Pleased to share that our paper “RedCoast: A Lightweight Tool to Automate Distributed Training of LLMs on Any GPU/TPUs” won a best demo paper runner up at #NAACL2024. Congrats Bowen Tan , Zhiting Hu , and the entire team!

💡Divergence thinking💡 is a hallmark of human creativity and problem-solving 🤖Can LLMs also do divergent reasoning to generate diverse solutions🤔? Introducing Flow-of-Reasoning (FoR) 🌊, a data-efficient way of training LLM policy to generate diverse, high-quality reasoning

On BlocksWorld, FoR produces both more diverse and higher-quality reasoning trajectories than CoT, Tree-of-Thoughts, RAP (MCTS), Supervised Finetuning (SFT), and PPO.

Fascinating idea of formulating LLM divergent thinking as sampling reasoning paths _proportional_ to reward functions (instead of just _maximizing_ reward)

With the context lengths of LLMs going beyond millions of tokens...📈 How to fully exploit their abilities with a super detailed prompt?🤔 LLM Reasoner now supports PromptAgent🤖: Let it find the best prompt for your task with advanced search methods🤩 shorturl.at/zU0Xk

"With long context LLMs comes long prompts"👇 People typically just write 1- or 2-sentence quick prompts when using an LLM for a task. How to create 1- or 2-page long prompts to boost performance? 🔥PromptAgent automatically writes long prompts for you!🔥 Without need of the

Optimizing pages-long expert-level prompts automatically 👇 It's fascinating that _prompt optimization_ can be formulated as a _planning_ problem: - Treat the LLM as a world model🌎 - We want a prompt, as a plan trajectory, that thrives in this world - So we do strategic

A higher quality version of the "Flow of Reasoning" animation 🎬🎞️🎥📽️

Excited to share that our paper “LLM Reasoners: New Evaluation, Library, and Analysis of Step-by-Step Reasoning with Large Language Models” is accepted to Conference on Language Modeling 2024! arxiv.org/abs/2404.05221 #LLMReasoners #COLM #COLM24

Happy to share that our agent engineering framework OpenAgents (github.com/xlang-ai/OpenA…) and LM tool-using survey(arxiv.org/2403.15452) are accepted by #COLM! See you at UPenn. And on a personal update, I feel excited and lucky that I will start my PhD at UCSD this fall

Finally a news that’s been brewing for years. Halıcıoğlu Data Science Institute and San Diego Supercomputer Center are coming together in forming the School of Computing, Information, and Data Science (SCI-DS), the 12th school in UC San Diego All credits go to Rajesh K. Gupta for making this happen.

📢Our amazing team are presenting two papers at #ICML2024. Join them to explore LLMs for chemistry reasoning and Controllable Jailbreaking LLMs 🌟 I'll miss being there, but hope everyone enjoys the conference! 🔥

Llama 3.1 is a huge advancement in open weight models - congrats AI at Meta team! The tech report introduces a branching concept: create an expert model using an unfinished checkpoint and new data. LLM360 has 500+ ckpts to branch from across 3 models. Get them on Hugging Face!**

Really excited about the #ACL2024 outstanding paper award to our work on Multi-Modal Theory-of-Mind evaluation! Congrats to Chuanyang Jin @ ACL 2024 and Tianmin Shu who led the work! Check out more: - chuanyangjin.com/mmtom-qa - maitrix.org

Rajesh K. Gupta, founding director of HDSI, is appointed Interim Dean of the School of Computing, Information, & Data Sciences at UC San Diego! 🎉 Join us in congratulating him as he leads SCIDS into a future of innovation and excellence. #HDSI #UCSD #Leadership #DataScience

Excited to announce the official leaderboard for our MMToM-QA benchmark hosted on Maitrix.org

Enjoyed this paper on LMs, world models and agent models by Zhiting Hu and Tianmin Shu TLDR: for reasoning tasks, it’s a useful abstraction to treat LMs as simulators (“backends”) that simulate agent models and world models arxiv.org/abs/2312.05230

Very interesting work of simulating digital game with a DOOM "world model"! 🌴Pandora we created earlier is a general-domain world model aiming to simulate diverse worlds, including digital games, interactively controlled by natural language. A larger and better version is