Peifeng Wang

I am a Research Scientist at Salesforce AI, where I mostly work on development of LLM-as-Judge and LLM post-training for enhancing reasoning. I did my PhD in Computer Science at the University of Southern California (USC), where I was advised by Prof. Muhao Chen and Prof. Xiang Ren.

Email / GitHub / Google Scholar / LinkedIn / CV

Research

I'm broadly interested in natural language processing, with a particular focus on reasoning and developing explainable artificial intelligence systems. My current research interests include:

Complex reasoning with large language models
Developing faithful explanations for AI models
Plan-ahead natural language generation
Integration of structured knowledge into language models

	Direct Judgement Preference Optimization Peifeng Wang, Austin Xu, Yilun Zhou, Caiming Xiong, Shafiq Joty preprint, 2024 arxiv / blog / A multifaceted generative judge model that generalizes cross various evaluation tasks
	DOMINO: A Dual-System for Multi-step Visual Language Reasoning Peifeng Wang, Olga Golovneva, Armen Aghajanyan, Xiang Ren, Muhao Chen, Asli Celikyilmaz, Maryam Fazel-Zarandi preprint, 2023 arxiv / code / A dual-system for answering a complex question over a chart step-by-step with LLM and a vision module.
	SCOTT: Self-Consistent Chain-of-Thought Distillation Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, Xiang Ren Annual Meeting of the Association for Computational Linguistics, 2023 Outstanding Paper Award arxiv / code / blog / A faithful knowledge distillation method that learns a small, self-consistent Chain-of-Thought model from a large teacher model.
	PINTO: Faithful Language Reasoning Using Prompt-Generated Rationales Peifeng Wang, Aaron Chan, Filip Ilievski, Muhao Chen, Xiang Ren International Conference on Learning Representations, 2023 arxiv / code / An LM pipeline that rationalizes via prompt-based learning, and learns to faithfully reason over rationales.
	Contextualized Scene Imagination for Generative Commonsense Reasoning Peifeng Wang, Jonathan Zamora, Junfeng Liu, Filip Ilievski, Muhao Chen, Xiang Ren International Conference on Learning Representations, 2022 arxiv / code / An Imagine-and-Verbalize framework, which learns to imagine a scene knowledge graph (SKG), and leverage the SKG as a constraint when generating a plausible scene description.
	Do Language Models Perform Generalizable Commonsense Inference? Peifeng Wang, Filip Ilievski, Muhao Chen, Xiang Ren Annual Meeting of the Association for Computational Linguistics - Findings, 2021 arxiv / code / An analysis over the ability of LMs to perform generalizable commonsense inference, in terms of knowledge capacity, transferability, and induction
	Learning to Deceive Knowledge Graph Augmented Models via Targeted Perturbation Mrigank Raman, Siddhant Agarwal, Peifeng Wang, Aaron Chan, Hansen Wang, Sungchul Kim, Ryan Rossi, Handong Zhao, Nedim Lipka, Xiang Ren International Conference on Learning Representations, 2021 arxiv / code / An analysis over the effects of strategically perturbed knowledge graphs (KGs) on KG-augmented model performance on downstream tasks.
	Scalable Multi-Hop Relational Reasoning for Knowledge-Aware Question Answering Yanlin Feng, Xinyue Chen, Bill Yuchen Lin, Peifeng Wang, Jun Yan, Xiang Ren Conference on Empirical Methods in Natural Language Processing, 2020 arxiv / code / A novel knowledge-aware approach that equips pre-trained language models with a multi-hop graph relation network.
	Connecting the Dots: A Knowledgeable Path Generator for Commonsense Question Answering Peifeng Wang, Nanyun Peng, Filip Ilievski, Pedro Szekely, Xiang Ren Conference on Empirical Methods in Natural Language Processing - Findings, 2020 arxiv / code / A general commonsense QA framework augmented with a knowledgeable path generator, which learns to connect a pair of concepts in text with a dynamic, and potentially novel, multi-hop relational path.
	Logic Attention Based Neighborhood Aggregation for Inductive Knowledge Graph Embedding Peifeng Wang, Jialong Han, Chenliang Li, Rong Pan AAAI Conference on Artificial Intelligence, 2019 arxiv / code / An inductive knowledge embedding model, Logic Attention Network, which aggregates entity neighbors with both rules- and network-based attention weights.
	Incorporating GAN for Negative Sampling in Knowledge Representation Learning Peifeng Wang, Shuangyin Li, Rong pan AAAI Conference on Artificial Intelligence, 2018 arxiv / A knowledge representation learning framework based on Generative Adversarial Networks (GANs) that aims at obtaining high-quality negative samples for more effective knowledge embedding.