Shangbin Feng
Allen School Ph.D. student Shangbin Feng aims to build a more open and democratic artificial intelligence future. To that end, his research focuses on model collaboration, where “multiple AI models, trained on different data, by different people, and thus possess diverse skills and strengths, collaborate, compose and complement each other.”
In December, Feng was named among the 2026 class of NVIDIA Graduate Fellows in recognition of his work. The NVIDIA Graduate Fellowship program supports graduate students from around the world whose outstanding research puts them at the forefront of accelerated computing and is relevant to the company’s interests.
“Through model collaboration, I aim to spearhead a modular, compositional, decentralized, and participatory AI future — that everyone everywhere could have a say in the future of AI by contributing data, models, or natural language feedback reflecting their interests and priorities, and building a compositional AI system from the bottom up with their decentralized contributions,” said Feng, who is advised by Allen School professor Yulia Tsvetkov.
Feng used model collaboration techniques to enhance the reliability and trustworthiness of large language models (LLMs). Despite evolving efforts to expand the LLMs’ knowledge base, they still run into knowledge gaps, or missing or outdated information. To help models abstain from generating low-confidence outputs, he and his team introduced two novel, robust multi-LLM collaboration-based approaches where LLMs probe other LLMs for knowledge gaps, either cooperatively or competitively. When multiple LLMs are working together in cooperation, one LLM employs other models to give feedback on the proposed answer, and it then synthesizes all the outputs into an overall abstain decision. In a competitive setting, the LLM is challenged by other models with conflicting information, and it has to decide whether to abstain or not. The team’s paper describing this approach received an Outstanding Paper Award at ACL 2024, the conference organized by the Association for Computational Linguistics.
