One Size Does Not Fit All! Dynamic Retriever and Generator Selection for RAG
研究通过动态选择检索器和生成器来优化RAG系统的资源分配问题，提出DRAG框架以适应不同查询需求，提高效率。
0 citationsRead paper
Selected work
Representative Papers
Automatic Generation of Titles for Research Papers Using Language Models
This study addresses the time-consuming and challenging task of academic paper title generation by proposing an automated approach that produces titles directly from abstracts. The authors introduce SpringerSSAT, a novel and publicly released dataset in the social sciences, and conduct a systematic evaluation of several large language models—including PEGASUS-large, LLaMA-3-8B, and GPT-3.5-turbo—for this task. Experimental results demonstrate that fine-tuned PEGASUS-large consistently outperforms other models across multiple metrics such as ROUGE, METEOR, and BERTScore, generating titles that are accurate, concise, and practically useful. The study also highlights the potential of ChatGPT in producing creative and engaging titles, suggesting complementary strengths among different model architectures for scholarly title generation.
0 citationsRead paper
An Evaluation of Context Length Extrapolation in Long Code via Positional Embeddings and Efficient Attention
This work addresses the challenge that large language models, constrained by fixed context lengths, struggle to effectively process code sequences exceeding their training length. The authors propose a zero-shot inference method that requires no fine-tuning, leveraging optimized positional encoding and efficient attention mechanisms to systematically evaluate the context-length extrapolation capabilities of existing techniques on long code completion tasks. For the first time, they provide a comprehensive comparison of multiple positional embedding schemes and attention strategies in ultra-long code scenarios, revealing significant differences in their extrapolation performance. This study offers empirical evidence and practical guidance for enhancing the ability of large models to handle extremely long code sequences.
0 citationsRead paper
TaxoAlign: Scholarly Taxonomy Generation Using Language Models
Existing automated academic survey generation methods lack systematic structural alignment between generated taxonomies and those crafted by human experts, resulting in insufficient semantic coherence and hierarchical consistency. To address this, we propose TaxoAlign—a three-stage taxonomy generation framework jointly guided by topic modeling and instruction tuning, integrating topic-aware representation learning, large language model–driven hierarchical expansion, and structure-aware refinement. To enable rigorous evaluation, we introduce CS-TaxoBench, the first high-quality, expert-annotated benchmark for computer science taxonomies, and design the first automated, quantitative evaluation framework for structural alignment and semantic coherence. Experimental results demonstrate that TaxoAlign significantly outperforms all baselines in both automated metrics and human evaluations, achieving breakthrough improvements in hierarchical plausibility, cross-level semantic consistency, and expert alignment.
0 citationsRead paper
A Cegar-centric Bounded Reachability Analysis for Compositional Affine Hybrid Systems
Addressing the dual challenges of semantic preservation and state-space explosion in bounded reachability analysis for compositional affine hybrid systems, this paper proposes a novel counterexample-guided abstraction refinement (CEGAR) framework. Instead of explicitly constructing the product automaton, our approach performs discrete abstract search to identify counterexamples and drives iterative state-space refinement. At the abstraction level, we employ step-wise compositional semantics to accelerate search; at the refinement level, we integrate shallow compositional semantics with symbolic reachability analysis. Continuous states are compactly represented via support functions, and intermediate results are cached for reuse. We implement the method in the tool SAT-Reach. Experimental evaluation demonstrates significant improvements in scalability and computational efficiency for compositional hybrid systems, providing a more practical and automated verification approach for complex hybrid systems.
0 citationsRead paper
Recent publications
Latest Papers
One Size Does Not Fit All! Dynamic Retriever and Generator Selection for RAG
研究通过动态选择检索器和生成器来优化RAG系统的资源分配问题，提出DRAG框架以适应不同查询需求，提高效率。
0 citationsRead paper
Automatic Generation of Titles for Research Papers Using Language Models
This study addresses the time-consuming and challenging task of academic paper title generation by proposing an automated approach that produces titles directly from abstracts. The authors introduce SpringerSSAT, a novel and publicly released dataset in the social sciences, and conduct a systematic evaluation of several large language models—including PEGASUS-large, LLaMA-3-8B, and GPT-3.5-turbo—for this task. Experimental results demonstrate that fine-tuned PEGASUS-large consistently outperforms other models across multiple metrics such as ROUGE, METEOR, and BERTScore, generating titles that are accurate, concise, and practically useful. The study also highlights the potential of ChatGPT in producing creative and engaging titles, suggesting complementary strengths among different model architectures for scholarly title generation.
0 citationsRead paper
An Evaluation of Context Length Extrapolation in Long Code via Positional Embeddings and Efficient Attention
This work addresses the challenge that large language models, constrained by fixed context lengths, struggle to effectively process code sequences exceeding their training length. The authors propose a zero-shot inference method that requires no fine-tuning, leveraging optimized positional encoding and efficient attention mechanisms to systematically evaluate the context-length extrapolation capabilities of existing techniques on long code completion tasks. For the first time, they provide a comprehensive comparison of multiple positional embedding schemes and attention strategies in ultra-long code scenarios, revealing significant differences in their extrapolation performance. This study offers empirical evidence and practical guidance for enhancing the ability of large models to handle extremely long code sequences.
0 citationsRead paper
TaxoAlign: Scholarly Taxonomy Generation Using Language Models
Existing automated academic survey generation methods lack systematic structural alignment between generated taxonomies and those crafted by human experts, resulting in insufficient semantic coherence and hierarchical consistency. To address this, we propose TaxoAlign—a three-stage taxonomy generation framework jointly guided by topic modeling and instruction tuning, integrating topic-aware representation learning, large language model–driven hierarchical expansion, and structure-aware refinement. To enable rigorous evaluation, we introduce CS-TaxoBench, the first high-quality, expert-annotated benchmark for computer science taxonomies, and design the first automated, quantitative evaluation framework for structural alignment and semantic coherence. Experimental results demonstrate that TaxoAlign significantly outperforms all baselines in both automated metrics and human evaluations, achieving breakthrough improvements in hierarchical plausibility, cross-level semantic consistency, and expert alignment.
0 citationsRead paper
A Cegar-centric Bounded Reachability Analysis for Compositional Affine Hybrid Systems
Addressing the dual challenges of semantic preservation and state-space explosion in bounded reachability analysis for compositional affine hybrid systems, this paper proposes a novel counterexample-guided abstraction refinement (CEGAR) framework. Instead of explicitly constructing the product automaton, our approach performs discrete abstract search to identify counterexamples and drives iterative state-space refinement. At the abstraction level, we employ step-wise compositional semantics to accelerate search; at the refinement level, we integrate shallow compositional semantics with symbolic reachability analysis. Continuous states are compactly represented via support functions, and intermediate results are cached for reuse. We implement the method in the tool SAT-Reach. Experimental evaluation demonstrates significant improvements in scalability and computational efficiency for compositional hybrid systems, providing a more practical and automated verification approach for complex hybrid systems.
0 citationsRead paper