Foundation models
Efficient adaptation and learning dynamics of large models.
POSTDOCTORAL RESEARCH FELLOW · SIMON FRASER UNIVERSITY
Chuiyang Meng
I study how intelligent systems learn, reason, and collaborate, from the behavior of individual models to collective intelligence.
I am a Postdoctoral Research Fellow at Simon Fraser University. I received my PhD in Electrical and Computer Engineering from the University of British Columbia in June 2026, advised by Prof. Vincent W.S. Wong.
My work connects foundation models and mechanistic interpretability with multi-agent systems, distributed learning, and communication networks.
Research interests
Mechanistic interpretability
Causal mechanisms and emerging specialization inside neural networks.
Multi-agent systems
Communication, collective reasoning, and the spread of errors.
Distributed & edge learning
Federated and split learning across heterogeneous devices.
Communications & networks
Resource allocation and coordination in wireless networks.
News
-
Social Circuits and CORTEX now on arXiv
Our preprints on multi-agent echo chambers and learning shared and specialized functions in dense language models are now available on arXiv.
-
Joined Simon Fraser University
I joined Simon Fraser University as a Postdoctoral Research Fellow.
-
ASFL accepted by IEEE TMC
Our adaptive model splitting and resource allocation framework for split federated learning was accepted by IEEE Transactions on Mobile Computing.
-
Successfully defended my PhD
I successfully defended my PhD dissertation at the University of British Columbia, advised by Prof. Vincent W.S. Wong.
Earlier updates (5)
-
FLoRG accepted at ICLR 2026
Our paper on federated fine-tuning with low-rank Gram matrices and Procrustes alignment was accepted at ICLR 2026.
-
Internship at the Vector Institute
I joined the Responsible AI Team as an Applied Machine Learning Intern, working on deepfake video detection and robust machine learning.
-
ZorBA accepted at IEEE INFOCOM 2026
Our paper on zeroth-order federated fine-tuning of LLMs with heterogeneous block activation was accepted at IEEE INFOCOM 2026.
-
GNN-based resource allocation in IEEE TMC
Our paper on GNN-based resource allocation for decentralized federated learning appeared in IEEE Transactions on Mobile Computing.
-
Paper at IEEE GLOBECOM 2023
Our work on GNN-based neighbor selection and resource allocation for decentralized federated learning appeared at IEEE GLOBECOM 2023.
Selected research
All publications →Social Circuits behind Multi-agent Echo Chambers
arXiv · 2026 Submitted to ICLR 2027
Receiver-state interventions trace how messages correct errors or reinforce them, guiding selective communication between agents.
PDF arXiv
BibTeX
@misc{meng2026socialcircuits,
title = {{Social Circuits behind Multi-agent Echo Chambers}},
author = {Chuiyang Meng and Wenlu Yu and Ming Tang and Cheng Li},
year = {2026},
note = {Submitted to ICLR 2027},
url = {https://arxiv.org/abs/2609.34444},
eprint = {2609.34444},
archivePrefix = {arXiv}
}
View figure
Figure 1 · Error correction and echo chambers
Social Circuits behind Multi-agent Echo Chambers
Agents can agree because evidence corrects a shared mistake, or because repeated claims reinforce it. This example contrasts these two outcomes. Social Circuits traces message effects through receiver activations, and Circuit-Guided Deliberation learns which message combinations help the receiver.
Read the paper (PDF) ↗ Full-size figure ↗CORTEX: Learning to Share and Specialize in Dense Language Models
arXiv · 2026 Submitted to ICLR 2027
Domain-conditioned learning dynamics shape shared and specialized functions within a single dense language model.
PDF arXiv
BibTeX
@misc{meng2026cortex,
title = {{CORTEX: Learning to Share and Specialize in Dense Language Models}},
author = {Chuiyang Meng and Ming Tang and Vincent W.S. Wong},
year = {2026},
note = {Submitted to ICLR 2027},
url = {https://arxiv.org/abs/2609.34449},
eprint = {2609.34449},
archivePrefix = {arXiv}
}
View figure
Figure 1(b) · Learning shared and specialized functions with CORTEX
CORTEX: Learning to Share and Specialize in Dense Language Models
CORTEX uses domain-conditioned gradients and cross-domain gradient similarity to identify which parameters support shared or specialized functions. These assignments evolve during training, allowing functional modularity to emerge within one dense model.
Read the paper (PDF) ↗ Full-size figure ↗FLoRG: Federated Fine-tuning with Low-rank Gram Matrices and Procrustes Alignment
ICLR · 2026 Acceptance rate: 27%
Gram-matrix aggregation and Procrustes alignment make federated fine-tuning more consistent and communication-efficient.
PDF arXiv
BibTeX
@inproceedings{meng2026florg,
title = {{FLoRG: Federated Fine-tuning with Low-rank Gram Matrices and Procrustes Alignment}},
author = {Chuiyang Meng and Ming Tang and Vincent W.S. Wong},
year = {2026},
booktitle = {International Conference on Learning Representations (ICLR)},
url = {https://arxiv.org/abs/2602.17095}
}
View figure
Figure 1 · The FLoRG framework
FLoRG: Federated Fine-tuning with Low-rank Gram Matrices and Procrustes Alignment
Clients fine-tune a single low-rank matrix, and the server aggregates its Gram matrix to avoid the aggregation error introduced by separately averaging LoRA factors. Procrustes alignment stabilizes the recovered factor between rounds, reducing decomposition drift.
Read the paper (PDF) ↗ Full-size figure ↗ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning
IEEE TMC · 2026 Vol. 25, no. 10
Adapting where a model is split balances learning progress, training delay, and energy consumption.
PDF arXiv
BibTeX
@article{meng2026asfl,
title = {{ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning}},
author = {Chuiyang Meng and Ming Tang and Vincent W.S. Wong},
year = {2026},
journal = {IEEE Transactions on Mobile Computing (TMC)},
volume = {25},
number = {10},
url = {https://arxiv.org/abs/2603.04437}
}
View figure
Figure 1 · Adaptive split federated learning
ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning
ASFL divides model training between resource-constrained clients and a central server. An online optimization algorithm jointly adapts the model split and resource allocation to improve convergence under long-term delay and energy constraints.
Read the paper (PDF) ↗ Full-size figure ↗Tackling Resource Allocation for Decentralized Federated Learning: A GNN-based Approach
IEEE TMC · 2025 Vol. 24, no. 10
A graph neural network coordinates neighbor selection and resource allocation for decentralized learning.
PDF Publisher
BibTeX
@article{meng2025gnntmc,
title = {{Tackling Resource Allocation for Decentralized Federated Learning: A GNN-based Approach}},
author = {Chuiyang Meng and Ming Tang and Mehdi Setayesh and Vincent W.S. Wong},
year = {2025},
journal = {IEEE Transactions on Mobile Computing (TMC)},
volume = {24},
number = {10},
doi = {10.1109/TMC.2025.3562834}
}
View figure
Figure 3 · The proposed GNN model for decentralized federated learning
Tackling Resource Allocation for Decentralized Federated Learning: A GNN-based Approach
The model combines a client encoder for node and edge features with a local model encoder for compressed model parameters. A neighbor-selection module chooses aggregation partners, while a decoder outputs resource-allocation decisions for each client.
Read the paper (PDF) ↗ Full-size figure ↗Experience & service
My experience includes responsible AI research at the Vector Institute and teaching assistantships at UBC and Johns Hopkins.
Get in touch
For research discussions and collaboration, please reach out by email.
chuiyang_meng@sfu.ca