Satori: Reinforcement Learning with Chain-of-Action-Thought Enhances LLM Reasoning via Autoregressive Search
FreeFreeFree tier
Inputs: textOutputs: text
About Satori: Reinforcement Learning with Chain-of-Action-Thought Enhances LLM Reasoning via Autoregressive Search
Satori is a 7B parameter large language model designed to enhance reasoning capabilities through autoregressive search, enabled by a novel Chain-of-Action-Thought (COAT) reasoning framework. It employs a two-stage post-training paradigm: format tuning to internalize COAT reasoning, followed by large-scale self-improvement via reinforcement learning. Trained on open-source models and data, Satori achieves state-of-the-art performance on mathematical reasoning benchmarks and demonstrates strong generalization to out-of-domain tasks. All code, data, and models are fully open-sourced.
Key Features
Chain-of-Action-Thought (COAT) reasoning framework for structured step-by-step reasoning
Two-stage training: format tuning then reinforcement learning self-improvement
Autoregressive search at inference time for self-reflection and strategy exploration
7B parameter model trained on open-source models and data
Fully open-sourced code, data, and models
State-of-the-art performance on mathematical reasoning benchmarks
Strong generalization to out-of-domain tasks
Pros & Cons
Pros
- Achieves state-of-the-art results on math reasoning benchmarks
- Fully open-source, enabling reproducibility and community extension
- Novel training paradigm that internalizes search capabilities without external verifiers
- Demonstrates strong generalization to out-of-domain tasks
- Efficient 7B parameter size balances performance and computational requirements
Cons
- Primarily validated on mathematical reasoning; performance on other domains less established
- Requires two-stage training pipeline which may be resource-intensive for smaller teams
- Limited documentation beyond the research paper; production use may need additional engineering
Best For
Mathematical reasoning and problem solvingScientific reasoning and theorem provingComplex logical deduction tasksOut-of-domain reasoning generalization evaluation
FAQ
What is Satori?
Satori is a 7B parameter open-source LLM that enhances reasoning via Chain-of-Action-Thought (COAT) and reinforcement learning, achieving state-of-the-art on math benchmarks.
Is Satori open-source?
Yes, the code, data, and models are fully open-sourced as stated in the paper.
What makes Satori different from other reasoning LLMs?
Satori introduces Chain-of-Action-Thought (COAT) reasoning and a two-stage training paradigm (format tuning then RL self-improvement) to internalize autoregressive search capabilities within a single model.
What benchmarks does Satori perform well on?
Satori achieves state-of-the-art performance on mathematical reasoning benchmarks and shows strong generalization to out-of-domain tasks.
How can I access Satori?
The paper indicates code, data, and models are fully open-sourced. Links are available via the arXiv page or associated GitHub repository.