Publications

Preprint

Dynamic Rewarding with Prompt Optimization Enables Tuning-free Self-Alignment of Language Models

Self-MoE: Towards Compositional Large Language Models with Self-Specialized Experts

2024

LLM Reasoners: New Evaluation, Library, and Analysis of Step-by-Step Reasoning with Large Language Models

LLM Reasoners is a library to enable LLMs to conduct complex reasoning, with advanced reasoning algorithms. It approaches multi-step reasoning as planning and searches for the optimal reasoning chain, which achieves the best balance of exploration vs exploitation with the idea of "World Model" and "Reward."

PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization

Tired of manual prompt engineering? PromptAgent offers the first principled framework to formalize the problem of API-based prompt optimization (state, action, reward, etc); also the first to benchmark exploration efficiency and show the transferability of optimized prompts. Targeting for expert-level prompting, there are many exciting directions ahead of PromptAgent!

GPT Is Becoming a Turing Machine: Here Are Some Ways to Program It

Through appropriate prompting, GPT models can be triggered to perform iterative behaviors necessary to execute (rather than just write or recall) programs that involve loops, including several popular algorithms found in computer science curricula, e.g., logical deduction, bubble sort, longest common subsequence, etc.

2023

Reasoning with Language Model is Planning with World Model

LLMs lack internal world models for effective reasoning. Reasoning via Planning (RAP) reformulates LLM reasoning as a planning problem, thus incorporating an external world model and principled planning seamlessly. This is a new framework applicable across varying tasks and an exciting direction for LLM augmentation research.

ToolkenGPT: Augmenting Frozen Language Models with Massive Tools via Tool Embeddings

ToolkenGPT augments LLMs with massive tools/APIs by representing tools as tokens (“toolken”) and enabling tool calls in the same way as generating regular words. ToolkenGPT is super efficient for learning massive tools, as plugging in new tools is as easy as learning embeddings.

ThinkSum: Probabilistic Reasoning Over Sets Using Large Language Models

We propose a two-stage probabilistic inference paradigm, ThinkSum, to improve LLMs' abilities of reasoning over multiple objects in two steps, Think (e.g., retrieval of associations) and Sum (e.g., aggregation of results), which beats chain-of-thought prompting in hard BIG-bench tasks.

Multitask Prompt Tuning Enables Parameter-Efficient Transfer Learning

We propose Multitask Prompt Tuning (MPT) to exploit the rich cross-task knowledge for more efficient and generalizable transfer learning. MPT learns a single transferrable soft prompt through the use of a novel combination of prompt decomposition and prompt distillation.

Entity Tracking via Effective Use of Multi-Task Learning Models

How to transfer multi-task knowledge from pre-training to niche downstream tasks, such as entity tracking on the procedural text? We show that you can reach STOA performance by simply fine-tuning T5 but with specialized QA prompts and task-specific decoding.

2022

Toward Knowledge-Centric NLP: Acquisition, Representation, Transfer, and Reasoning

Coherence Boosting: When Your Pretrained Language Model is Not Paying Enough Attention

We demonstrate that large language models have insufficiently learned the effect of distant words on next-token prediction. We present Coherence Boosting, an inference procedure that increases an LM’s focus on a long context, which greatly improves NLG and NLU tasks.

Knowledge Transfer between Structured and Unstructured Sources for Complex Question Answering

We study knowledge transfer for multi-hop reasoning processes between structured (Knowledge Base) and unstructured (text corpus) knowledge. We design SimultQA unifying KBQA and TextQA systems and leverage it to study how reasoning is transferred between two knowledge sources.

2021

Bootstrapping a User-Centered Task-Oriented Dialogue System

We build TacoBot, a task-oriented dialogue system for the inaugural Alexa Prize TaskBot Challenge, to assist users in multi-step cooking and home improvement tasks. We propose several data augmentation methods, such as GPT-3 simulation, to bootstrap neural dialogue systems into new domains and make them more robust to noise user initiatives.

Modeling Context Pair Interaction for Pairwise Tasks on Graphs

We propose to explicitly model context interactions for pairwise prediction tasks on graphs, which consist of two perspectives, node-centric and pair-centric. We also propose to pre-train pair embeddings to facilitate the pair-centric model.

2020

Rationalizing Medical Relation Prediction from Corpus-level Statistics

We propose a self-interpretable framework to rationalize the neural relation prediction based on corpus-level statistics. This framework is inspired by human cognitive theory about recall and recognition, which provides structured knowledge triplets as rationales.

Graph Embedding on Biomedical Networks: Methods, Applications, and Evaluations

We benchmark 11 representative graph embedding methods on five important biomedical tasks. We verify the effectiveness of recent graph embedding methods and provide general guidelines for their usage.

2019

SurfCon: Synonym Discovery on Privacy-Aware Clinical Data

We propose to discover structured knowledge--synonyms--from the privacy-aware text corpus and present a novel framework to leverage both surface form and context information to discover out-of-distribution synonyms.

Before 2019

A Comprehensive Study of StaQC for Deep Code Summarization

We examine three popular datasets mined from Stack Overflow on the code summarization task and show that StaQC (Stack Overflow Question-Code pairs) helps achieve substantially better results.

Hessian Regularized Sparse Coding for Human Action Recognition

@inproceedings{liu2015hessian,
                  title={Hessian regularized sparse coding for human action recognition},
                  author={Liu, Weifeng and Wang, Zhen and Tao, Dapeng and Yu, Jun},
                  booktitle={International Conference on Multimedia Modeling},
                  pages={502--511},
                  year={2015},
                  organization={Springer}
                  }
                  

We propose Hessian regularized sparse coding (HessianSC) for action recognition, which can preserve the local geometry well and steer the sparse coding varying linearly along the manifold of data distribution.