Binance Accelerator Program - AI Research Scientist (LLM Reasoning & Post-Training)

About the Role

You'll work alongside senior research scientists on problems at the frontier of LLM reasoning, post-training methodology, and agentic AI — in one of the few environments where your models interact with live global markets at scale.

This isn't a support or literature-review role. You'll run experiments, form independent hypotheses, implement ideas from recent papers, and work closely with engineering teams to understand how research behaves under real production constraints — 24/7, zero-downtime, hundreds of millions of users.

Who may apply

Current university students (Masters, PHD in AI track) or recent graduates who don't mind starting as intern. 

Responsibilities

Design and run experiments in reasoning model training, post-training alignment, test-time compute scaling, and systematic model evaluation — grounded in financial and crypto-native problem settings
Implement model variants, training pipelines (including RLVR-based approaches), and evaluation frameworks in PyTorch and the Hugging Face ecosystem
Synthesize recent work from NeurIPS, ICML, ICLR, and ACL to sharpen active research directions — not just track the field, but translate it into testable ideas
Apply LLM reasoning to crypto-native data: on-chain signals, market microstructure, and multi-modal market intelligence — research opportunities that don't exist anywhere else
Maintain rigorous experiment tracking and reproducibility standards (W&B or equivalent)
Partner with applied engineering to understand how research translates into production systems — and what constraints actually matter

Requirements

Currently pursuing a Master's or PhD in Machine Learning, Computer Science, Mathematics, or a related field (preferably graduating between 2026 to 2028)
Strong Python and PyTorch fundamentals; C++ or Rust exposure is a bonus 
Comfortable using AI-assisted development tools as a natural part of your research workflow — not as a crutch, but as leverage
Solid grounding in transformer architectures, LLM pretraining, and the shift toward reasoning-capable models
You form opinions about research, not just summaries of it