While large language models (LLMs) have made significant progress in expanding their context windows, they still face great challenges in effectively organizing and utilizing long-term memory to maintain conversation consistency and coherence. Summarizing historical conversations has achieved remarkable performance, which, however, loses conversational trajectory and association, making it difficult to precisely combine memories from different sessions in response to current queries. To address this, we propose the Dynamic Memory Forest (DMF), a novel Consolidation-then-Growth framework for long-term open-domain conversation, which simulates the consolidation and growth processes of human memory by dynamically organizing long-term conversation histories into a memory forest of memory trees. To be specific, inspired by the principles of synaptic consolidation and plasticity from Cognitive Science, we first consolidate each session into memory units that preserve thematic coherence ("Consolidation"). Then, we first structure these units into memory trees and then grow the forest by dynamically connecting them through an evolutionary grafting mechanism, called Group Relative Voting Optimization, which mimics synaptic connection to decide whether a new memory tree should be grafted onto the existing forest or grow independently ("Growth"). For retrieval, we design an Entropy-Driven Memory Walk, constructing a logically coherent memory path via a navigation policy that prioritizes exploring high-entropy nodes. Experiments on three long-term conversation datasets show that our DMF significantly outperforms baselines in enhancing response generation for LLMs.

Cai Ke, Yiming Du, Bin Liang, Yifan Xiang, Lin Gui, Zhongyang Li, Baojun Wang, Yue Yu, Hui Wang, Kam-Fai Wong, Ruifeng Xu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Aspect-Based Sentiment Analysis (ABSA) is a fine-grained sentiment analysis task and has become a significant task with real-world scenario value. The challenge of this task is how to generate an effective text representation and construct an end-to-end model that can simultaneously detect (target, aspect, sentiment) triples from a sentence. Besides, the existing models do not take the heavily unbalanced distribution of labels into account and also do not give enough consideration to long-distance dependence of targets and aspect-sentiment pairs. To overcome these challenges, we propose a novel end-to-end model named Prior-BERT and Multi-Task Learning (PBERT-MTL), which can detect all triples more efficiently. We evaluate our model on SemEval-2015 and SemEval-2016 datasets. Extensive results show the validity of our work in this paper. In addition, our model also achieves higher performance on a series of subtasks of target-aspect-sentiment detection. Code is available at https://github.com/CQUPT-CaiKe/PBERT-MTL.

Structural graph clustering (SCAN) is a popular clustering technique. Using the concept of$\epsilon$-neighborhood, SCAN defines the core vertices that uniquely determine the clusters of a graph. Most existing studies assume that the graph processed by SCAN contains no controlled edges. Few studies, however, have focused on manipulating SCAN by injecting edges. Manipulation of SCAN can be used to assess its robustness and lay the groundwork for developing robust clustering algorithms. To fill this gap and considering the importance of the$\epsilon$-neighborhood for SCAN, we propose a problem, denoted as MN, for manipulating SCAN. The MN problem aims to maximize the$\epsilon$-neighborhood of the target vertex by inserting some edges. On the theoretical side, we prove that the MN problem is both NP-hard and APX-hard, and also is non-submodular and non-monotonic. On the algorithmic side, we design an algorithm by focusing on how to select vertices to join$\epsilon -$neighborhood and thus avoid enumerating edges to report a solution. As a result, our algorithm bypasses the non-monotonicity nature of the MN problem. Extensive experiments on real-world graphs show that our algorithm can effectively solve the proposed MN problem.