ML Tea: Planning and Problem-Solving with General, Scalable Neuro-Symbolic Models

Speaker: Yongchao Chen

Title: Planning and Problem-Solving with General, Scalable Neuro-Symbolic Models

Abstract: Foundation models excel at large-scale learning but underuse their potential for search—critical for digital agents and physical robots requiring symbolic reasoning and optimization. Pure text-based reasoning in LLMs or waypoints-based TAMP in VLAs remain limited in both reliability and efficiency. To address this, I explore augmenting LLMs with symbolic computation through three approaches: integrating external planners and tools, steering models to generate code as planners, and unifying text, code, and tool-use modes. Using supervised fine-tuning, multi-stage curriculum learning with GRPO, and tool-augmented test-time scaling, our models CodeSteer and R1-Code-Interpreter—downloaded over 300k times on HuggingFace—demonstrate strong performance. Our TUMIX method further boosts Gemini-2.5-pro from 21.6 to 34.1 on the Humanity Last Exam (HLE) benchmark.

Speaker Bio: Yongchao is a fifth-year Ph.D. student in Electrical Engineering at Harvard SEAS and MIT LIDS. His research focuses on Neuro-Symbolic Foundation Models for Reasoning and Planning, advised by Prof. Chuchu Fan and Prof. Nicholas Roy at MIT, and co-advised by Prof. Na Li at Harvard. Yongchao has conducted research at Google Research and DeepMind, Microsoft Research, and the MIT-IBM Watson AI Lab. His work has been featured by MIT News Spotlight.