Olive Song

In her December 2025 talk and January 2026 interview, Olive Song described her reinforcement-learning and model-evaluation work at MiniMax, including making coding agents adapt to different environments and meet expert developers’ expectations. MiniMax’s July 1, 2026 post identified her as its research lead in RL and reported her conversation with Hugging Face co-founder Thomas Wolf at AI Engineer World’s Fair about sparse attention, native multimodal training, and open weights.[2][4]

Her earlier December 2025 AI Engineer talk and January 2026 interview with Ksenia Se explain the reinforcement-learning practices behind that work. Both were republished in The Cognitive Revolution’s February 2026 episode.[4]

Song described expert developers as participants in MiniMax’s model-training cycle, contributing to problem definition, bug fixing, repository refactoring, rewards, and evaluation. Researchers and developers examined reinforcement-learning results together: a rising score prompted inspection of the actual behavior, and developers could identify problems that required a fix or more training data. For Song, alignment meant defining expert expectations and training toward them, rather than allowing a model to pursue its final goal through dangerous actions.[4]

In her presentation on MiniMax M2, Song explained interleaved thinking as repeated reasoning and tool interaction. After receiving a tool response, the model considers whether it has enough information and decides what to do next. This gives it opportunities to respond to errors and unexpected results during long tasks. She later qualified the mechanism: it was not full continual learning, but an approach on the path toward it.[4]

Her team initially expected training with many tools to prepare the model for unseen tools. Song said that changing the agent scaffold nevertheless exposed failures to generalize. The team broadened its definition of generalization to cover the model’s entire operational space—tool information, system and user prompts, chat templates, environments, and tool responses—and designed data perturbation pipelines around those variations.[4]

Song also recalled an implementation problem during MiniMax M1 experiments: accuracy had stopped improving despite the team’s theoretical expectations. They examined log probabilities layer by layer and traced the gap to numerical precision.[4]

Song said she tries new open models on their release day, including midnight launches, using personal questions collected across reasoning, mathematics, proofs, report writing, and agentic tasks. She distinguished that informal exploration from professional evaluation: “five questions isn't enough.” Fair comparisons, she argued, need diverse test sets, enough questions within each domain, and repeated runs because model behavior is not perfectly stable.[4]

Song supported open weights as a way for researchers and the community to build better models together. She pointed to self-deployment, fine-tuning, and keeping data on a user’s own infrastructure as benefits, while acknowledging that deployment and commercial strategy were outside her expertise. She also recognized a safety limit: MiniMax could evaluate and align a model before release, but could not control every use once its weights were public.[4]

In the January 2026 interview, Song identified better collaboration with experts and developers as her own next-generation research goal. That emphasis connected her evaluation work to a concrete question: whether a model could work effectively with people whose expectations extended beyond benchmark scores.[4]

Before those public accounts of her MiniMax work, Song graduated from NYU’s MSDS program in 2024. Looking back on the transition from school and lab work to industry, she said she had expected a routine of reading papers, developing ideas, and running experiments. Scaling data, compute, and teams instead brought unfamiliar problems that required reasoning from fundamentals and substantial engineering. As she put it, “engineering is very, very, very important.”[3][4]