Pruning is a promising approach to reduce the high inference cost of large language models (LLMs), but it often comes at the expense of performance. Motivated by the "functional localization" theory in neuroscience, we hypothesize that LLMs contain task-specific expert activation paths, where specific subsets of neurons are co-activated for particular tasks. This structure allows selective activation to preserve task performance while improving inference efficiency. We introduce Sparse Expert Activation Pruning (SEAP), a training-free pruning method for large language models. SEAP identifies task-relevant activation paths by analyzing the clustering patterns of hidden states and neuron activations on a multi-task calibration dataset. Cross-task transfer evaluations confirm the existence of such expert activation structures. SEAP constructs task-aware pruning masks by leveraging a task-expert calibration dataset, which provides representative samples across diverse tasks to reveal their activation signatures. It then employs a lightweight task router to dynamically select relevant computation paths based on the input task. This design significantly reduces inference cost without compromising accuracy. Experimental results show that SEAP retains model performance with only a 1.5% drop on most tasks at 20% sparsity, and at 50% sparsity, it surpasses strong pruning baselines such as WandA and FLAP by over 20%. These results highlight SEAP as a scalable and effective solution for efficient LLM inference.

Retrieval-augmented generation (RAG) provides an efficient solution for expanding the knowledge boundaries of large language models (LLMs), where the indexing serves as a compass to guide LLMs in locating query-relevant external knowledge. Nevertheless, current indexing methods commonly encounter a critical challenge: native indexing is convenient to construct, but it usually disrupts contextual associations and constrains the expressive capacity of rich knowledge. Conversely, knowledge indexing can structure contextual knowledge, but it is often based on preset schemas that limit its generalizability. To address it, we propose a universal and flexible knowledge indexing called pseudo-graph (PG) indexing. During the indexing construction phase, we use the advanced LLMs to transform the knowledge of each raw text into a concise and structured mind map, organizing intra-document knowledge. Subsequently, independent mind maps are linked by associating highly relevant topics or consistent facts across documents, thereby establishing inter-document knowledge connections. Eventually, using the resulting knowledge network PG as the knowledge indexing can circumvent the challenges associated with schema design reliant on preset knowledge and relationship types. During the knowledge retrieval phase, we develop a PG knowledge retriever to mimic human note-reviewing, adaptively navigating and recalling query-relevant knowledge from PG. Experimental results demonstrate that retrieving relevant pseudo-subgraphs from the PG via PG indexing and retriever significantly improves performance in fact-based Q&A, hallucination correction, and two multi-document Q&A tasks, achieving$F1_{QE}$improvements of 15.85%, 8.12%, 3.34%, and 5.73%, respectively, and outperforming the state-of-the-art baseline KGP-LLaMA. Our code is available at:https://github.com/IAAR-Shanghai/PGRAG.

Xun Liang, Simin Niu, Zhiyu Li, Sensen Zhang, Hanyu Wang, Feiyu Xiong, Zhaoxin Fan, Bo Tang, Jihao Zhao, Jiawei Yang, Shichao Song, Mengwei Wang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Retrieval-augmented citation generation (RACG) helps users trust the large language model output by retrieving evidence from reliable sources. However, most current RACG research focuses on single-language tasks, particularly in English, and overlooks the need for cross-lingual evidence retrieval and utilization in real-world applications. To address this issue, we introduce a plug-and-play Retrieval-Augmented Multilingual Citation Generation method (RAMCG) which uses a multilingual retriever to identify relevant evidence from a multilingual knowledge base. The evidence is then combined with the query and processed by a multilingual citation generator. The result is citations that are both accurate and comprehensive. Experiments show that RAMCG outperforms baseline methods in multilingual citation generation and is well-suited for practical use.

Graph Meta-learning methods have improved the performance of few-shot node classification by means of applying meta-learning to the data in non-Euclidean domains. However, most works focus on adopting a single domain, ignoring the fact that tasks in various domains may be distinct, which can cause overfitting problems and thus limit generalizability. To tackle this challenge, we propose a novel Graph Meta-learning framework called Feature-Enhanced Cross-domain Graph Meta-learning that consists of two crucial modules: 1) A feature information extraction module that aims to capture discriminative node importance and simulate various node feature distributions under distinct domains; 2) A heterogeneous graph encoder module that leverages the enhanced node features and topological information to generate task-specific node embeddings with simple fine-tuning. Moreover, we meta-learn the parameters involved to ensure the generalizability in the unseen domains. Results show that our method markedly outperforms the existing state-of-the-art methods.

The continuous advancement of large language models (LLMs) has brought increasing attention to the critical issue of developing fair and reliable methods for evaluating their performance. Particularly, the emergence of cheating phenomena, such as test set leakage and prompt format overfitting, poses significant challenges to the reliable evaluation of LLMs. As evaluation frameworks commonly use Regular Expression (RegEx) for answer extraction, models may adjust their responses to fit formats easily handled by RegEx. Nevertheless, the key answer extraction module based on RegEx frequently suffers from extraction errors. Furthermore, recent studies proposing fine-tuned LLMs as judge models for automated evaluation face challenges in terms of generalization ability and fairness. This paper comprehensively analyzes the entire LLM evaluation chain and demonstrates that optimizing the key answer extraction module improves extraction accuracy and enhances evaluation reliability. Our findings suggest that improving the key answer extraction module can lead to higher judgment accuracy and improved evaluation efficiency compared to the judge models. To address these issues, we propose xFinder, a novel evaluator for answer extraction and matching in LLM evaluation. As part of this process, we create a specialized dataset, the Key Answer Finder (KAF) dataset, to ensure effective model training and evaluation. Generalization tests and real-world evaluations show that the smallest xFinder model, with only 500 million parameters, achieves an average extraction accuracy of 93.42\%. In contrast, RegEx accuracy in the best evaluation framework is 74.38\%. The final judgment accuracy of xFinder reaches 97.61\%, outperforming existing evaluation frameworks and judge models.

Cross-domain Graph Meta-learning (CGML) has shown its promise, where meta-knowledge is extracted from few-shot graph data in multiple relevant but distinct domains. However, several recent efforts assume target data available, which commonly does not established in practice. In this paper, we devise a novel Cross-domain Data Augmentation for Graph Meta-Learning (CDA-GML), which incorporates the superiorities of CGML and Data Augmentation， has addressed intractable shortcomings of label sparsity, domain shift, and the absence of target data simultaneously. Specifically, our method simulates instance-level and task-level domain shift to alleviate the cross-domain generalization issue in conventional graph meta-learning. Experiments show that our method outperforms the existing state-of-the-art methods.

Xun Liang, Shichao Song, Simin Niu, Zhiyu Li, Feiyu Xiong, Bo Tang, Yezhaohui Wang, Dawei He, Cheng Peng, Zhonghao Wang, Haiying Deng. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.