🥳🥳🥳🥳

📢Check out our recent work in mathematical reasoning! Our Multi-View Fine-Tuning (MinT🌿) method achieves the SOTA performance of 7B models without LLM teachers, by harnessing different datasets, solution styles, and even noisy data in training. Link: arxiv.org/pdf/2307.07951…

#LLMs excel in various tasks, yet autoregressive #LLMs have limitations. We invite you to explore alternatives like non-autoregressive editing models with us. Join our tutorial (Google Eric Malmi & UCR) at #KDD2023 Aug 8th 10am-1pm, room 202c. Info: kdd2023-text-editing.github.io

Stabilizing RLHF through Advantage Model and Selective Rehearsal paper page: huggingface.co/papers/2309.10… Large Language Models (LLMs) have revolutionized natural language processing, yet aligning these models with human values and preferences using RLHF remains a significant

Instructing LLMs to solve math word prob better: - Mask a condition, Predict the condition with initial answer, Verify the answer, Rectify to avoid wrong answer(s) (AAAI'24): wzy6642.github.io/prp.github.io/ - Identify and Ignore Irrelevant Conditions (NAACL'24): wzy6642.github.io/I3C.github.io/

The code of our CoL is available now. Feel free to use it in your own project! Thanks to Yuyang Bai for organizing the code package. Code: github.com/QingkaiZeng/Ch… Please contact me (qingkaizeng.github.io) if you have any questions about our work.

🔎How to unleash LLM's inherent ability to detect and rectify incorrect responses without external feedback? 👀Check out our latest pre-print and find out how to progressively identify and correct false responses. Link: arxiv.org/abs/2405.14092

🚨🚨 We are hiring! RT appreciated! Prof. Rui Song (song-ray.github.io) and I will recruit post-doc scientists through Amazon’s post-doc program (amazon.science/postdoctoral-s…).

Just arrived in Philadelphia for Conference on Language Modeling ! Feel free to reach out!

Just arrived in Boise for ACM CIKM 2024 ! Feel free to reach out!

📄 paper: arxiv.org/abs/2402.04401 🔎 code: github.com/TamSiuhin/OPPU 🥳Joint work w/ QingkaiZeng Yijun Tian Zheyuan Liu Bing Yin Meng Jiang

How do we define 'long' chain-of-thought (CoT) reasoning? Is it about hundreds, thousands, or even more reasoning tokens for a question?

👀 How to effectively leverage the expertise of diverse models? ✨ Optimize graphs of LLMs with swarm intelligence! 👉🏻 Introducing Heterogeneous Swarms, jointly optimizing the roles and weights of multi-LLM systems for collaborative gains! 📄 Paper: arxiv.org/abs/2502.04510

Exciting news at the Lucy Family Institute for Data & Society! The Foundation Models and Applications Lab has launched with co-directors Meng Jiang and Xiangliang Zhang from University of Notre Dame's Department of Computer Science and Engineering. Learn more: lucyinstitute.nd.edu/news-events/20…

Totally agree. Still, I dislike how we split papers into “main vs. findings,” “oral vs. poster,” or “spotlight vs. others.” What’s truly valuable is solid, reproducible, real-world work. Collaboration > competition, always.