Thrilled to receive the KDD Dissertation Award Runner-Up, for my PhD works on Neural-Symbolic Reasoning.
Sincerely thanks to my PhD advisors
@YizhouSun
and
@kaiwei_chang
, my letter supporters
@yisongyue
and
@jhamrick
. Thanks to the award committee
@kdd_news
for such honor.
Interested in LLM + Tool-Use, via Tree-Search?
This afternoon in
#NeurIPS2023
,
#215
, I'll present "AVIS: Autonomous Visual Information Seeking with Large Language Model Agent" ()
Feel free to drop by and chat.
🤔 How to let Large Language Models (LLMs) agent utilize diverse tools via Tree Search 🔍?
In AVIS, we enable LLM Agent to dynamically traverse a transition graph with self-critic (when one path is not informative, backtrack to previous state). This achieves SOTA VQA result.
Today on the blog, read all about AVIS — Autonomous Visual Information Seeking with Large Language Models — a novel method that iteratively employs a planner and reasoner to achieve state-of-the-art results on visual information seeking tasks →
Can LLMs play a hidden-identity board game "Renaissance Avalon"?
Check out:
Code:
In this work, we built a game engine AvalonBench, consisting of several fixed rule baselines. We found ChatGPT 3.5 still cannot beat simple rules.
How to control LLM behavior with LLM-as-a-judge?
Check our paper: "Self-Control of LLM Behaviors by Compressing Suffix Gradient into Prefix Controller"
Website:
Paper:
Code:
Excited to receive the
#SoCalNLP
Best Paper Award for our paper "Empowering Language Models with Knowledge Graph Reasoning for Question Answering". The paper link is:
Thanks to the organizers and all the great collaborators!
Our
@MegagonLabs
Best Paper Award winner was "Empowering Language Models with Knowledge Graph Reasoning for Question Answering" by Ziniu Hu et al from UCLA!
Paper link:
Thank you to award sponsor
@MegagonLabs
for supporting our event! (4/4)
Can LLM be the "world model🌍" to predict the future?
Our new benchmark evaluate LLM Agents for global international events (conflict ⚔️vs mediation🤝)
We provide agents of Python APIs to interact with diverse tools and knowledge sources, and SoTA GPT-4o + code achieves 32% F1.
📢New LLM Agents Benchmark!
Introducing 🌟MIRAI🌟: A groundbreaking benchmark crafted for evaluating LLM agents in temporal forecasting of international events with tool use and complex reasoning!
📜 Arxiv:
🔗 Project page:
🧵1/N
🎹 How to make music diffusion model a really useful tool for music composition?
Check our work on non-differential rule guided diffusion: .
It controls the music generation by rules (e.g. chord progression) in a training-free, plug-and-play manner.
Excited to share our work on symbolic music generation: !
We introduce a symbolic music generator with non-differentiable rule guided diffusion models, enabling musicians to effectively use it as a compositional tool.
Website: . 🧵👇
Hi
#NeurIPS2022
This afternoon 4pm at poster 211, we'll present our work "Improving Multi-Task Generalization via Regularizing Spurious Correlation". If you're interested in unique challenges of out-of-distribution generalization for Multi-Task Learning, please come and chat!
Interested in how Spurious Correlation affects Multi-Task Generalization (especially out-of-distribution setting)?
Check out our
#NeurIPS2022
spotlight paper:
I will present the poster at Hall J 211 on Thursday 2-4pm (Dec 1). Please drop by and chat!
#CVPR2023
This afternoon from 4pm, I will present our CVPR ✨highlight paper, REVEAL, at Exhibit Halls ABC 264.
If you're interested in augmenting large Visual-Language model with external and up-to-date knowledge, please drop by and chat~
Learn how REVEAL, an end-to-end retrieval-augmented visual-language model that learns to use multi-source multi-modal data to answer knowledge-intensive queries, achieves state-of-the-art results on visual question answering and image caption tasks.
Interested in how Spurious Correlation affects Multi-Task Generalization (especially out-of-distribution setting)?
Check out our
#NeurIPS2022
spotlight paper:
I will present the poster at Hall J 211 on Thursday 2-4pm (Dec 1). Please drop by and chat!
🧐 Can neural simulators softly satisfy multiple physical constraints?
Check out:
We propose TANGO, a physics-informed GraphODE that injects time-reversal symmetry and in the meanwhile numerically benefits various dynamical systems.
Interested in building 🚀fast and efficient
#GraphNeuralNetworks
for large-scale data?
Check out our recent survey () led by Shichang, which provides a clear taxonomy of GNN acceleration, and suggest future
directions.
The game requires both complicated decision-making and language skills (including cooperation, deception and deduction). We hope it can serve as a test-bed for future research of LLM Agent.
Thanks for the contributions by
@JonathanMLight
, Min Cai and
@shengs1123
Self-Control enable fine-grained control for a wide range of tasks, including emotional modulation, ensuring harmlessness, and enhancing complex reasoning (GSM8K, comparable with CoT-Decoding), and achieve 0 error on privacy leakage task.
@dezhou
Yeah it can. The tree search framework is quite general, one concurrent work tree-of-thought has many interesting results on pure text problem.
Major differences to ToT are: 1) a transition graph as prior to narrow down search space; 2) working memory to avoid taking same paths.
The searched gradient for each task can also provide insights on interpreting LLM: different tasks apply different layers of transformers, meaning such behavior is "stored" at different places of Transformer.
MT-CRL could improve multi-task generalization (especially with distribution shift), and we show that it could indeed help alleviate spurious correlation. Without MT-CRL, the children's movie recommendation could be associated with violent words. MT-CRL could address this issue.