九问ScienceDiscovery实现树搜索驱动RSI，加速科学发现，小时级写出通用积分器，低成本找出物理科学规律
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - openJiuwen’s open-source ScienceDiscovery platform uses tree search to autonomously evolve and evaluate scientific code with off-the-shelf LLMs, without model training or parameter tuning. It reports gains across numerical integration, code optimization, and symbolic equation discovery at relatively low time and inference cost.
- Each iteration selects a prior artifact, asks an LLM to rewrite it, evaluates it in a sandbox, and updates ancestor scores; revisiting older branches balances exploitation and exploration.
- A 236-version, two-hour search produced a general oscillatory-integral solver that achieved 0.07% average relative error on 19 scoring problems and generalized to 19 held-out problems.
- On AlgoTune’s 154 numerical tasks, two runs averaged a reported 2.279× speedup; a separate 48-expansion search improved hypergeometric-function accuracy on 1,000 unseen inputs.
- On LLM-SRBench symbolic regression, the system recovered correct equations for 41.4% of 111 tasks using an average of 16.5 model calls per task.
Sources (1)
九问ScienceDiscovery实现树搜索驱动RSI，加速科学发现，小时级写出通用积分器，低成本找出物理科学规律
TL;DR - openJiuwen’s open-source ScienceDiscovery platform uses tree search to autonomously evolve and evaluate scientific code with off-the-shelf LLMs, without model training or parameter tuning. It reports gains across numerical integration, code optimization, and symbolic equation discovery at relatively low time and inference cost.
- Each iteration selects a prior artifact, asks an LLM to rewrite it, evaluates it in a sandbox, and updates ancestor scores; revisiting older branches balances exploitation and exploration.
- A 236-version, two-hour search produced a general oscillatory-integral solver that achieved 0.07% average relative error on 19 scoring problems and generalized to 19 held-out problems.
- On AlgoTune’s 154 numerical tasks, two runs averaged a reported 2.279× speedup; a separate 48-expansion search improved hypergeometric-function accuracy on 1,000 unseen inputs.
- On LLM-SRBench symbolic regression, the system recovered correct equations for 41.4% of 111 tasks using an average of 16.5 model calls per task.