Olive Song leads reinforcement learning research at MiniMax. She works on training and evaluating models for coding and agent tasks, where success depends on completing a sequence of actions rather than producing one plausible answer.[2][1]
Start with Intelligence with Everyone: RL at MiniMax →
GitHub 
X · olive_jy_song Career
MiniMax identifies Song as its research lead in reinforcement learning in a post about her conversation with Thomas Wolf at AI Engineer. The discussion covered sparse attention, training text and vision together from the beginning, and the reasons for releasing model weights. Her GitHub profile uses the name Jiayuan (Olive) Song and reports BA and MS study at NYU Courant.[2][3]
Contributions
Song presents MiniMax M2’s coding environments and developer feedback as training and evaluation inputs. A model must reason again after a tool responds and continue through a longer task. The environment, prompts and tools are part of training: good performance in one setup does not guarantee dependable behavior after the scaffold changes.[1]
Ideas
In an interview, Song recounts investigating reinforcement learning progress that had stalled despite a theoretically sound approach. Her team inspected layer-level probabilities and traced the problem to numerical precision. The investigation shows why numerical implementation matters alongside mathematical design.[1]
Tools, writing & talks
Writing, research & talks
Start here
Intelligence with Everyone: RL at MiniMaxcognitiverevolution.aiA combined presentation and interview covering developer-informed reinforcement learning, agent robustness and debugging training experiments.
Begin with the presentation portion for the agent-training problem, then the interview for a concrete debugging example. Basic familiarity with language models is enough to follow the discussion.
Speaker and interview guest[1]Inside MiniMax M3: an open-weight model with a 1M-token window. Covers its MSA block-retrieval sparse attention and GPU-aware kernels, why DSA did not fit grouped-query attention, and why text and vision trained together from step zero gave more stable results.
Catalog workOriginal source