MetaStone-S1 - A reflective generative large-scale model launched by MetaStone Technology

What is MetaStone-S1?

MetaStone-S1 is a reflective generative large-scale model launched by MetaStone Technology, which for the first time integrates deep reasoning with the ability to self-select reasoning chains. The core of the model uses a self-supervised reflective paradigm, based on a policy model and process scoring model (SPRM) with a shared backbone. It can evaluate the quality of reasoning steps in real time with only 53M additional parameters, without the need for manual annotation. The model supports Long-CoT reinforcement learning to generate ultra-long reasoning chains, outperforming similar models in mathematical (AIME), coding (LiveCodeBench), and Chinese reasoning (C-EVAL) tasks. It is open-sourced in versions 1.5B, 7B, and 32, achieving high performance with low inference costs and pushing reasoning intelligence towards a new stage of "self-correction."

Main functions of MetaStone-S1

• Deep reasoning generation functionMetaStone-S1 can generate extremely long and complex reasoning chains (Long-CoT), making it particularly suitable for solving challenging reasoning tasks such as mathematical proofs and programming algorithms.
• Intelligent inference chain optimizationThe model incorporates a self-supervised process scoring mechanism (SPRM), which supports the automatic identification and elimination of erroneous steps in the reasoning process, significantly improving the accuracy of the final answer.
• Multi-level reasoning modeIt offers three working modes: Low (fast response), Medium (balancing accuracy and speed), and High (deep thinking) to meet the reasoning needs of different scenarios.
• Open source and scalabilityIt offers fully open-source models in three sizes: 1.5B, 7B, and 32B, along with supporting tools, enabling developers to further optimize the model's inference capabilities in specific domains.

MetaStone-S1 Technical Principles

• Dual-head shared architectureBased on the design of a shared backbone network between the Policy Model and the Process Scoring Model (SPRM), a Generation Head and a Scoring Head are deployed in parallel on the Transformer layer. The former is responsible for generating the inference chain, while the latter scores each inference step in real time based on self-supervised learning.
• Self-monitoring process rewardsWe propose the SPR Loss (Self-supervised Process Reward Loss) algorithm, which uses the correctness of the final answer as a weak supervision signal and automatically generates step-level pseudo-labels based on a noise filtering mechanism to train the process scoring model and get rid of the dependence on manual annotation.
• Dynamic reasoning for optimal selectionIn the inference phase, Test-Time Scaling technology is used: First, multiple candidate inference chains are generated (e.g., 32 chains are generated in High mode), the total path score is calculated using SPRM, and finally the optimal path is selected to continue generating, forming a closed loop of "generation-evaluation-selection".
• Joint optimization mechanismBased on the GRPO reinforcement learning algorithm, the policy model and SPRM are optimized simultaneously. The policy model maximizes the answer accuracy, while the SPRM uses contrastive learning to distinguish between high-quality and low-quality inference steps. The two share gradients and form a co-evolution.
• Emergent capability regulationThe Scaling Law of Design Thinking Length and Model Performance controls the computational load (parameter count × number of thinking tokens) by adjusting the number of rollouts, 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 libraryhttps://huggingface.co/MetaStoneTec
• arXiv technical paperhttps://arxiv.org/pdf/2507.00195

Application scenarios of MetaStone-S1

• Educational intelligenceAs an "AI tutor," it accurately answers math/physics competition questions and generates interactive explanations of the solution path.
• Legal Intelligence FieldWe provide in-depth analysis of the logical relationships between contract terms, accurately identify potential legal risks, and offer revision suggestions that align with legal logic.
• Intelligent manufacturing fieldBased on multi-level causal reasoning, it can quickly locate the root cause of industrial equipment failures, generate the optimal maintenance plan, and significantly improve production efficiency.
• Academic writing fieldIt supports the derivation of formulas and the verification of theories in scientific research papers, ensuring the logical rigor of academic content.