XLearning-SCU/HERO-Trajectories

HERO Evaluation Trajectories

Doing More with Less Tokens: Hierarchical Reinforcement Learning for Efficient Coding Agents

🔍 Overview

HERO is a hierarchical reinforcement learning method for training token-efficient coding agents while prioritizing task resolution. This release contains 2,400 evaluation trajectories from HERO-4B, HERO-9B, and HERO-35B-A3B on SWE-bench Verified and SWE-bench Multilingual, using Claude Code as the agent scaffold.

Results

Each model includes 500 Verified tasks and 300 Multilingual tasks. These are individual recorded runs; the paper reports averages over three random seeds.

Benchmark	Model	Tasks	Resolved Rate	Token Usage (M)
Verified	HERO-4B	500	43.2%	3.5
Verified	HERO-9B	500	50.6%	2.8
Verified	HERO-35B-A3B	500	61.4%	2.4
Multilingual	HERO-4B	300	29.3%	4.2
Multilingual	HERO-9B	300	40.0%	3.6
Multilingual	HERO-35B-A3B	300	46.0%	2.7

Tokens are averages per task in millions (M), summed over agent calls, including repeated input context. Records with missing or zero usage are excluded from token averages; resolve rates include all tasks. Detailed usage and processing notes are in each summary.json.

Files

Each <benchmark>/<model>/ directory contains:

File	Contents
results.jsonl	Task outcomes, grading reports, token usage, and trajectory paths
predictions.jsonl	Predicted patches
summary.json	Aggregate results, evaluation settings, and processing notes
trajectories.tar.gz	Per-task JSON trajectories in hero-trace-v1 format

Loading the Traces

Extract an archive from the repository root:

tar -xzf verified/HERO-4B/trajectories.tar.gz -C verified/HERO-4B

Load an extracted trajectory:

import json
from pathlib import Path

path = next(Path("verified/HERO-4B/trajectories").glob("*.json"))
with path.open() as f:
    trace = json.load(f)
calls = trace["trajectory"]

The trajectory_file paths in results.jsonl are relative to the model directory.

Run Notes

The existing 35B run summaries report different checkpoints for Verified and Multilingual; checkpoint correspondence with the released weights remains unconfirmed. Some prediction patches were sanitized without re-evaluation, so resolution labels describe the original runs. The 35B Multilingual token statistics exclude auxiliary token-counting calls, as documented in its summary.json.

Source Licenses

SWE-bench Verified, SWE-bench Multilingual, and their source repositories retain their respective licenses. Please also cite the corresponding benchmarks when using these traces.

📖 Citation

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}
}