Zhehao Zhang

I am a second year Master student in Computer Science at Dartmouth College. Currently, I am a research intern at Stanford SALT Lab under the supervision of Diyi Yang. Previously, I worked as a Research Intern at Adobe Research and Microsoft Research Lab – Asia. I received my bachelor's degree in Artificial Intelligence Honor Class at Shanghai Jiao Tong University.

My research interests lie in Natural Language Processing (NLP).

I am actively seeking a PhD position in NLP in the USA for Fall 2025.

Please feel free to contact me by email!

Mail  /  Résumé  /  Google Scholar  /  Github 

News📢

• 2024 Sept 25th: The DARG paper is accepted to NeurIPS 2024. See you in Vancouver!
• 2024 Sept 20th: One first-author paper on vision language model for code generation is accepted to EMNLP 2024. See you in Miami!
• 2024 Sept 5th: Honored to give a talk on Recent Advances in Synthetic Data for Foundation Models at Stanford SALT lab. Slides are available here.
• 2024 June 25th: A new preprint is released! Please check DARG which is a dynamic evaluation framework aiming to augment current reasoning benchmarks from the level of reasoning graphs.
• 2024 Mar 13th: One first-author paper on LLMs for hierarchical table analysis is accepted to NAACL 2024. See you in Mexico City!
• 2024 Mar: Honored to give a talk on Augmented Language Models at TRIP Lab at Dartmouth, hosted by Prof. Yaoqing Yang. Recording and the slide are available.
• 2024 Feb: I will join Adobe Research as Research Intern in this summer. See you in San Jose and bay area!
• 2023 Dec 14th: One first-author papers from my undergraduate is accepted to ICASSP 2024.
• 2023 Oct 27th: The paper titled "Can Large Language Models Transform Computational Social Science?" is accepted to Computational Linguistics.
• 2023 Oct 7th: Two first-author papers from my undergraduate are accepted to EMNLP 2023. See you in Singapore!

Research Interests🔍

My current research interests lie in multiple fields in NLP including:

• Language Agent
• Synthetic Data and Dynamic Evaluation of LLMs
• Computational social science (NLP for social good)
• Multi-Modal Large Language Model

TL; DR: We propose a dynamic evaluation framework named DARG, aimed at augmenting current reasoning benchmarks from the level of reasoning graphs. We evaluate 15 SOTA LLMs and observe a consistent performance decrease for all LLMs as the complexity level increases, along with increasing biases on some datasets.

TL; DR: We introduce VipAct, a multi-agent framework that combines visual-language models (VLMs) with vision expert models to enhance fine-grained visual perception, showing significant improvements over current baselines by leveraging multi-agent collaboration and specialized tools for visual perception tasks.

TL; DR: We explore the capabilities of vision-language models (VLMs) in generating Python code templates to reproduce academic data visualizations, finding that while closed-source models like GPT-4-Vision show promise in reproducing complex charts, open-source alternatives are less effective, especially for sophisticated visuals.

TL; DR: We provide a comprehensive survey of personalized large language models (LLMs), offering a unifying taxonomy across usage, granularity, techniques, datasets, and evaluation methods, while identifying critical challenges and open research questions in the field.

TL; DR: We propose a tool-augmented LLM framework named $E^5$ that contains 5 stages to solve the challenging real-life hierarchical table analysis task which achieves an 85.08 exact match score and 93.11 GPT-4-Eval Score. We also introduce $F^3$ which significantly reduces the token length while maintaining useful information to analyze such huge tables.

TL; DR: We systematically evaluate LLMs' reasoning ability on tabular data and establish a comprehensive taxonomy on operation and reasoning types for table analysis. hen, we propose CRT-QA, a dataset of complex reasoning QA over tables. We propose ARC which effectively utilizes table analysis tools to solve table reasoning tasks without manually-annotated exemplars.

TL; DR: We analyze the generation process of HS biases from a causal view and identify two confounders that cause the biases. Propose Multi-Task Intervention and Data-Specific Intervention to mitigate them.

TL; DR: We provide a road map for using LLMs as CSS tools and contribute a set of prompting best practices and an extensive evaluation pipeline to measure the zero-shot performance of 13 language models on 24 representative CSS benchmarks.

Dartmouth College, Master of Science in Computer Science, 2023 - 2025 (expected)

Shanghai Jiao Tong University, B.Eng. in Artificial Intelligence (Honor Class), 2019 - 2023

Social and Language Technologies (SALT) lab at Stanford, Research Intern

Adobe Research, Research Intern

Data, Knowledge, and Intelligence group at Microsoft Research Lab – Asia, Research Intern