EdotEnv是由Rui和Michael创建的YC S26初创公司，推出了一个基于量化交易工作流的自进化强化学习环境[1]。该平台旨在训练和评估大语言模型的研究能力，利用金融市场的自然演进特性作为基准[1]。
该环境采用真实市场数据而非合成数据[1]，包含特征构建、投资组合设计和策略回测等多个任务步骤[1]。通过这些任务，平台可以评估模型的长期规划和持续学习能力[1]。EdotEnv指出，现有最先进的模型存在倾向于广泛浅层搜索而非深度迭代、高阶推理能力无法提升性能，以及缺乏交易理解能力等问题[1]。
该公司已开源示例任务库，目标客户为AI实验室、研究人员和企业，供其用于训练自有Agent[1]。
EdotEnv, a Y Combinator S26 startup founded by Rui and Michael, has unveiled a reinforcement learning environment designed to train and assess the research abilities of large language models through quantitative trading workflows.[1] The platform leverages real market data and the natural evolution of financial markets as a benchmark, enabling LLMs to complete tasks including feature engineering, portfolio design, and strategy backtesting to evaluate their long-term planning and continuous learning capabilities.[1]
The startup has identified several limitations in state-of-the-art models when applied to this domain: they tend toward broad, shallow searches rather than deep iterative exploration, their advanced reasoning capabilities fail to improve performance in these tasks, and they lack genuine understanding of trading mechanics.[1] EdotEnv's environment encompasses multiple task stages, including feature construction, backtesting tools, and execution tools, positioning itself as a training and evaluation platform for AI laboratories, researchers, and enterprises developing their own agents.[1] The company has already open-sourced sample task libraries to support adoption.[1]