Zhicheng Cai蔡志成

Advancing High-performance GPU Kernels Generation with Large-scale Agentic RL System通过大规模 Agentic RL 系统推进高性能 GPU Kernel 生成

• A reliable sandbox with isolated workspaces and rubric-based GRM safeguards against reward hacking.可靠的沙箱环境，结合隔离工作空间与基于评分准则的 GRM 防护，抵御奖励黑客行为。
• A stable and effective RL training framework, combining Ratchet-raw process rewards with a suite of stabilization techniques.稳定高效的强化学习训练框架，将 Ratchet-raw 过程奖励与一系列训练稳定化技术相结合。
• High-throughput infrastructure based on fully asynchronous rollouts and multi-token prediction.基于全异步 rollout 与多 token 预测的高吞吐基础设施。

Scaling Agentic RL to a 1T LLM with 300 turns, 256K context trained on 1,024 GPUs over 150 steps.在 1,024 张 GPU 上训练 150 步，将 Agentic RL 扩展至 1T LLM、300 轮交互与 256K 上下文。

One of the world's largest-scale RL-trained agents.世界上规模最大的 RL 训练智能体之一。

Achieving world-leading GPU kernel generation performance取得世界领先的 GPU Kernel 生成性能

Accomplish in two hours the kernel optimization that takes human experts several days.仅用两小时，完成需要人类专家数天才能完成的 Kernel 优化。

FLEX (Forward Learning from Experience) introduces a new learning paradigm for AIFLEX（Forward Learning from Experience）提出 AI 全新学习范式

It enables agents to continuously self-evolve through experience, fundamentally overturning the traditional model-training framework based on gradient backpropagation and taking a step toward artificial general intelligence (AGI).让智能体通过经验持续自我进化，突破依赖梯度反向传播的传统模型训练框架，向通用人工智能（AGI）迈进一步。

Continuous self-evolution for performance enhancement during deployment在部署中持续自进化，提升任务表现

• Olympiad-level mathematics: enhance agent performance from 40% to 63% on AIME25.奥赛级数学：在 AIME25 上将智能体表现从 40% 提升至 63%。
• Chemical retrosynthesis: enhance agent performance from 20% to 30% on USPTO50k.化学逆合成：在 USPTO50k 上将智能体表现从 20% 提升至 30%。
• Protein fitness prediction: enhance agent performance from 46% to 60% on ProteinGym.蛋白质适应度预测：在 ProteinGym 上将智能体表现从 46% 提升至 60%。
• Substantial improvements in reasoning and knowledge utilization.显著提升推理与知识运用能力。

Knowledge transfer across agents to build collective intelligence跨智能体知识迁移，构建群体智能

• The experience library is portable and reusable. It can be seamlessly transferred across different agents, enabling two powerful effects that transcend traditional model boundaries.经验库具备可迁移、可复用的特性，可在不同智能体之间无缝传递，带来超越传统模型边界的两种强大效应。
• These inheritance properties open a path to universal experience libraries: reusable knowledge modules that enhance entire AI ecosystems without individual training, enabling a collaborative intelligence paradigm.这种继承能力为通用经验库开辟道路：通过可复用的知识模块，无需逐一训练即可增强整个 AI 生态，形成协作智能范式。