研究：赋予研究智能体完全自主性可提升数学发现能力
Does giving research agents full autonomy add capability
Does giving research agents full autonomy add capability
赋予研究代理完全自主权是否能增加能力
Cambridge, Hong Kong Univ and other labs find that freedom to choose directions yields mathematical proofs and generalizations no evaluator asked for.
剑桥、香港大学及其他实验室发现，让代理自由选择方向，能产生评估者未曾要求的数学证明和推广。
Right now the industry consensus is: models are unreliable, so wrap them in scaffolding. Fixed pipelines, orchestrators, assigned roles, tight evaluators, human-designed search loops.
目前行业共识是：模型不可靠，因此要用脚手架包裹它们。固定流水线、编排器、分配角色、严格评估器、人工设计的搜索循环。
AlphaEvolve is the strongest version of that, an evolutionary search loop where the model is a mutation operator inside machinery a human built.
AlphaEvolve 是该共识的最强版本，一种进化搜索循环，其中模型作为人类构建的机器中的变异算子。
This paper is kind of saying that consensus is now wrong, or at least expiring. Their claim is that current models are good enough to be treated as researchers rather than as components, and that the scaffolding is costing you more than it buys.
这篇论文似乎在说，这一共识现在已错误，或至少正在失效。他们声称，当前模型已足够优秀，应被视为研究人员而非组件，而脚手架的成本大于其收益。
The gain, as per this paper, appears to come from agents spending their budget on structure rather than search, which helps where a theorem can prune the space and not where the best construction is irregular.
根据这篇论文，增益似乎来自代理将预算花在结构而非搜索上，这有助于定理能剪枝空间的情况，而非最佳构造不规则的情况。
– arxiv. org/abs/2608.23691
– arxiv.org/abs/2608.23691
Title: "Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment"
标题：“开放世界多代理环境中的自主数学发现”
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力