Zheng Zhao

PhD Candidate

I am a final-year PhD candidate in Natural Language Processing at the UKRI Centre for Doctoral Training, University of Edinburgh. I work under the supervision of Shay Cohen and Bonnie Webber and am affiliated with the Institute for Language, Cognition and Computation (ILCC) in the School of Informatics. I am also an active member of The Cohort and EdinburghNLP research groups.

My doctoral research focuses on uncovering and understanding the inner workings of large language models (LLMs). Specifically, I investigate how these complex systems learn, adapt, and represent information across different domains, languages, and tasks.

During my PhD, I have explored topics including:

  • Domain learning in language models
  • Multilingual representation analysis
  • Instruction tuning and multi-task learning
  • Contextual understanding and representation learning

These research topics allow me to systematically dissect the computational processes underlying modern language models, providing insights into their learning mechanisms, representational capabilities, and limitations. In addition to my core PhD research, I have worked on topics including discourse analysis, model alignment, personalization, reducing model hallucination, temporal grounding, and summarization.

Google Scholar Semantic Scholar GitHub Twitter LinkedIn ORCID
Education
  • University of Edinburgh
    University of Edinburgh
    PhD in Natural Language Processing
    Sep. 2020 - Present
  • University of Edinburgh
    University of Edinburgh
    Masters by Research in Natural Language Processing
    Sep. 2019 - Aug. 2020
  • University of Edinburgh
    University of Edinburgh
    BEng in Artificial Intelligence and Software Engineering
    Sep. 2015 - Jul. 2019
Experience
  • Amazon AGI
    Amazon AGI
    Applied Scientist Intern
    Sep. 2024 - present
  • Amazon Alexa AI
    Amazon Alexa AI
    Applied Scientist Intern
    Jun. 2023 - Nov. 2023
  • Goldman Sachs
    Goldman Sachs
    Summer Analyst
    Jun. 2018 - Aug. 2018
News
2024
I am attending NeurIPS 2024 in Vancouver to present our work on LLM alignment.
Dec.
I am attending EMNLP 2024 in Miami to present our work on uncovering where multi-task learning happens in instruction-tuned LLMs.
Nov.
I am attending NAACL 2024 in Mexico City to present our papers on temporal grounding and context understanding in LLMs.
Jun.
2023
I am presenting two papers on multilingual summarization and multilingual representation analysis at EMNLP 2023 in Singapore.
Dec.
I am starting my internship at Amazon Alexa AI in Cambridge, UK.
Jun.
Selected Publications  view all
Layer by Layer: Uncovering Where Multi-Task Learning Happens in Instruction-Tuned Large Language Models
Layer by Layer: Uncovering Where Multi-Task Learning Happens in Instruction-Tuned Large Language Models

Zheng Zhao, Yftah Ziser, Shay Cohen

EMNLP 2024

This paper investigates how instruction-tuned LLMs internally process different tasks, finding that their layers organize into three functional groups: early layers for general features, middle layers for task-specific transitions, and final layers for refinement.

[Paper] [Code/Data] [Poster]

Layer by Layer: Uncovering Where Multi-Task Learning Happens in Instruction-Tuned Large Language Models
Layer by Layer: Uncovering Where Multi-Task Learning Happens in Instruction-Tuned Large Language Models

Zheng Zhao, Yftah Ziser, Shay Cohen

EMNLP 2024

This paper investigates how instruction-tuned LLMs internally process different tasks, finding that their layers organize into three functional groups: early layers for general features, middle layers for task-specific transitions, and final layers for refinement.

[Paper] [Code/Data] [Poster]

Spectral Editing of Activations for Large Language Model Alignment
Spectral Editing of Activations for Large Language Model Alignment

Yifu Qiu, Zheng Zhao, Yftah Ziser, Anna Korhonen, Edoardo Ponti, Shay Cohen

NeurIPS 2024

We introduce Spectral Editing of Activations (SEA), a novel inference-time method to adjust large language models' internal representations, improving truthfulness and reducing bias. SEA projects input representations to align with positive examples while minimizing alignment with negatives, showing superior effectiveness, generalization, and efficiency compared to existing methods with minimal impact on other model capabilities.

[Paper] [Code/Data]

Spectral Editing of Activations for Large Language Model Alignment
Spectral Editing of Activations for Large Language Model Alignment

Yifu Qiu, Zheng Zhao, Yftah Ziser, Anna Korhonen, Edoardo Ponti, Shay Cohen

NeurIPS 2024

We introduce Spectral Editing of Activations (SEA), a novel inference-time method to adjust large language models' internal representations, improving truthfulness and reducing bias. SEA projects input representations to align with positive examples while minimizing alignment with negatives, showing superior effectiveness, generalization, and efficiency compared to existing methods with minimal impact on other model capabilities.

[Paper] [Code/Data]

Enhancing Contextual Understanding in Large Language Models through Contrastive Decoding
Enhancing Contextual Understanding in Large Language Models through Contrastive Decoding

Zheng Zhao, Emilio Monti, Jens Lehmann, Haytham Assem

NAACL 2024 Oral

This work introduces a novel approach integrating contrastive decoding with adversarial irrelevant passages as negative samples to enhance robust context grounding during generation and operates at inference time without requiring further training.

[Paper] [Code/Data]

Enhancing Contextual Understanding in Large Language Models through Contrastive Decoding
Enhancing Contextual Understanding in Large Language Models through Contrastive Decoding

Zheng Zhao, Emilio Monti, Jens Lehmann, Haytham Assem

NAACL 2024 Oral

This work introduces a novel approach integrating contrastive decoding with adversarial irrelevant passages as negative samples to enhance robust context grounding during generation and operates at inference time without requiring further training.

[Paper] [Code/Data]

A Joint Matrix Factorization Analysis of Multilingual Representations
A Joint Matrix Factorization Analysis of Multilingual Representations

Zheng Zhao, Yftah Ziser, Bonnie Webber, Shay Cohen

EMNLP (Findings) 2023

This work presents an analysis tool based on joint matrix factorization for comparing latent representations of multilingual and monolingual models, and finds the factorization outputs exhibit strong associations with performance observed across different cross-lingual tasks.

[Paper] [Code/Data] [Poster]

A Joint Matrix Factorization Analysis of Multilingual Representations
A Joint Matrix Factorization Analysis of Multilingual Representations

Zheng Zhao, Yftah Ziser, Bonnie Webber, Shay Cohen

EMNLP (Findings) 2023

This work presents an analysis tool based on joint matrix factorization for comparing latent representations of multilingual and monolingual models, and finds the factorization outputs exhibit strong associations with performance observed across different cross-lingual tasks.

[Paper] [Code/Data] [Poster]

Understanding Domain Learning in Language Models Through Subpopulation Analysis
Understanding Domain Learning in Language Models Through Subpopulation Analysis

Zheng Zhao, Yftah Ziser, Shay Cohen

BlackboxNLP 2022

We examine how different domains are represented in neural network architectures, focusing on the relationship between domains, model size, and training data. Using subpopulation analysis with SVCCA on Transformer-based language models, we compare models trained on multiple domains versus a single domain. Our findings show that increasing model capacity differently affects domain information storage in upper and lower layers, with larger models embedding domain-specific information similarly to separate smaller models.

[Paper] [Code/Data] [Poster]

Understanding Domain Learning in Language Models Through Subpopulation Analysis
Understanding Domain Learning in Language Models Through Subpopulation Analysis

Zheng Zhao, Yftah Ziser, Shay Cohen

BlackboxNLP 2022

We examine how different domains are represented in neural network architectures, focusing on the relationship between domains, model size, and training data. Using subpopulation analysis with SVCCA on Transformer-based language models, we compare models trained on multiple domains versus a single domain. Our findings show that increasing model capacity differently affects domain information storage in upper and lower layers, with larger models embedding domain-specific information similarly to separate smaller models.

[Paper] [Code/Data] [Poster]

Reducing Quantity Hallucinations in Abstractive Summarization
Reducing Quantity Hallucinations in Abstractive Summarization

Zheng Zhao, Shay Cohen, Bonnie Webber

EMNLP (Findings) 2020

Abstractive summaries often hallucinate unsupported content, but our system, Herman, mitigates this by verifying specific entities like dates and numbers, improving summary accuracy and earning higher ROUGE scores and human preference.

[Paper] [Code/Data] [Poster]

Reducing Quantity Hallucinations in Abstractive Summarization
Reducing Quantity Hallucinations in Abstractive Summarization

Zheng Zhao, Shay Cohen, Bonnie Webber

EMNLP (Findings) 2020

Abstractive summaries often hallucinate unsupported content, but our system, Herman, mitigates this by verifying specific entities like dates and numbers, improving summary accuracy and earning higher ROUGE scores and human preference.

[Paper] [Code/Data] [Poster]

All publications