My long-term research goal is recursive self-improvement (RSI): agents that learn from experience to improve their capabilities and their ability to learn and improve. I am interested in the loop between task design, environments, evaluation, and updates to model weights, memory, tools, and harnesses. One question within this goal is learning from extremely delayed real-world rewards: how can agents learn and provide credible evidence of improvement when true outcomes are unavailable in time and historical supervision risks leaking future information?

我的长期研究目标是实现递归式自我改进（RSI）：让智能体从经验中提升能力，并进一步提升自身学习与改进的能力。我关注任务设计、环境、评价，以及模型参数、记忆、工具与 harness 的更新如何形成闭环。其中一个问题是真实世界超级延迟奖励下的学习：当真实标签来不及用于训练、历史监督又容易泄漏未来信息时，如何学习并提供可信的改进证据？

One possible route is learning in accelerated simulations that preserve the mechanisms relevant to our decisions, then testing whether the resulting improvements transfer to the real world.

一条可能的路径是在可加速的模拟环境中学习：保留影响决策的关键机制，再检验其中学到的改进能否迁移到现实。