What makes Metastone-S1 the main reflective generative model for AI reasoning?

Researchers from Metastone-Ai and USTC introduce a reflective generative model, Metastone-S1, which achieves the performance of Openai O3-Mini thanks to a new reflective generative form.

Key innovations

Reflective

• Modeling of policy and unified awards: Metastone-S1 incorporates the policy model (to generate reasoning trajectories) and the processes' reward model (PRM) at a step in a single architecture, using shared parameters. This implementation requires only a light addition (as little as 53 m of parameters for the verifier in the main model 32b), considerably reducing calculation costs compared to conventional autonomous PRMs.
• Self-supervised process reward (SPRM): SPRM eliminates the need for data labeled in terms of expensive process. It takes advantage of a self-supervised loss function which only uses the accuracy of the final response to judge the quality of the intermediate reasoning stages, supported by a dynamic weighting mechanism to filter the noisy labels.

Redefined test scale (TTS)

Traditional LLMs often improve via settings parameters during training. Metastone -S1 adopts a separate approach – TTS – by increasing the performance of inference thanks to an increased calculation depth rather than a simple increase in the size of the model:

• Internal tts: Extends the chain of thoughts for a deeper sequential problem solving, but can cause substantial calculation costs.
• External tts: Generates several paths of reasoning in parallel and selects the best PRMs. This generally requires additional models and separate labeling.
• Metastone-S1 approach: Combines the two paradigms into a single architecture, offering an effective and precise selection of trajectory with a minimum of additional resources.

Performance and comparative analysis

Metastone-S1 is available in three sizes (parameters 1.5b, 7b and 32b). The largest, Metastone-S1-32B, corresponds or surpasses the main proprietary and open source models, including Openai O3-Mini, on key reasoning and mathematics references.

Each size demonstrates high scaling properties and effective use of parameters. For example, Metastone-S1-1.5B surpasses comparable size models on mathematical tasks, while sizes 7B and 32B effectively evolve with TTS capacity and strategy.

Efficiency and “moment aha”

• Minimum general costs: The integration of the SPRM just adds a fraction of parameters compared to traditional PRMs (for example, 26m vs 72b), which gives advanced results between tasks.
• Aha moment: The analysis of the formation reveals a distinct point where the model begins to clearly mark the correct reasoning paths compared to incorrect reasoning paths, leading to improved discrimination and final performance.
• Scaling law: Metastone -S1's performance increases logarithmically with the calculation budget (size of the model × reasoning tokens), tray around the sampling of the best of 32 – an effective compromise for deployment.

Flexible reasoning modes

To balance the performance and use of resources, Metastone-S1 offers three TTS inference modes:

• Bottom (k = 2): The fastest inference for quick responses.
• Middle (K = 8): Better precision with a moderate calculation.
• High (k = 32): Maximum depth for difficult tasks.

Conclusion

With its new reflective generative structure, Metastone-S1 unifies problem solving and the verification of the solution in a single effective framework. By achieving the performance of Openai O3-Mini with much less resources, he demonstrates that innovation in LLM architecture can compete with a gross-source scaling-opening up new avenues for the advancement of AI reasoning and accessibility

Asif Razzaq is the CEO of Marktechpost Media Inc .. as a visionary entrepreneur and engineer, AIF undertakes to exploit the potential of artificial intelligence for social good. His most recent company is the launch of an artificial intelligence media platform, Marktechpost, which stands out from its in-depth coverage of automatic learning and in-depth learning news which are both technically solid and easily understandable by a large audience. The platform has more than 2 million monthly views, illustrating its popularity with the public.