Research Scientist, Stitch

Job description

Define and pursue a long-term applied research agenda to overcome fundamental limitations of LLMs in long-horizon reasoning, planning, and tool use for complex software engineering tasks. Design and implement novel agentic architectures, exploring frontiers such as automated workflow optimization, multi-agent decomposition, and self-improvement. Explore techniques for scalable memory and context management that enable reasoning over large codebases (dynamic, structurally-aware context and memory). Lead research into synthetic data generation, verification methods/environments and rewards for building data and capability flywheels for Gemini improvement. Develop high-fidelity evaluation frameworks that measure agent impact to software quality and developer productivity. Minimum Qualifications: PhD in Computer Science, a related field, or equivalent practical experience. 2 years of experience leading a research agenda. Experience with modern ML frameworks such as JAX or PyTorch. Experience in one or more of the following areas: agentic AI (planning, tool use, memory), reinforcement learning (RLHF, RLVF, RLGF, offline RL). One or more scientific publication submissions for conferences, journals, or public repositories (such as CVPR, ICCV, NeurIPS, ICML, ICLR, etc.). Preferred Qualifications: 2 years of coding experience. 1 year of experience leading research efforts and influencing other researchers. Experience designing and implementing novel evaluation benchmarks for LLMs, generative code models or agentic systems. Experience with program analysis, program synthesis, automated program repair, formal methods, or designing developer tools. Experience building, training, and fine-tuning large language models.