What is MetaStone-S1
MetaStone-S1 is a reflective generative large model launched by Yuanshi Technology, integrating deep reasoning and reasoning chain self-screening capabilities for the first time. The model's core uses a self-supervised reflection paradigm, based on a shared-backbone policy model and a process scoring model (SPRM), adding only 53M parameters to evaluate the quality of reasoning steps in real time without manual annotation. The model supports Long-CoT reinforcement learning, generating ultra-long reasoning chains, and surpasses similar models in mathematics (AIME), code (LiveCodeBench), and Chinese reasoning (C-EVAL) tasks. It open-sources 1.5B, 7B, and 32B versions, achieving high performance with low inference cost and advancing reasoning intelligence toward a new stage of "self-correction".
Main Features of MetaStone-S1
- Deep Reasoning Generation: MetaStone-S1 can generate ultra-long and complex reasoning chains (Long-CoT), especially suitable for solving high-difficulty reasoning tasks such as mathematical proofs and programming algorithms.
- Intelligent Reasoning Chain Optimization: The model has a built-in self-supervised process scoring mechanism (SPRM), supporting automatic identification and elimination of erroneous steps in the reasoning process, significantly improving the accuracy of the final answer.
- Multi-level Reasoning Modes: Provides three working modes: Low (fast response), Medium (balanced accuracy and speed), and High (deep thinking), meeting the reasoning needs of different scenarios.
- Open-source and Extensible Features: Fully open-sources three scale models 1.5B/7B/32B and accompanying tools, supporting developers to further optimize the model's reasoning capabilities in specific domains.
Technical Principles of MetaStone-S1
- Dual-head Shared Architecture: Based on the design of a policy model (Policy Model) and a process scoring model (SPRM) sharing the backbone network, a generation head (Generation Head) and a scoring head (Scoring Head) are deployed in parallel on the Transformer layer. The former is responsible for generating reasoning chains, and the latter scores each reasoning step in real time based on self-supervised learning.
- Self-supervised Process Reward: Proposes the SPR Loss (Self-supervised Process Reward Loss) algorithm, using the correctness of the final answer as a weak supervision signal, and based on a noise filtering mechanism automatically generates step-level pseudo-labels to train the process scoring model, getting rid of dependence on manual annotation.
- Dynamic Reasoning Selection: In the inference stage, uses Test-Time Scaling technology: first generates multiple candidate reasoning chains (e.g., High mode generates 32), uses SPRM to calculate the total score of the paths, and finally selects the optimal path to continue generation, forming a closed loop of "generation-evaluation-selection".
- Joint Optimization Mechanism: Based on the GRPO reinforcement learning algorithm, simultaneously optimizes the policy model and SPRM, where the policy model maximizes answer correctness, and SPRM uses contrastive learning to distinguish high-quality/low-quality reasoning steps. The two share gradients to form collaborative evolution.
- Emergent Ability Regulation: Designs a Scaling Law for thinking length and model performance, based on adjusting the number of rollouts to control computation (parameter count × thinking token count), achieving a smooth transition from fast response (Low) to deep thinking (High).
MetaStone-S1 Project Address
- GitHub Repository：https://github.com/MetaStone-AI/MetaStone-S1
- HuggingFace Model Library：https://huggingface.co/MetaStoneTec
- arXiv Technical Paper：https://arxiv.org/pdf/2507.00195
Application Scenarios of MetaStone-S1
- Educational Intelligence: As an "AI tutor", it accurately solves mathematics/physics competition problems and generates interactive explanations of problem-solving paths.
- Legal Intelligence Field: In-depth analysis of the logical relationships among contract clauses, precise identification of potential legal risk points, and provision of revision suggestions that conform to legal logic.
- Intelligent manufacturing field: Based on multi-level causal reasoning, quickly locate the root causes of industrial equipment failures, generate optimal maintenance plans, and significantly improve production efficiency.
- Academic writing field: Supports formula derivation and theoretical verification in research papers, ensuring the logical rigor of academic content.