Email: chuanx [at] nagoya-u.jp

Phone: +81-6-6105-6502 

Address:

Graduate School of Information Science and Technology

Osaka University

1-5, Yamadaoka, Suita, Osaka 565-0871

Japan

[DBLP Entry] [Google Scholar]

News

• PAG: A projection-augmented graph approach to approximate nearest neighbor search (ANNS), super fast in both indexing and searching (5x HNSW), scaling up to high dimensionality (e.g., 3072) and retrieval size K (10 to 1000). Index size can be adjusted to meet memory requirement. [code]

• ShapleyLaw: multilingual scaling laws for language models. 

• PKF: (ICLR 2026 Oral) angle testing for fast MIPS and ANNS. [code]

• ELLMob: (ICLR 2026) An event-driven approach to human mobility generation using LLMs. [code]

• FleS: (ICLR 2026) self-gated activation model for discriminative pattern recognition. 

• ASL: An adaptive layer-wise token pruning approach to KV cache reduction. [code]

• LFP: (EMNLP 2025) We introduced the task of legal fact prediction and constructed a dataset for 657 litigation cases in China, covering 10 representative types of civil cases. [benchmark] 

• SAS: We are excited to announce the launch of SAS (Smart Agent Survey), a software designed for automated surveys utilizing LLM agents. Users can upload their questionnaires in various formats (DOCX, PDF, etc.). The software employs LLM agents to simulate respondent answers, making it a valuable tool for marketing strategies and policy-making. 

• We argue that LLM-based social simulations need clear boundaries to meaningfully contribute to social science research.

• We analyze 7 security challenges that must be solved in cross-domain multi-agent LLM systems.

• VLM & LLM fusion: We are excited to share some fascinating discoveries on combining Vision-Language Models (VLMs) with large language models (LLMs).