Overview
We fine-tuned Qwen3-30B-A3B-Thinking with Agent Context Compilation (ACC) — a method that converts multi-turn agent trajectories (Search, SWE, SQL) into long-context QA pairs for direct supervised fine-tuning. Unlike standard agent SFT that masks tool responses, ACC assembles scattered evidence across turns into a single context, enabling explicit supervision of long-range dependency modeling.
Table with columns: Benchmark, Score, Δ vs Base| Benchmark | Score | Δ vs Base |
|---|
| MRCR | 68.28 | +18.09 |
| GraphWalks | 77.51 | +7.59 |
| GPQA-Diamond | 70.20 | +2.49 |
| MMLU-Pro | 76.00 | +1.50 |
Results on MRCR and GraphWalks are comparable to Qwen3-235B-A22B despite ~8× fewer active parameters. General capabilities are preserved.
Training Data
- Dataset: groundhogLLM/ACC-dataset
- Size: 10,802 compiled trajectories (Search: 3,369; SWE: 4,368; SQL: 3,065)
- Context length: 2K – 128K tokens
- Training seq length: 131,072 tokens
- Epochs: 4
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "groundhogLLM/ACC-Qwen3-30B-A3B"
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
messages = [{"role": "user", "content": "Your long-context question here..."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0]))
Citation
If you use this model, please cite:
@misc{su2026acccompilingagenttrajectories,
title={ACC: Compiling Agent Trajectories for Long-Context Training},
author={Qisheng Su and Zhen Fang and Shiting Huang and Yu Zeng and Yiming Zhao and Kou Shi and Ziao Zhang and Lin Chen and Zehui Chen and Lijun Wu and Feng Zhao},
year={2026},
eprint={2605.21850},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2605.21850},
}
