About Me 🌲
Hi, I am a third-year Ph.D. student at the University of New South Wales (UNSW Sydney), supervised by Dr. Dong Gong. Before that, I received my MPhil degree from the Australian Institute for Machine Learning (AIML), University of Adelaide, supervised by A/Prof. Qi Wu and Dr. Yuankai Qi. I obtained my Bachelor’s degree from Beijing Jiaotong University, where I was advised by Prof. Runmin Cong.
My research focuses on continual learning of foundation models, particularly Dynamic MoE for learning when, where, and how much capacity to add, and which experiences should drive learning. More broadly, I ask: when a trained model encounters new experience, what should change, where should it be stored, and what should drive the update?
Building on this question, I am currently exploring how models can learn from experience through updates to their parameters, recurrent states, and representations of the world:
- Continual and self-improving learning — how a pretrained model can continue acquiring capabilities through post-training, test-time learning, and interaction, while deciding what to update and which experience should drive that update.
- Recurrent sequence modeling — how a fixed-size hidden state can compress and preserve long histories, what information should be written, retained, or overwritten, and whether the state-update rule itself can be learned rather than manually designed.
- World models and interactive simulation — how generative models can learn compact, persistent representations of physical or digital environments from interaction, enabling agents to predict, act, and continue learning over long horizons.
News 🔥
Aug 2026CoDyRA is accepted to EMNLP 2026 Findings. Congratulations to Jeff!
Forgetting in continual post-training grows with LoRA rank — a sparsity penalty on per-component importance keeps each update at the rank its task needs.Aug 2026Released QwenReasoning, an open-source evaluation harness for test-time reasoning.
Soft, entropy-gated and CoT decoding for Qwen3 / Qwen3.5 / Qwen3.8, on HF and vLLM.May 2026Recognized as an ICML 2026 Gold Reviewer!May 2026MoRAM is accepted to ICML 2026. Congratulations to Jeff!
No router — each expert is a rank-1 key–value memory that activates on its own key, so capacity grows as memory atoms instead of coarse experts a router must tell apart.Feb 2026DyMoE is accepted to CVPR 2026. Thanks to all collaborators!
Frozen experts still forget: old and ambiguous tokens in new-task data teach little but drift onto the new experts — DyMoE filters them out token by token.Jun 2025SAME is accepted to ICCV 2025. Congratulations to Gengze!
One agent for seven navigation tasks — experts routed by the agent's state, not by the instruction's granularity.Dec 2024New preprint! Check CoDyRA on arXiv!
Forgetting grows with update rank, so each LoRA learns only the rank it needs — rank minimization as an implicit forgetting regularizer.Dec 2024New preprint! Check MambaCL on arXiv!
Meta-trained so continual learning happens at test time in the forward pass — Mamba's constant-size state is what gets updated, no gradient steps and no growing KV cache.Oct 2024Recognized as an ACM MM 2024 Outstanding Reviewer!Dec 2023WebVLN is accepted to AAAI 2024. Thanks to all collaborators!
Vision-and-language navigation on shopping websites.Jul 2023MG-VLN is accepted to ACM MM 2023. Thanks to all collaborators!
Backtracking to passed correct locations via first-person video grounding for navigation.
Research 🌴
Continual Learning
CVPR On Token’s Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models
CVPR 2026
CVPR 2026
TL;DR 
In an LVLM whose MoE grows task by task, freezing the old experts and their routers does not prevent forgetting, because of routing drift, in which training on new-task data reassigns old-task tokens to the new experts.
Ambiguous tokens, scored similarly by the old and the new experts, cause most of the drift and contribute least to learning the new task. DyMoE therefore filters tokens before assignment, restricting reassignment to the ambiguous ones and to the router. This reduces forgetting while preserving new-task performance.
Ambiguous tokens, scored similarly by the old and the new experts, cause most of the drift and contribute least to learning the new task. DyMoE therefore filters tokens before assignment, restricting reassignment to the ambiguous ones and to the router. This reduces forgetting while preserving new-task performance.
Preprint Learning Mamba as a Continual Learner: Meta-learning Selective State Space Models for Continual Learning
Preprint
Preprint
TL;DR
A recurrent model replaces the unbounded KV cache of attention with a fixed-size state, so at each step its update rule has to decide what to keep and what to overwrite.
We study that update rule as a learning rule: what the state should write, and how that writing should be trained. The rule is meta-learned rather than hand-designed, and training is regularized on what the state keeps. Across Mamba, Longhorn and Gated DeltaNet, as well as RWKV-4 and -7, rules closer to the delta rule learn better from a non-stationary stream.
We study that update rule as a learning rule: what the state should write, and how that writing should be trained. The rule is meta-learned rather than hand-designed, and training is regularized on what the state keeps. Across Mamba, Longhorn and Gated DeltaNet, as well as RWKV-4 and -7, rules closer to the delta rule learn better from a non-stationary stream.
ICML Little By Little: Continual Learning via Incremental Mixture of Rank-1 Associative Memory Experts
ICML 2026
ICML 2026
TL;DR 
Taking a whole LoRA module as one expert is a coarse unit of growth, since each new expert overlaps with the existing ones and routing becomes less reliable as they accumulate. Read as a linear associative memory, an expert decomposes into rank-1 key–value pairs, each responding only to its own key. Capacity then grows one memory at a time, and no router is required.
EMNLP Take Only What You Need: Rank Minimization as an Implicit Forgetting Regularizer in Continual Learning
EMNLP 2026 Findings
EMNLP 2026 Findings
TL;DR 
The best rank shifts with the module and the task, and a bound explains why: forgetting scales with both the rank of an update, how many directions it uses, and its norm, how far it moves along each. One sparsity penalty on per-component importance presses both at once, leaving every update at the rank its task needs and mergeable straight back into the backbone — no task ID, no replay.
Reasoning
Project QwenReasoning: Test-Time Reasoning Methods for Qwen3, Qwen3.5 and Qwen3.8
Open-source project, 2026
Open-source project, 2026
TL;DR 
Training-free decoding that skips the argmax: the probability-weighted embedding goes back into the model, so the next step reads the full distribution rather than one sampled token. Soft, entropy-gated and CoT variants share an entry point, on HF and a vLLM fork, over five math and science benchmarks.
Embodied Navigation
ACM MM Mind the Gap: Improving Success Rate of Vision-and-Language Navigation by Revisiting Oracle Success Routes
ACM MM 2023
ACM MM 2023
TL;DR
Across four VLN agents on R2R and REVERIE, oracle success runs up to 9% above success — the agent reaches the target and walks on. The stop is therefore recoverable from a trajectory an off-the-shelf agent already produced: read the route as a first-person video and ground the instruction in it to locate the frame to stop at.
AAAI WebVLN: Vision-and-Language Navigation on Websites
AAAI 2024
AAAI 2024
TL;DR 
One of the early benchmarks for language-guided web agents, moving navigation from 3D houses to tree-structured web pages and replacing turn-by-turn instructions with questions to answer. The agent reads the raw HTML as well as the rendering, which exposes text and structure the page never displays. I contributed to the task formulation and was responsible for constructing the WebVLN-v1 benchmark and adapting embodied navigation methods to the web.
ICCV SAME: Learning Generic Language-Guided Visual Navigation with State-Adaptive Mixture of Experts
ICCV 2025
ICCV 2025
TL;DR 
Category-level search and step-by-step instruction following are built as separate tasks, though both read an instruction, look around, and pick a move. SAME routes its experts on the agent's navigation state rather than on the instruction's granularity, covering seven language-guided navigation tasks with one agent.
Services
Conference Reviewer
ICML '24 '25 '26 Gold Reviewer NeurIPS '24 '25 '26 ICLR '25 '26 '27
CVPR '24 '25 '26 ICCV '25 ECCV '24 '26
ACM MM '24 Outstanding Reviewer AAAI '25 '27 ICDM '24
Journal Reviewer
IJCV IEEE-TNNLS IEEE-TCSVT Neurocomputing CAAI-TIT