yuuxia/acts-controller

1---2license: apache-2.03library_name: transformers4pipeline_tag: text-generation5tags:6- reasoning7- agent8- chain-of-thought9---10 11# ACTS: Agentic Chain-of-Thought Steering Controller12 13This 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](https://huggingface.co/papers/2606.03965).14 15ACTS 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.16 17## Resources18- **Paper:** [Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning](https://huggingface.co/papers/2606.03965)19- **Repository:** [Andree-9/ACTS](https://github.com/Andree-9/ACTS)20- **SFT Data:** [yuuxia/controller-sft-data](https://huggingface.co/datasets/yuuxia/controller-sft-data)21 22## Quick Start Inference23 24To use this controller to steer a reasoner, follow the setup instructions in the [GitHub repository](https://github.com/Andree-9/ACTS) and run the following command:25 26```bash27conda activate slime28./scripts/run_acts_inference.sh \29    --controller yuuxia/acts-controller \30    --reasoner   deepseek-ai/DeepSeek-R1-Distill-Qwen-7B \31    --benchmark  aime2024 \32    --budget     1000033```34 35## Citation36 37```bibtex38@misc{xia2026acts,39      title={Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning},40      author={Yu Xia and Zhouhang Xie and Xin Xu and Byungkyu Kang and Prarit Lamba and Xiang Gao and Julian McAuley},41      year={2026},42      eprint={2606.03965},43      archivePrefix={arXiv},44      primaryClass={cs.CL},45      url={https://arxiv.org/abs/2606.03965},46}47```