Guangyu Yang

About Me

I am a third-year PhD student in Information Engineering at the University of Cambridge, working in the Machine Intelligence Lab under the supervision of Prof. Bill Byrne. My PhD is fully funded by the Cambridge Trust.

I hold an MPhil in Machine Learning and Machine Intelligence from Cambridge. Before that, I obtained my BS from UC San Diego with a double major in Computer Science and Cognitive Science and a minor in Mathematics.

Research Interests

I study the safety and alignment of language and multimodal models, with a focus on LLM/VLM post-training. My research interests also include Retrieval-Augmented Generation (RAG), Vision-Language-Action (VLA) models, Minimum Bayes Risk (MBR) decoding, and neural machine translation.

Grants & Awards

UKRI AI Research Resource (AIRR) Gateway Award

2026

Project Co-Lead (with Prof. Bill Byrne as Project Lead)

10,000 GPU-hours on Isambard-AI; funded by UK Research and Innovation (UKRI).

Lambda Research Grant Program

2026

Lead Investigator

Cambridge International Scholarship

2024

Cambridge Commonwealth, European and International Trust

Full funding for PhD studies at the University of Cambridge.

Publications (Google Scholar)

Equal contribution Corresponding author

2026

ATPO teaser comparing fixed and controllable moderation policies.

Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization

Guangyu Yang (equal contribution), Jingbiao Mei (equal contribution), Mingsheng Sun (equal contribution), Jinghong Chen, Yingtong Bu, Pengda Qin, Da Chen (corresponding author), Bill Byrne (corresponding author)

NeurIPS 2026

Trains video safety models to detect multiple harmful categories while allowing moderation policies to adjust the balance between false alarms and missed detections.

Paper Project Code Models

BERAG: Bayesian Ensemble Retrieval-Augmented Generation for Knowledge-based Visual Question Answering

Jinghong Chen, Jingbiao Mei, Guangyu Yang, Bill Byrne

arXiv preprint

Combines evidence from individually retrieved documents through Bayesian weighting, enabling visual question answering with clearer source attribution and shorter contexts.

Paper

ATM-Bench teaser showing personal reference grounding, memory updates, and location-aware memory questions.

According to Me: Long-Term Personalized Referential Memory QA

Jingbiao Mei (corresponding author), Jinghong Chen, Guangyu Yang, Xinyu Hou, Margaret Li, Bill Byrne (corresponding author)

PALM @ NeurIPS 2026 (Workshop)

Introduces a benchmark and structured memory approach for assistants answering personal questions across years of photos, videos, emails, and conflicting evidence.

Paper Code

CycleVLA robot demonstrations showing failure prediction, backtracking, retry, and success.

CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding

Chenyang Ma, Kai Lu, Guangyu Yang (corresponding author), Jiuming Liu, Shitong Xu, Bill Byrne, Ioannis Havoutis, Niki Trigoni, Andrew Markham

arXiv preprint

Enables robots to anticipate failures, backtrack to earlier subtasks, and select retry actions using minimum Bayes risk decoding.

Paper Project Code

Retrieval-Augmented Defense pipeline overview.

Retrieval-Augmented Defense: Adaptive and Controllable Jailbreak Prevention for Large Language Models

Guangyu Yang, Jinghong Chen, Jingbiao Mei, Weizhe Lin, Bill Byrne

ACL 2026 (Main)

Detects jailbreak attempts using retrieved attack examples, allowing defenses to incorporate new attacks without retraining and adjust the balance between safety and usefulness.

Paper Project Code

2025

RA-HMD system diagram showing the multimodal backbone, contrastive training, and retrieval-augmented KNN classification.

Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection

Jingbiao Mei, Jinghong Chen, Guangyu Yang, Weizhe Lin, Bill Byrne

EMNLP 2025 (Main, Oral)

Combines retrieval and contrastive learning to adapt multimodal models for hateful meme detection while retaining their broader image and language capabilities.

Paper Code

On Extending Direct Preference Optimization to Accommodate Ties

Jinghong Chen, Guangyu Yang, Weizhe Lin, Jingbiao Mei, Bill Byrne

NeurIPS 2025

Extends direct preference optimization to model ties between responses, allowing training to use both clear preferences and equally rated examples.

Paper Code

2024

Direct Preference Optimization for Neural Machine Translation with Minimum Bayes Risk Decoding

Guangyu Yang, Jinghong Chen, Weizhe Lin, Bill Byrne

NAACL 2024 (Main)

Uses automatically ranked translations to train multilingual models with preference optimization, improving translation quality without extra computation at translation time.

Paper Code

2023

Multilingual Models in Neural Machine Translation

Guangyu Yang

MPhil Thesis

Paper

Education

University of Cambridge

Oct 2024 - Present

PhD, Information Engineering

Machine Intelligence Lab; Supervisor: Prof. Bill Byrne

University of Cambridge

Oct 2022 - Aug 2023

MPhil, Machine Learning and Machine Intelligence

Specialized in speech and language processing; Supervisor: Prof. Bill Byrne

UC San Diego

Oct 2018 - Jun 2022

BS, Computer Science and Cognitive Science (double major)

Minor in Mathematics; member of the Caledonian Honors Society at Muir College; graduated cum laude.

Industry Experience

RedNote (Xiaohongshu)

Apr 2026 - Oct 2026

Multimodal LLM Research Intern

Develop ATPO for controllable multi-label video safety detection using GRPO-based reinforcement learning and adaptive reward design.

To0 Space

Sep 2023 - Sep 2024

Machine Learning Engineer

To0 Space is a start-up developing diffusion-based models for high-fidelity architectural rendering.

Fine-tuned Stable Diffusion (SD) and Vision Transformer (ViT) models for architectural rendering. Built a pipeline using VLMs and LLMs to automate image labelling and captioning for SD training.
Applied retrieval-augmented generation (RAG), ControlNet, and IP-Adapter at inference time for architectural rendering.

Teaching Experience

University of Cambridge

Teaching Assistant

Lent 2026

MLMI 8: Large Language Model Applications

MPhil Dissertation Co-supervisor

2025–2026

Jailbreaking LLMs for Copyrighted Content

UC San Diego

Instructional Assistant

Spring 2022

COGS 118A: Supervised Machine Learning Algorithms