yuuxia/acts-controller

ACTS: Agentic Chain-of-Thought Steering Controller

This repository contains a controller agent checkpoint for ACTS (Agentic Chain-of-Thought Steering), presented in the paper Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning.

ACTS is a framework where a lightweight controller agent adaptively steers a frozen reasoner (such as DeepSeek-R1) step-by-step under a thinking-token budget. By formulating reasoning steering as a Markov decision process, the controller chooses a reasoning strategy and a short steering phrase at each step to enable controllable accuracy–efficiency trade-offs.

Resources

• —Paper: Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning
• —Repository: Andree-9/ACTS
• —SFT Data: yuuxia/controller-sft-data

Quick Start Inference

To use this controller to steer a reasoner, follow the setup instructions in the GitHub repository and run the following command:

conda activate slime
./scripts/run_acts_inference.sh \
    --controller yuuxia/acts-controller \
    --reasoner   deepseek-ai/DeepSeek-R1-Distill-Qwen-7B \
    --benchmark  aime2024 \
    --budget     10000

Citation

@misc{xia2026acts,
      title={Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning},
      author={Yu Xia and Zhouhang Xie and Xin Xu and Byungkyu Kang and Prarit Lamba and Xiang Gao and Julian McAuley},
      year={2026},
      eprint={2606.03965},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.03965},
}