Portfolio

About Me: I'm an AI researcher and MSc student in Computer Science. My current work sits at the intersection of multimodal LLM and post-training, with a focus on evidence-grounded reasoning.

Hello, I'm Zihao

A researcher focused on the intersection of multimodal LLM and post-training to advance the frontier of reliable, human-centered AI.

Research Interests

LLM post-training, multimodal LLM, evidence-grounded reasoning, LLM agents

Beyond Research

I enjoy giving talks on topics like LLM internals, attention mechanisms, and foundation models (see the Blog section).

Cuttlefish (ICML 2026)

Designed an entropy-guided adapter that focuses on instruction-relevant parts of multimodal entities, solving the longstanding gap of LLMs treating all modality positions as equally important.

ICML 2026Multimodal LLMAll-AtomStructure-Aware Reasoning

DetailsPaperCodeModelDataset

EDT-Former (ICLR 2026)

Built a dynamic connector that scales representation slots based on the input complexity, bridging the gap of fixed-size connectors that lose detail for large multimodal inputs.

ICLR 2026Graph-LLMMultimodalFoundation Model

DetailsPaperCodeModelDataset

MuMo (NeurIPS 2025)

Built a multimodal encoder that first stabilizes noisy 3D modality into a reliable prior, solving the longstanding instability in multimodal embeddings that incorporate 3D information.

NeurIPS 2025MultimodalMolecularRepresentation Learning

DetailsPaperCodeModelDatasetPosterSlides

Piccolo2 (SenseTime 2024)

General text embedding model with multi-task hybrid loss. SOTA on CMTEB with average score 70.95, surpassing all prior BERT-scale models.

Tech Report 2024Text EmbeddingFoundation ModelChinese NLP

DetailsPaperCodeModel

Piccolo-GPT (SenseTime 2024)

Unified LLM for both text embedding and generation via mixed contrastive-generative loss and two-stage training. Top-10 on CMTEB (score 68) with a single model powering full RAG pipelines.

Tech Report 2024LLMRAGText Embedding

DetailsPaperCodeModel

EC-Prune (Workshop)

Model-agnostic, data-centric graph pruning that removes boundary-noisy outer-layer nodes via an entropy–correlation dual criterion.

ICML 2026Graph PruningSpatiotemporalTransfer Learning

DetailsCode