All Speakers
Olive Song (Jiayuan Song)

Olive Song (Jiayuan Song)

LLM Algorithm Engineer

MiniMax

Expert Biography

Olive Song (Jiayuan Song)

LLM Algorithm Engineer at MiniMax specializing in large language model development and optimization. Contributed to the MiniMax-01 foundation model and M2 efficient agentic model, with research background in self-supervised learning and applied representation learning from NYU.

About MiniMax

MiniMax is an AI company based in Shanghai, China, founded in December 2021 by computer vision veterans from SenseTime. With over 150 million users globally, MiniMax has raised $1.15 billion in funding and achieved a $2.5 billion valuation in 2024, becoming one of China’s “AI Tiger” companies.

Key Products:

  • MiniMax M2 - Open-source MoE model (230B total/10B active parameters) achieving 2x faster inference than Claude Sonnet at 8% of cost
  • MiniMax-01 - Foundation model with lightning attention enabling 1M token context windows
  • Top Agentic Performance - Leading scores on τ²-Bench (77.2), BrowseComp (44.0), FinSearchComp-global (65.5)

Background & Expertise

Olive Song holds an M.S. in Data Science from NYU Center for Data Science and dual B.A. degrees in Mathematics and Computer Science from NYU Courant. Her research focuses on self-supervised learning, video prediction, and multimodal dialogue systems, with contributions to the VSTAR multimodal dialogue dataset (ACL 2023).

At MiniMax, Olive has played a key role in developing post-training strategies that enable M2’s exceptional performance on coding and agent tasks, exemplifying the company’s approach of tight research-engineering collaboration.

Key Research & Projects

Papers:

MiniMax Resources:

Conference Appearance

Event: AI Engineering Code Summit 2025 Date: November 20, 2025 Time: 12:20 PM Session: Building Voice-Activated Agents

Presented on MiniMax M2’s capabilities and training innovations for agentic AI, covering:

  • Coding Excellence - Scaled expert development with GitHub/internet data and feedback-driven training
  • Interleaved Thinking - Training for dynamic environments with tool-environment coupling
  • Agent Generalization - Perturbation resilience and tool scaffold adaptation
  • Multi-Agent Scalability - Cost-effective distributed coordination
  • Research-Product Integration - Co-located teams enabling rapid iteration from research to production

The session emphasized MiniMax’s organizational advantage: researchers and engineers working side-by-side with 150M users providing real-world feedback for continuous model improvement.


Last Updated: November 24, 2025

Stay Updated

Get the Latest AI Engineering Insights

Join the Focus.AI newsletter for curated research, analysis, and perspectives on the evolving AI landscape.

No spam. Unsubscribe anytime.

CLASSIFIED_FILES

USER: AUTHORIZED

[ EMPTY DRAWER ]

No documents have been filed.