STILL (Slow Thinking with LLMs) is a series of technical research efforts by the RUCAIBox team centered on slow-thinking reasoning systems for large language models, with the goal of matching o1-like reasoning systems. The project has open-sourced research reports, code, models, and training data on GitHub.
[Main Research Directions and Results]
- LLM Exploration Mechanism (August 2025): Systematically analyzes the exploration mechanisms of large models in RLVR, quantifies the differences between SFT and RL, reveals the entropy-performance dynamic relationship, and proposes strategies to improve exploration and training efficiency; the code has been open-sourced.
- Code-Integrated Reasoning (June 2025): Proposes an improved tool-augmented reinforcement learning method that balances exploration and stability, performs excellently on five mathematical benchmarks, and provides an in-depth analysis of this mechanism.
- ICPC-Eval (May 2025): A new benchmark based on 118 ICPC competition problems, evaluating the reasoning ability of large models in real competitive programming scenarios, proposing the iterative repair metric Refine@K, and open-sourcing the dataset, evaluation code, and paper.
- SimpleDeepSearcher (April 2025): Stimulates the autonomous web search ability of large models through knowledge distillation and self-distillation, filters 871 high-quality training samples, outperforms existing reinforcement learning methods on complex information retrieval tasks, and open-sources models and datasets (0.5k and 0.8k).
- STILL-3 (March 2025): An empirical study on stimulating and improving R1-like reasoning models, systematically experimenting with various factors affecting RL training, and exploring the improvement of reasoning performance through tool operation, with open-sourced code and training data.
- OlymMATH (March 2025): A benchmark containing 200 olympiad-level mathematics problems, covering algebra, geometry, number theory, and combinatorics, providing Chinese and English bilingual versions and difficulty grading, and open-sourcing the dataset, evaluation code, and paper.
- R1-Searcher (March 2025): A two-stage, outcome-based two-stage reinforcement learning method that enables large models to autonomously call external search systems during reasoning, without requiring process rewards or cold-start distillation, and open-sources models (Qwen-2.5-7B-Base-RL, Llama-3.1-8B-Instruct-RL), code, and data.
- Model release: STILL-3-Tool-32B (March 2025) utilizes Python code-assisted reasoning, AIME 2024 accuracy 81.70%; STILL-3-1.5B-preview (January 2025) is a 1.5-billion-parameter slow-thinking reasoning model, AIME accuracy 39.33%; Virgo (January 2025) is a multimodal slow-thinking reasoning model based on Qwen2-VL-72B-Instruct.
- Hallucination mitigation (January 2025): proposed the HaluSearch framework, combining tree search algorithms with a dynamic system switching mechanism, inspired by dual-process theory, to reduce large model hallucinations during reasoning.
- Early work: STILL-1 (November 2024) used reward-guided tree search to assist reasoning; STILL-2 (December 2024) preliminarily reproduced the slow-thinking reasoning system and released a technical report.
[How to Use] Research materials show that the project provides open-source code, models, training data, and evaluation benchmarks, some of which are released via GitHub and HuggingFace. Specific installation commands, dependency configuration, and running steps are not yet provided.
[Use Cases] Suitable for academic and engineering research scenarios such as large model reasoning research, reinforcement learning training, mathematics and code reasoning evaluation, information retrieval augmentation, and hallucination mitigation.
[Limitations] The project is positioned as technical research; the degree of productization, charging model, platform support, and detailed usage documentation are not yet provided.
