Doing More with Less Tokens: Hierarchical Reinforcement Learning for Efficient Coding Agents
HERO-35B-A3B is post-trained from Qwen3.5-35B-A3B using HERO (HiErarchical ReinfOrcement learning) to improve token efficiency while prioritizing task resolution.
| Item | Description |
|---|---|
| Base model | Qwen/Qwen3.5-35B-A3B |
| Parameters | 35B total, approximately 3B active (MoE) |
| Architecture and tokenizer | Inherited from Qwen3.5-35B-A3B |
| Training | HERO on 640 multilingual SWE tasks from SWE-Gym, Multi-SWE-bench, and SWE-rebench |
| Intended use | Repository-level coding agents |
| Format | Hugging Face weights and tokenizer |
HERO combines capability-based efficiency gating, resolution-first clipping, and efficiency credit at trajectory and turn levels. Qwen3.5 is the backbone; this release contains the HERO post-trained weights.
Follow the preparation guide in the HERO code repository. From that repository's root,
with this model saved under ../models/HERO-35B-A3B/:
MODEL=../models/HERO-35B-A3B bash eval/run_eval_swebench_verified.sh
MODEL=../models/HERO-35B-A3B bash eval/run_eval_swebench_multilingual.sh
The released evaluation scripts use the following defaults for both SWE-bench Verified and SWE-bench Multilingual:
| Setting | Value |
|---|---|
| Agent scaffold | Claude Code |
| Inference backend | SGLang |
| Temperature | 0.6 |
| Top-p | 0.95 |
| Context length | 131,072 tokens |
| Maximum output per call | 16,000 tokens |
| Maximum agent turns | 200 |
| Automatic context compaction | Disabled |
| Tools | Default tools, excluding WebFetch, WebSearch, and Agent |
See the evaluation script for configuration options.
Apache-2.0. See LICENSE. We acknowledge the Qwen team for the base model.
If you find HERO useful, please cite our paper:
@misc{li2026hero,
title = {Doing More with Less Tokens: Hierarchical Reinforcement Learning for Efficient Coding Agents},
author = {Haobin Li and Liang Jiang and Zhenyu Huang and Mouxing Yang and Xi Peng},
year = {2026},
eprint = {2609.38885},
archivePrefix = {arXiv},
primaryClass = {cs.SE},
url = {https://arxiv.org/abs/2609.38885}
}
model-00131-of-00132.safetensors 1.0 GBmodel-00132-of-00132.safetensors 1.0 GBmodel-00055-of-00132.safetensors 537 MBmodel-00068-of-00132.safetensors 537 MBmodel-00019-of-00132.safetensors 537 MBmodel-00032-of-00132.safetensors 537 MBmodel-00045-of-00132.safetensors 537 MBmodel-00058-of-00132.safetensors 537 MBmodel-00071-of-00132.safetensors 537 MBmodel-00084-of-00132.safetensors 537 MBmodel-00097-of-00132.safetensors 537 MBmodel-00110-of-00132.safetensors 537 MBmodel-00087-of-00132.safetensors 537 MBmodel-00100-of-00132.safetensors 537 MBmodel-00011-of-00132.safetensors 537 MBmodel-00014-of-00132.safetensors 537 MBmodel-00024-of-00132.safetensors 537 MBmodel-00027-of-00132.safetensors 537 MBmodel-00037-of-00132.safetensors 537 MBmodel-00050-of-00132.safetensors 537 MBmodel-00063-of-00132.safetensors 537 MBmodel-00079-of-00132.safetensors 537 MBmodel-00089-of-00132.safetensors 537 MBmodel-00092-of-00132.safetensors 537 MBmodel-00102-of-00132.safetensors 537 MBmodel-00105-of-00132.safetensors 537 MBmodel-00008-of-00132.safetensors 537 MBmodel-00022-of-00132.safetensors 537 MBmodel-00035-of-00132.safetensors 537 MBmodel-00048-of-00132.safetensors 537 MBmodel-00061-of-00132.safetensors 537 MBmodel-00074-of-00132.safetensors 537 MBmodel-00009-of-00132.safetensors 537 MBmodel-00021-of-00132.safetensors 537 MBmodel-00034-of-00132.safetensors 537 MBmodel-00047-of-00132.safetensors 537 MBmodel-00060-of-00132.safetensors 537 MBmodel-00073-of-00132.safetensors 537 MBmodel-00086-of-00132.safetensors 537 MBmodel-00099-of-00132.safetensors 537 MBmodel-00090-of-00132.safetensors 537 MBmodel-00103-of-00132.safetensors 537 MBmodel-00003-of-00132.safetensors 537 MBmodel-00018-of-00132.safetensors 537 MBmodel-00031-of-00132.safetensors 537 MBmodel-00044-of-00132.safetensors 537 MBmodel-00057-of-00132.safetensors 537 MBmodel-00070-of-00132.safetensors 537 MBmodel-00083-of-00132.safetensors 537 MBmodel-00096-of-00132.safetensors 537 MBmodel-00109-of-00132.safetensors 537 MBmodel-00012-of-00132.safetensors 537 MBmodel-00025-of-00132.safetensors 537 MBmodel-00038-of-00132.safetensors 537 MBmodel-00051-of-00132.safetensors 537 MBmodel-00064-of-00132.safetensors 537 MBmodel-00054-of-00132.safetensors 537 MBmodel-00067-of-00132.safetensors 537 MBmodel-00015-of-00132.safetensors 537 MBmodel-00028-of-00132.safetensors 537 MBmodel-00080-of-00132.safetensors 537 MBmodel-00093-of-00132.safetensors 537 MBmodel-00106-of-00132.safetensors 537 MBmodel-00123-of-00132.safetensors 537 MBmodel-00006-of-00132.safetensors 537 MBmodel-00113-of-00132.safetensors 537 MBmodel-00126-of-00132.safetensors 537 MBmodel-00040-of-00132.safetensors 537 MBmodel-00076-of-00132.safetensors 537 MBmodel-00115-of-00132.safetensors 537 MBmodel-00118-of-00132.safetensors 537 MBmodel-00128-of-00132.safetensors 537 MBmodel-00112-of-00132.safetensors 537 MBmodel-00125-of-00132.safetensors 537 MBmodel-00116-of-00132.safetensors 537 MBmodel-00129-of-00132.safetensors 537 MBmodel-00005-of-00132.safetensors 537 MBmodel-00122-of-00132.safetensors 537 MBmodel-00077-of-00132.safetensors 537 MBmodel-00041-of-00132.safetensors 537 MBmodel-00119-of-00132.safetensors 537 MBmodel-00002-of-00132.safetensors 537 MBmodel-00001-of-00132.safetensors 533 MBmodel-00013-of-00132.safetensors 532 MBmodel-00026-of-00132.safetensors 532 MBmodel-00039-of-00132.safetensors 532 MBmodel-00056-of-00132.safetensors 532 MBmodel-00069-of-00132.safetensors 532 MBmodel-00085-of-00132.safetensors 532 MBmodel-00098-of-00132.safetensors 532 MBmodel-00078-of-00132.safetensors 532 MBmodel-00111-of-00132.safetensors 532 MBmodel-00124-of-00132.safetensors 532 MBmodel-00016-of-00132.safetensors 519 MBmodel-00029-of-00132.safetensors 507 MBmodel-00081-of-00132.safetensors 507 MBmodel-00094-of-00132.safetensors 507 MBmodel-00107-of-00132.safetensors 507 MBmodel-00042-of-00132.safetensors 507 MBmodel-00120-of-00132.safetensors 507 MBmodel-00053-of-00132.safetensors 499 MBmodel-00066-of-00132.safetensors 499 MBmodel-00052-of-00132.safetensors 473 MBmodel-00065-of-00132.safetensors 473 MBmodel-00017-of-00132.safetensors 466 MBmodel-00030-of-00132.safetensors 466 MBmodel-00082-of-00132.safetensors 466 MBmodel-00095-of-00132.safetensors 466 MBmodel-00108-of-00132.safetensors 466 MBmodel-00043-of-00132.safetensors 466 MBmodel-00121-of-00132.safetensors 466 MBmodel-00091-of-00132.safetensors 465 MBmodel-00104-of-00132.safetensors 465 MBmodel-00010-of-00132.safetensors 465 MBmodel-00023-of-00132.safetensors 465 MBmodel-00036-of-00132.safetensors 465 MBmodel-00049-of-00132.safetensors 465 MBmodel-00062-of-00132.safetensors 465 MBmodel-00088-of-00132.safetensors 465 MBmodel-00101-of-00132.safetensors 465 MBmodel-00075-of-00132.safetensors 465 MBmodel-00020-of-00132.safetensors 465 MBmodel-00033-of-00132.safetensors 465 MBmodel-00046-of-00132.safetensors 465 MBmodel-00059-of-00132.safetensors 465 MBmodel-00072-of-00132.safetensors 465 MBmodel-00007-of-00132.safetensors 465 MBmodel-00117-of-00132.safetensors 465 MBmodel-00114-of-00132.safetensors 465 MBmodel-00127-of-00132.safetensors 465 MBmodel-00004-of-00132.safetensors 465 MBmodel-00130-of-00132.safetensors 427 MBQwen/Qwen3.5-35B-A3B