Institution profile

National Taiwan University

Academic institutionasia · tw
Official website
Research library682linked papers
Opportunities0open roles
Selected work

Representative Papers

Retrieval-Augmented Generation for Natural Language Processing: A Survey

Jul 18, 2024arXiv.org

To address hallucination, knowledge staleness, and poor domain adaptability in large language models (LLMs), this paper conducts a systematic study of retrieval-augmented generation (RAG). We propose a full-stack RAG framework encompassing retriever design (dense, sparse, and hybrid), query rewriting, context fusion, LLM fine-tuning, and prompt engineering. We introduce the first taxonomy for dynamic knowledge updating and establish a multidimensional evaluation benchmark that balances academic rigor with industrial practicality. Additionally, we release a structured RAG knowledge graph and fully reproducible open-source code. Our contributions significantly enhance RAG’s robustness and maintainability in real-world deployments, providing both theoretical foundations and engineering best practices for knowledge-enhanced generative systems.

15 citationsRead paper

Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning

Jan 14, 2026

This work addresses the high inference latency and efficiency–performance trade-off inherent in existing Vision-Language-Action (VLA) methods that rely on explicit chains of thought. To overcome this limitation, we propose Fast-ThinkAct, a novel framework that introduces, for the first time, a verbalizable implicit chain of thought. By leveraging knowledge distillation from a teacher model and preference-guided optimization, Fast-ThinkAct efficiently transfers language and visual planning capabilities into embodied control policies. The approach maintains strong long-horizon planning, few-shot adaptation, and failure recovery abilities while substantially reducing computational overhead. Evaluated across multiple embodied manipulation and reasoning benchmarks, Fast-ThinkAct achieves up to an 89.3% reduction in inference latency without compromising task success rates or generalization performance, consistently outperforming prior methods.

3 citations1 influentialRead paper

Semi-Autonomous Mathematics Discovery with Gemini: A Case Study on the Erd\H{o}s Problems

Jan 29, 2026

This work proposes a semi-autonomous discovery framework that integrates artificial intelligence with human expertise to investigate 700 mathematical conjectures labeled as “open” in Bloom’s Erdős Problem Database. Leveraging the Gemini large language model for natural language reasoning and automated literature comparison as an initial screening step, candidate solutions are subsequently evaluated by domain experts for correctness and novelty. The study reveals that many problems deemed “open” stem not from intrinsic difficulty but from challenges in literature retrieval—termed “information occlusion.” Among the 13 problems successfully resolved, five yielded novel AI-generated solutions, while eight were traced to previously published results. This research represents the first large-scale demonstration of human–AI collaborative verification in mathematical conjectures and highlights the risk of “unconscious plagiarism” inherent in AI-assisted scholarly discovery.

2 citationsRead paper

ASMR: Augmenting Life Scenario using Large Generative Models for Robotic Action Reflection

Jun 16, 2025

To address inaccurate multimodal user intent understanding by domestic service robots under few-shot conditions, this paper proposes a semantics-controllable dialogue–scene image co-generation framework for data augmentation. To overcome the bottlenecks of scarce real-world multimodal data and high annotation costs, our method integrates large language models (LLMs) for contextualized dialogue modeling and reasoning, and leverages Stable Diffusion to synthesize high-fidelity, semantically aligned environment images. This establishes an end-to-end synthetic data generation and fine-tuning pipeline. To the best of our knowledge, this is the first framework enabling joint, controllable generation of linguistic intent and visual context. Experimental results demonstrate substantial improvements in action selection accuracy on real-world benchmark datasets, achieving state-of-the-art (SOTA) performance. The results validate that synthetically generated multimodal data effectively enhances downstream models’ generalization capability across modalities.

2 citationsRead paper
Recent publications

Latest Papers