe.

Stage 2: Single-policy integration with MOPD

The student starts from the common base and generates its own trajectories. Each training task is routed to its corresponding category expert for token-level supervision. Reference-anchored extrapolation augments imitation, using a positive teacher–reference gap gate. The student receives no direct environment-reward term during this pure distillation stage.

Neither expert development nor integration uses an external model to provide solution trajectories or action targets.

Here, integration means on-policy distillation, not arithmetic averaging of expert weights. Inference uses one model; category labels and separate expert models are not required for deployment.

Intended use: research on repository-level issue resolution, coding agents, long-horizon post-training, expert development, and policy distillation.

📊 Experimental Results

Expanded evaluation

Logics-SWE-Qwen3.6-27B improves on Qwen3.6-27B across four software-engineering and coding evaluations. Scores are three-round means (%); gains are in percentage points and calculated before rounding.

The figure also shows a Qwen3.6-35B-A3B student distilled using the same three 27B category experts. Qwen baseline scores are from our own evaluations under the same settings as Logics-SWE.

Pro-618 in the original paper

The original paper reports 58.04 ± 0.20% for Logics-SWE-Qwen3.6-27B and 52.64 ± 0.28% for Qwen3.6-27B on Pro-618, a 618-task subset of SWE-bench Pro. Values are the mean ± population standard deviation over three evaluation rounds.

🚀 Quickstart

Transformers: text-generation smoke test

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Logics-MLLM/Logics-SWE-Qwen3.6-27B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

messages = [{
    "role": "user",
    "content": "A parser crashes on an empty input file. Describe how you would "
               "investigate the cause, implement a fix, and test for regressions.",
}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=True,
        temperature=1.0,
        top_p=0.95,
        top_k=20,
    )
completion = output[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(completion, skip_special_tokens=True))

The example follows the Transformers chat-template interface. Its short generation budget is for demonstration only. Repository-level agent use additionally requires the matching system prompt, tool schema, action parser, sandbox, and verifier. Model-generated commands should not be executed automatically outside an isolated, permission-controlled environment.

📦 Training Data and Release Scope

The public Logics-SWE-Env-2.5K release provides 2,553 software-engineering task instances and their matching environment-image references. It is a curated subset of the training tasks used in the paper; it does not contain the complete training corpus or agent trajectories.

The AgenticBigBang GitHub repository additionally provides the label taxonomy, SWE Labeler code, anonymous training manifests, benchmark outcomes, aggregate results, and training and evaluation configuration specifications.

📚 Citation

If you use this model, please cite:

@misc{zhao2026onetomoremoretoone,
  title         = {One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents},
  author        = {Jie Zhao and Ziyu Jiang and Suhang Zheng and Minghui Shan and Xiaoxiao Xu and Lin Qu},
  year          = {2026},
  eprint        = {2609.23377},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2609.23377}
}

🙏 Acknowledgements

We thank the Qwen, ROLL, R2E-Gym, SWE-agent, SWE-bench, SWE-bench Pro, and SWE-bench Multilingual contributors, and the research communities behind MiniRL, MOPD, and ExOPD.

License

The model weights are released under the Apache License 2.0.