LLM Agent Research Scientist 大模型智能体研究科学家

About the Role

As an LLM Agent Research Scientist, you'll define and lead research that pushes the limits of what LLM-powered agents can do. You'll design agent architectures, reasoning and planning methods, tool-use and multi-agent frameworks, and the post-training recipes (including RL) that make agents capable and dependable. You'll also train models with high-end GPUs on large, high-quality datasets. You'll sit at the intersection of frontier research and product, directly shaping how LLM agents transform the way people create and get work done in Canva.

作为大模型智能体研究科学家，你将定义并主导不断突破智能体能力上限的研究。你将设计智能体架构、推理与规划方法、工具调用与多智能体框架，以及让智能体更强大、更可靠的后训练方案（包括 RL）。你也会在大规模、高质量的数据集上，用高端 GPU 训练模型。你将处在前沿研究与产品的交汇点，直接影响大模型智能体如何改变人们在 Canva 上创作与工作的方式。

At the moment, this role is focused on

• Research Direction: Determining and leading agent research initiatives aligned with the frontier of LLM/agent research and Canva's product goals.

• Model Innovation: Creating novel agent capabilities through post-training and RL—developing reward modeling, synthetic data, and environment design—and running experiments with high-end GPUs on large, high-quality datasets to validate hypotheses and translate results into shippable improvements.

• Harness Innovation: Designing and building the agent harness that lets agents execute complex, long-horizon tasks reliably in production, and exploring harness techniques for self-improvement.

• Evaluation: Creating rigorous benchmarks and evaluation methods for agent correctness, safety, robustness, and real-world task success.

• Product Impact: Partnering closely with engineering, product, design, and data teams to land frontier research capabilities in real Canva product experiences.

• 研究方向：定义并主导契合大模型/智能体研究前沿与 Canva 产品目标的研究方向。

• 模型创新：通过后训练和 RL 打造全新的智能体能力——涵盖 reward modeling、合成数据、环境设计等——并在大规模、高质量数据集上用高端 GPU 开展实验，验证假设，把结果转化为可落地的改进。

• Harness 创新：设计并构建 agent harness，让智能体能在生产环境中稳定地完成复杂的长链路任务，并探索让智能体自我改进的 harness 技术。

• 评测：围绕智能体的正确性、安全性、鲁棒性和真实任务完成情况，构建严谨的基准与评测方法。

• 产品影响：与工程、产品、设计和数据团队紧密协作，将前沿研究能力落地到真实的 Canva 产品体验中。