About
Linquan Wu is a PhD Student in the ASI Lab. of Department of Computer Science, City University of Hong Kong.
Linquan Wu works on multimodal and retrieval-augmented reasoning, applying large models to software and data intelligence. In LaViT he diagnoses a "Perception Gap" in multimodal distillation — students mimic a teacher's text while attending to divergent visual regions — and aligns latent visual thoughts via autoregressive reconstruction of the teacher's visual semantics and attention trajectories, gaining up to +16.9% on complex reasoning and letting a compact 3B model outperform GPT-4o. With HumanEval-V he co-built a 253-task benchmark turning diagrams into working Python, finding top models reach only 36.8% pass@1 and that vision-to-code alignment, not coding skill, is the bottleneck. In Transient-RAG he breaks the static-indexing GraphRAG paradigm by constructing query-specific ephemeral graphs with Semantic-Sentence Personalized PageRank and hypothetical edges, achieving state-of-the-art on HotpotQA, 2WikiMultiHopQA and ifQA while eliminating index maintenance.
Research Interests
Publications (6)
-
CoRR2026
-
Int.2026ConferenceInt. Conf. Computer Supported Cooperative Work in Design (CSCWD)
-
Int.2026ConferenceInt. Conf. Computer Supported Cooperative Work in Design (CSCWD)