Intern-S2
A scientific multimodal large model developed by Shanghai AI Laboratory. It possesses large-scale recursive reasoning and human-computer interaction capabilities, is aimed at frontier scientific research scenarios such as mathematics, and supports long-horizon rigorous reasoning.
Timeline (last 90 days)
Xu Disheng's team, relying on the Scholar Science multimodal large model S2 (Scholar-S2), completed the proof of the three-dimensional Ghys conjecture through human-machine collaboration.
In August 2026, Scholar-S2, which had taken shape through internal iteration, and its supporting systems possessed large-scale recursive reasoning capability and highly flexible human-machine interaction capability, and the first round of reasoning produced a 400-page derivation report.
The research team adhered to the technical route of integrating general and specialized capabilities, broke through the asynchronous long-horizon reasoning reinforcement learning training method, solved problems such as exponential accumulation of errors, long-term state maintenance, and self-correction, and developed a self-evolution algorithm that distills Harness capabilities back into the model.
On benchmarks such as IMO-Proof and AdvancedMathBench, Scholar-S2 is overall ahead of existing open-source models, and on some metrics it surpasses top closed-source models such as Gemini 3.1 Pro.
During the tackling process, Scholar-S2 successively completed 7 batches of continued deduction, the problem graph expanded from 15 nodes to 389 nodes, the total output reached over a thousand pages of rigorous derivation, and at the same time dozens of entirely new lines of argument were generated.
All Facts (5)
In August 2026, Scholar-S2, which had taken shape through internal iteration, and its supporting systems possessed large-scale recursive reasoning capability and highly flexible human-machine interaction capability, and the first round of reasoning produced a 400-page derivation report.
50%UnverifiedThe research team adhered to the technical route of integrating general and specialized capabilities, broke through the asynchronous long-horizon reasoning reinforcement learning training method, solved problems such as exponential accumulation of errors, long-term state maintenance, and self-correction, and developed a self-evolution algorithm that distills Harness capabilities back into the model.
50%UnverifiedOn benchmarks such as IMO-Proof and AdvancedMathBench, Scholar-S2 is overall ahead of existing open-source models, and on some metrics it surpasses top closed-source models such as Gemini 3.1 Pro.
50%UnverifiedXu Disheng's team, relying on the Scholar Science multimodal large model S2 (Scholar-S2), completed the proof of the three-dimensional Ghys conjecture through human-machine collaboration.
50%UnverifiedDuring the tackling process, Scholar-S2 successively completed 7 batches of continued deduction, the problem graph expanded from 15 nodes to 389 nodes, the total output reached over a thousand pages of rigorous derivation, and at the same time dozens of entirely new lines of argument were generated.
50%