Mixture-of-Experts (MoE) architectures have recently become a more prevalent choice for large language models (LLMs) than dense architectures due to their superior performance. However, billions of parameters bring MoE LLMs a huge cost for deployment and inference. To address these issues, knowledge distillation (KD) has become a widely adopted technique to compress LLMs. Existing KD methods for LLMs can be divided into dense-to-dense and moe-to-dense distillation. Dense-to-dense distillation transfers knowledge between single dense LLMs, while moe-to-dense distillation attempts to transfer knowledge between the MoE LLMs and the dense LLMs. However, the architectural mismatch prevents the student from fully absorbing knowledge when distilling MoE LLMs. To address this limitation, we investigate a new distillation setting, moe-to-moe, which aims to fully leverage expert knowledge of teachers and enable the student to absorb it more effectively. Compared to dense-to-dense and moe-to-dense, moe-to-moe suffers from two imbalance issues. First, expert-coverage deficiency reflects an imbalanced knowledge transfer of teacher experts: traditional distillation utilizes only the few experts activated by the teacher router. Second, routing imbalance appears when the student routing distribution drifts from the teacher, which makes it difficult for students to learn how to distribute different experts. To overcome these issues, we propose a novel distillation framework for moe-to-moe, Balanced Distillation (B-Distill), which equally spreads teacher expertise across student experts while regularizing the student router toward teacher-consistent balance. First, to mitigate expert-coverage deficiency, we introduce Monte Carlo exploration, which stochastically perturbs router probabilities so every teacher and student expert is sampled without enlarging the search space. Second, to correct routing imbalance and avert load collapse, we propose an entropy-aware router distillation mechanism that aligns the student router with the teacher while curbing over-concentration. Experiments show that B-Distill outperforms baselines by up to 6.6% in Rouge-L.

Mixture of experts (MoE) dynamically routes inputs to specialized expert networks to scale model capacity with low inference overhead. However, the excessive parameter growth in MoE models poses challenges in low-resource settings. To address these issues, MoE with parameter-efficient fine-tuning (PEFT) methods have emerged as a lightweight adaptation paradigm that distributes knowledge among experts via multiple LoRA blocks. Existing MoE-PEFT methods can be broadly categorized into External and Internal PEFT methods. External PEFT methods incorporate lightweight models into existing MoE architectures without modifying their routing, which limits the model’s parameter efficiency. To overcome these issues, Internal PEFT methods integrate MoE architectures into PEFT, enabling minimal parameter overhead. However, they still face two major challenges: (1) lack of expert functional differentiation, resulting in overlapping specialization across modules, and (2) absence of a structured attribution mechanism to guide expert selection based on semantic relevance. To alleviate these challenges, we propose TopicLoRA, a novel three-stage framework that leverages topic knowledge as semantic anchors to guide expert allocation. Specifically, (1) to address expert redundancy, we construct a topic-level prior graph using Graph Neural Network-enhanced representation learning over Big-Bench categories, enforcing structural separation among expert embeddings, and (2) to introduce semantic attribution, we design a dual-loss training mechanism that softly aligns input-query relevance with topic-guided routing distributions via KL divergence. Extensive experiments on representative datasets (e.g., MMLU, GSM8K, Flanv2) demonstrate that TopicLoRA outperforms state-of-the-art PEFT baselines by 2.40% on average in accuracy. Notably, the maximum improvement is 4.21%. Furthermore, ablation studies demonstrate that our framework's robustness to intricate topics and input sequence variations, which stems from the dual-loss training mechanism.

Large Language Models (LLMs) have demonstrated strong capabilities across diverse tasks under the example-driven learning paradigm. However, in high-stakes domains such as emergency response and industrial safety, historical incidents are scarce, confidential, or both, while concise rule books are abundant. We formalize this underexplored setting as rule knowledge-driven reasoning and ask: Can LLMs reason reliably when rules are plentiful but examples are nearly absent? To study this question, we introduce RULER, an automatic benchmark that generates 32K rigorously verified questions from 1K expert-curated emergency response rules to probe three core abilities: rule memorization, single-rule application, and multi-rule complex reasoning. RULER is further equipped with a hallucination-aware evaluation suite and novel relational metrics. A comprehensive empirical study of five representative LLMs and five enhancement strategies shows that, even when models achieve reliable performance on rule memorization and single-rule application, multi-rule complex reasoning plateaus at 5.4 on a 10-point scale. To address this limitation, we propose RAMPS, a Rule knowledge-Aware Monte Carlo Tree Search Process-reward Supervision framework. RAMPS injects rule knowledge priors into MCTS, distills 12K step-level traces without human annotation, and trains an advantage-based reward model that scores candidate reasoning paths during beam search inference. Experimental results show that RAMPS significantly improves multi-rule complex reasoning performance to 7.7.

Relation extraction (RE) identifies semantic relations between entities in text, with existing methods falling into two main paradigms: discriminative and generative.Discriminative models encode sentences and entities into relation representations and classify the most likely relation, whereas generative models directly produce relation labels through sequence generation.Although the latter have benefited from recent advances in large language models (LLMs), their performance remains limited by bottlenecks.In this work, we present the systematic investigation of how discriminative models can support generative RE.We propose the Discriminative-to-Generative (D2G) framework, which first leverages discriminative models to produce a top-k set of candidate relations, and then integrates this knowledge into generative models via in-context or prompt learning.Extensive experiments on five benchmarks demonstrate that D2G consistently achieves state-of-the-art performance, with notable gains on long-tailed relation classes.

Chuanxin Zhang, Jiajun Liu, Yao He, Wenjun Ke, Peng Wang, Yankun Le, Sirui Liu, Zhaoyu Yang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Unified Information Extraction (UIE) aims to handle heterogeneous IE tasks within a single framework, but existing methods often suffer from inconsistent schema representation, implicitly intermediate reasoning and fullparameter adaptation, which limit generalization, interpretability and parameter efficiency.To address these issues, we propose UC-UIE (Universal Capabilities-based Unified Information Extractor), a unified framework based on Large Language Model (LLM), which introduces a unified frame-and-slots schema for IE tasks and explicitly decomposes IE reasoning into three universal capabilities: judging, locating, and associating.Furthermore, UC-UIE adopts a Low-Rank Adaptation (LoRA) based hierarchical Mixture-of-Experts (MoE) adapter to fine-tune LLMs for IE tasks, which explicitly models these three capabilities in a task-driven way while ensuring parameter efficiency.With only 1.24% trainable parameters, UC-UIE outperforms full-parameter tuning methods, showing excellent parameter efficiency.Zero-shot evaluation reveals its strong generalization ability to unseen domains and schemas, benefiting from unified schema representation and explicit capability decomposition.Further experiments validate that the hierarchical MoE adapter learns capability specialization and composition, which enhances both UIE performance and interpretability.

Existing knowledge graphs (KGs) inevitably contain outdated or erroneous knowledge that needs to be removed from knowledge graph embedding (KGE) models. To address this challenge, knowledge unlearning can be applied to eliminate specific information while preserving the integrity of the remaining knowledge in KGs. Existing unlearning methods can generally be categorized into exact unlearning and approximate unlearning. However, exact unlearning requires high training costs, while approximate unlearning faces two issues when applied to KGs due to the inherent connectivity of triples: (1) It fails to fully remove targeted information, as forgetting triples can still be inferred from remaining ones. (2) It focuses on local data for specific removal, which weakens the remaining knowledge in the forgetting boundary. To address these issues, we propose GraphDPO, a novel approximate unlearning framework based on direct preference optimization (DPO). Firstly, to effectively remove forgetting triples, we reframe unlearning as a preference optimization problem, where the model is trained by DPO to prefer reconstructed alternatives over the original forgetting triples. This formulation penalizes reliance on forgettable knowledge, mitigating incomplete forgetting caused by KG connectivity. Moreover, we introduce an out-boundary sampling strategy to construct preference pairs with minimal semantic overlap, weakening the connection between forgetting and retained knowledge. Secondly, to preserve boundary knowledge, we introduce a boundary recall mechanism that replays and distills relevant information both within and across time steps. We construct eight unlearning datasets across four popular KGs with varying unlearning rates. Experiments show that GraphDPO outperforms state-of-the-art baselines by up to 10.1% in MRR_Avg and 14.0% in MRR_F1. Further analysis confirms that GraphDPO more effectively removes target knowledge while preserving surrounding context.

Topic modeling aims to discover the distribution of topics within a corpus. The advanced comprehension and generative capabilities of large language models (LLMs) have introduced new avenues for topic modeling, particularly by prompting LLMs to generate topics and refine them by merging similar ones. However, this approach necessitates that LLMs generate topics with consistent granularity, thus relying on the exceptional instruction-following capabilities of closed-source LLMs (such as GPT-4) or requiring additional training. Moreover, merging based only on topic words and neglecting the fine-grained semantics within documents might fail to fully uncover the underlying topic structure. In this work, we propose a semi-supervised topic modeling method, LiSA, that combines LLMs with clustering to improve topic generation and distribution. Specifically, we begin with prompting LLMs to generate a candidate topic word for each document, thereby constructing a topic-level semantic space. To further utilize the mutual complementarity between them, we first cluster documents and candidate topic words, and then establish a mapping from document to topic in the LLM-guided assignment stage. Subsequently, we introduce a collaborative enhancement strategy to align the two semantic spaces and establish a better topic distribution. Experimental results demonstrate that LiSA outperforms state-of-the-art methods that utilize GPT-4 on topic alignment, and exhibits competitive performance compared to Neural Topic Models on topic quality. The codes are available at https://github.com/ljh986/LiSA.

Ziyu Shang, Jianghan Liu, Zhizhao Luo, Peng Wang, Wenjun Ke, Jiajun Liu, Zijie Xu, Guozheng Li. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Prompt optimization aims to construct effective prompts to fully leverage the capabilities of large language models (LLMs). The key challenge lies in analyzing the outputs of LLMs, identifying fine-grained errors as explicit evidence and guiding reverse prompt correction. However, existing dominant methods feed all evidence into LLMs, leading to excessive redundancy that misleads the models. Moreover, this redundant input could increase computational costs. To address the issue, we introduce a novel meta evidence-aware prompt optimization method (MEPO), which performs fine-grained analysis of the causal relationships within evidence using clustering and Bayesian inference. Initially, MEPO generates preliminary evidence to describe the drawbacks of the given prompt. Subsequently, the evidence is clustered and constructed into a directed acyclic graph (DAG) to produce meta evidence. Finally, we edit the initial prompt based on the obtained meta evidence and employ the beam search to select the best-performing prompt. Experimental results show that MEPO outperforms state-of-the-art baselines, achieving absolute F1 score improvements of 5.8% and 5.0% on two common-used Liar and Jailbreak datasets, respectively. The datasets and codes are available at https://anonymous.4open.science/r/MEPO-E1A0.

Knowledge graph reasoning (KGR) seeks to infer new factual triples from existing knowledge graphs (KGs). Recent methods have unified transductive and inductive reasoning by learning entity-independent representations through local neighboring structures. Nevertheless, these methods often encounter inefficiencies and rely on elaborate local structures without directly modeling the correlations between queries and various structures within KGs. In this paper, we propose a novel framework MulGA, which is designed to learn multi-granularity and adaptive embeddings for KGR. MulGA first employs connectivity subgraphs to uniformly and hierarchically represent query-related structures within KGs, such as triples, relation paths, and subgraphs, establishing the hierarchical relationship between structures at different granularities. Subsequently, we design a graph neural network-based multi-granularity embedding propagation module that unifies the message-passing process with the connectivity subgraph construction. This module obtains the query-related structural representations by all entities at multiple granularities, eliminating the need to explicitly extract any graph elements, thus addressing inefficiency issues. Moreover, we develop a structure-aware adaptive merging mechanism that assigns weights to different granularities and integrates them into cohesive subgraph-granularity representations for reasoning. The systematic experiments have been conducted on 15 benchmarks and MulGA achieves a significant improvement in MRR by an average of 0.5%-1.1% on transductive tasks and 0.2%-7.3% on inductive tasks than existing state-of-the-art methods. Moreover, MulGA exhibits faster convergence speed, smaller number of parameters, competitive inference time, and alleviates the over-smoothing prevalent in graph neural networks.

Traditional knowledge graph embedding (KGE) methods typically require preserving the entire knowledge graph (KG) with significant training costs when new knowledge emerges. To address this issue, the continual knowledge graph embedding (CKGE) task has been proposed to train the KGE model by learning emerging knowledge efficiently while simultaneously preserving decent old knowledge. However, the explicit graph structure in KGs, which is critical for the above goal, has been heavily ignored by existing CKGE methods. On the one hand, existing methods usually learn new triples in a random order, destroying the inner structure of new KGs. On the other hand, old triples are preserved with equal priority, failing to alleviate catastrophic forgetting effectively. In this paper, we propose a competitive method for CKGE based on incremental distillation (IncDE), which considers the full use of the explicit graph structure in KGs. First, to optimize the learning order, we introduce a hierarchical strategy, ranking new triples for layer-by-layer learning. By employing the inter- and intra-hierarchical orders together, new triples are grouped into layers based on the graph structure features. Secondly, to preserve the old knowledge effectively, we devise a novel incremental distillation mechanism, which facilitates the seamless transfer of entity representations from the previous layer to the next one, promoting old knowledge preservation. Finally, we adopt a two-stage training paradigm to avoid the over-corruption of old knowledge influenced by under-trained new knowledge. Experimental results demonstrate the superiority of IncDE over state-of-the-art baselines. Notably, the incremental distillation mechanism contributes to improvements of 0.2%-6.5% in the mean reciprocal rank (MRR) score. More exploratory experiments validate the effectiveness of IncDE in proficiently learning new knowledge while preserving old knowledge across all time steps.

Large language models (LLMs) have demonstrated impressive proficiency in information retrieval, while they are prone to generating incorrect responses that conflict with reality, a phenomenon known as intrinsic hallucination. The critical challenge lies in the unclear and unreliable fact distribution within LLMs trained on vast amounts of data. The prevalent approach frames the factual detection task as a question-answering paradigm, where the LLMs are asked about factual knowledge and examined for correctness. However, existing studies primarily focused on deriving test cases only from several specific domains, such as movies and sports, limiting the comprehensive observation of missing knowledge and the analysis of unexpected hallucinations. To address this issue, we propose OntoFact, an adaptive framework for detecting unknown facts of LLMs, devoted to mining the ontology-level skeleton of the missing knowledge. Specifically, we argue that LLMs could expose the ontology-based similarity among missing facts and introduce five representative knowledge graphs (KGs) as benchmarks. We further devise a sophisticated ontology-driven reinforcement learning (ORL) mechanism to produce error-prone test cases with specific entities and relations automatically. The ORL mechanism rewards the KGs for navigating toward a feasible direction for unveiling factual errors. Moreover, empirical efforts demonstrate that dominant LLMs are biased towards answering Yes rather than No, regardless of whether this knowledge is included. To mitigate the overconfidence of LLMs, we leverage a hallucination-free detection (HFD) strategy to tackle unfair comparisons between baselines, thereby boosting the result robustness. Experimental results on 5 datasets, using 32 representative LLMs, reveal a general lack of fact in current LLMs. Notably, ChatGPT exhibits fact error rates of 51.6% on DBpedia and 64.7% on YAGO, respectively. Additionally, the ORL mechanism demonstrates promising error prediction scores, with F1 scores ranging from 70% to 90% across most LLMs. Compared to the exhaustive testing, ORL achieves an average recall of 80% while reducing evaluation time by 35.29% to 63.12%.

The in-context learning (ICL) for relational triple extraction (RTE) has achieved promising performance, but still encounters two key challenges: (1) how to design effective prompts and (2) how to select proper demonstrations. Existing methods, however, fail to address these challenges appropriately. On the one hand, they usually recast RTE task to text-to-text prompting formats, which is unnatural and results in a mismatch between the output format at the pre-training time and the inference time for large language models (LLMs). On the other hand, they only utilize surface natural language features and lack consideration of triple semantics in sample selection. These issues are blocking improved performance in ICL for RTE, thus we aim to tackle prompt designing and sample selection challenges simultaneously. To this end, we devise a tabular prompting for RTE (TableIE) which frames RTE task into a table generation task to incorporate explicit structured information into ICL, facilitating conversion of outputs to RTE structures. Then we propose instructive in-context learning (I^2CL) which only selects and annotates a few samples considering internal triple semantics in massive unlabeled samples. Specifically, we first adopt off-the-shelf LLMs to perform schema-agnostic pre-extraction of triples in unlabeled samples using TableIE. Then we propose a novel triple-level similarity metric considering triple semantics between these samples and train a sample retrieval model based on calculated similarities in pre-extracted unlabeled data. We also devise three different sample annotation strategies for various scenarios. Finally, the annotated samples are considered as few-shot demonstrations in ICL for RTE. Experimental results on two RTE benchmarks show that I^2CL with TableIE achieves state-of-the-art performance compared to other methods under various few-shot RTE settings.

The immense parameter scale of large language models underscores the necessity for parameter-efficient fine-tuning methods. Methods based on Low-Rank Adaptation (LoRA) assume the low-rank characteristics of the incremental matrix and optimize the matrix obtained from low-rank decomposition. Although effective, these methods are constrained by a fixed and unalterable intrinsic rank, neglecting the variable importance of matrices. Consequently, methods for adaptive rank allocation are proposed, among which AdaLoRA demonstrates excellent fine-tuning performance. AdaLoRA conducts adaptation based on singular value decomposition (SVD), dynamically allocating intrinsic ranks according to importance. However, it still struggles to achieve a balance between fine-tuning effectiveness and efficiency, leading to limited rank allocation space. Additionally, the importance measurement focuses only on parameters with minimal impact on the loss, neglecting the dominant role of singular values in SVD-based matrices and the fluctuations during training. To address these issues, we propose SalientLoRA, which adaptively optimizes intrinsic ranks of LoRA via salience measurement. Firstly, during rank allocation, the salience measurement analyses the variation of singular value magnitudes across multiple time steps and establishes their inter-dependency relationships to assess the matrix importance. This measurement mitigates instability and randomness that may arise during importance assessment. Secondly, to achieve a balance between fine-tuning performance and efficiency, we propose an adaptive adjustment of time-series window, which adaptively controls the size of time-series for significance measurement and rank reduction during training, allowing for rapid rank allocation while maintaining training stability. This mechanism enables matrics to set a higher initial rank, thus expanding the allocation space for ranks. To evaluate the generality of our method across various tasks, we conduct experiments on natural language understanding (NLU), natural language generation (NLG), and large model instruction tuning tasks. Experimental results demonstrate the superiority of SalientLoRA, which outperforms state-of-the-art methods by 0.96\%-3.56\% on multiple datasets. Furthermore, as the rank allocation space expands, our method ensures fine-tuning efficiency, achieving a speed improvement of 94.5\% compared to AdaLoRA. The code is publicly available at https://github.com/Heyest/SalientLoRA.

Knowledge distillation for knowledge graph embedding (KGE) aims to reduce the KGE model size to address the challenges of storage limitations and knowledge reasoning efficiency. However, current work still suffers from the performance drops when compressing a high-dimensional original KGE model to a low-dimensional distillation KGE model. Moreover, most work focuses on the reduction of inference time but ignores the time-consuming training process of distilling KGE models. In this paper, we propose IterDE, a novel knowledge distillation framework for KGEs. First, IterDE introduces an iterative distillation way and enables a KGE model to alternately be a student model and a teacher model during the iterative distillation process. Consequently, knowledge can be transferred in a smooth manner between high-dimensional teacher models and low-dimensional student models, while preserving good KGE performances. Furthermore, in order to optimize the training process, we consider that different optimization objects between hard label loss and soft label loss can affect the efficiency of training, and then we propose a soft-label weighting dynamic adjustment mechanism that can balance the inconsistency of optimization direction between hard and soft label loss by gradually increasing the weighting of soft label loss. Our experimental results demonstrate that IterDE achieves a new state-of-the-art distillation performance for KGEs compared to strong baselines on the link prediction task. Significantly, IterDE can reduce the training time by 50% on average. Finally, more exploratory experiments show that the soft-label weighting dynamic adjustment mechanism and more fine-grained iterations can improve distillation performance.