Zero-shot Dialog State Tracking (zs-DST) is essential for enabling Task-Oriented Dialog Systems (TODs) to generalize to new domains without costly data annotation. A central challenge lies in the semantic misalignment between dynamic dialog contexts and static prompts, leading to inflexible cross-layer coordination, domain interference, and catastrophic forgetting. To tackle this, we propose Hierarchical Collaborative Low-Rank Adaptation (HiCoLoRA), a framework that enhances zero-shot slot inference through robust prompt alignment. It features a hierarchical LoRA architecture for dynamic layer-specific processing (combining lower-layer heuristic grouping and higher-layer full interaction), integrates Spectral Joint Domain-Slot Clustering to identify transferable associations (feeding an Adaptive Linear Fusion Mechanism), and employs Semantic-Enhanced SVD Initialization (SemSVD-Init) to preserve pre-trained knowledge. Experiments on multi-domain datasets MultiWOZ and SGD show that HiCoLoRA outperforms baselines, achieving SOTA in zs-DST.

LLM-based sequential recommenders usually cast next-item prediction as text generation, but this interface is poorly matched to full-catalog top-K ranking. We propose TATK, a Triple-Aware framework that couples Top-K Learning (TKL) with Knowledge-Grounded Verification (KGV) for LLM-based sequential recommendation. Top-K Learning combines context-aware metadata-KG prompt grounding with position-aware top-K rewards, aligning training with ranking utility; Knowledge-Grounded Verification then applies structure-aware reranking over the top-M candidates after a single LLM forward pass, using the same metadata-derived item graph. We evaluate TATK on Musical Instruments, CDs and Vinyl, and Video Games from Amazon Reviews 2023 under a matched R²ec-style full-catalog protocol. Experiments use Gemma-2-2B-It and Qwen2.5-3B-Instruct backbones, compare against sequential, generative, KG-augmented, and reasoning-enhanced baselines, and include component, reward-shape, sequence-perturbation, reranking, relation-quality, and candidate-pool diagnostics. TATK improves over the matched R²ec reproduction on all 36 reported metrics. On NDCG@10, it improves Qwen by 8.05%, 4.26%, and 3.78% on the three datasets, and improves Gemma by 27.03%, 10.52%, and 10.23%, while keeping inference within 1.17× of Base RecPO latency. The diagnostics show that structural evidence is most useful for recoverable top-M candidates with reliable KG support, and should be gated when metadata relations are sparse or noisy.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance across a wide range of multimodal tasks. Recent advances in personalized MLLMs enable effective capture of user-specific concepts, supporting both recognition of personalized concepts and contextual captioning. However, humans typically explore and reason over relations among objects and individuals, transcending surface-level information to achieve more personalized and contextual understanding. To this end, existing methods may face three main limitations: (1) Their training data lacks multi-object sets in which relations among objects are learnable, (2) Existing models often neglect the connections between different personalized concepts, thereby failing to perform reasoning over them, (3) Their experiments mainly focus on a single personalized concept, where evaluations are limited to recognition and captioning tasks. To address the limitations, (i) We present a new dataset named ReGraP, consisting of 120 sets of personalized knowledge. Each set includes images, Knowledge Graphs (KGs), and Chain-of-Thought Question-Answering (CoT QA) pairs derived from the KGs, enabling more structured and sophisticated reasoning pathways. (ii) We propose Reasoning enabled Graph-based Personalized Large Language and Vision Assistant (ReGraP-LLaVA), an MLLM trained with the corresponding KGs and CoT QA pairs, where soft and/or hard graph prompting methods are designed to align KGs within the model’s semantic space. (iii) We establish the ReGraP Benchmark, which contains diverse task types: Multiple-Choice, Fill-in-the-blank, True/False, and Descriptive questions in both open- and closed-ended settings. The proposed benchmark is designed to evaluate the relational reasoning and knowledge-connection capability of personalized MLLMs. We conduct experiments on the proposed ReGraP-LLaVA and other competitive MLLMs. Results show that the proposed model not only learns personalized knowledge but also performs relational reasoning in responses, achieving the SoTA performance compared with the competitive methods. All the codes and datasets are released at: https://anonymous.4open.science/r/ReGraP.

Text-based role-playing models can imitate character styles, but often fail to capture scene atmosphere and evolving tension, which are crucial for immersive applications such as VR games and interactive narratives. We study video-grounded role-playing dialogue and introduce EBM-RL (Eye-Brain-Mouth Reinforcement Learning), a decoupled GRPO-based framework that separates observation ( lt;perception gt;), reasoning ( lt;think gt;), and utterance generation ( lt;answer gt;). This design mimics the human See-Think-Speak process, enabling the model to ground dialogue in visual perception before reasoning and response generation. To optimize this See-Think-Speak process, EBM-RL integrates complementary rewards for scene-text alignment, perceptual-cognitive utility, answer faithfulness, and format consistency. Extensive experiments show that EBM-RL substantially outperforms text-only role-playing baselines and comparable-scale VLM baselines, while remaining competitive with stronger larger-scale VLMs on visual grounding and situational persona metrics. Moreover, EBM-RL demonstrates strong zero-shot transfer to out-of-domain VideoQA benchmarks without additional fine-tuning. We also release an open-source dataset for video-grounded role-playing dialogue.

Task oriented dialog systems often rely on static exploration strategies that do not adapt to dynamic dialog contexts, leading to inefficient exploration and suboptimal performance. We propose DyBBT, a novel dialog policy learning framework that formalizes the exploration challenge through a structured cognitive state space 𝒞 that captures dialog progression, user uncertainty, and slot dependency. DyBBT proposes a bandit-inspired meta-controller that dynamically switches between a fast intuitive inference (System 1) and a slow deliberative reasoner (System 2) based on real-time cognitive states and visitation counts. Extensive experiments on single- and multi-domain benchmarks show that DyBBT achieves SOTA performance in success rate, efficiency, and generalization, with human evaluations confirming that its decisions are well-aligned with expert judgment.

Traditional task-oriented dialog systems are unable to evolve from ongoing interactions or adapt to new domains after deployment, that is a critical limitation in real-world dynamic environments. Continual learning approaches depend on episodic retraining with human-curated data, failing to achieve autonomy lifelong improvement. While evolutionary computation and LLM driven self-improvement offer promising mechanisms for dialog optimization, they lack a unified framework for holistic, iterative strategy refinement. To bridge this gap, we propose DarwinTOD, a lifelong self-evolving dialog framework that systematically integrates these two paradigms, enabling continuous strategy optimization from a zero-shot base without task-specific fine-tuning. DarwinTOD maintains an Evolvable Strategy Bank and operates through a dual-loop process: online multi-agent dialog execution with peer critique, and offline structured evolutionary operations that refine the strategy bank using accumulated feedback. This closed-loop design enables autonomous continuous improvement without human intervention. Extensive experiments show that DarwinTOD surpasses previous state-of-the-art methods and exhibits continuous performance gains throughout evolution. Our work provides a novel framework for building dialog systems with lifelong self-evolution capabilities.

Large-scale generative models like DeepSeek-R1 and OpenAI-O1 benefit substantially from chain-of-thought (CoT) reasoning, yet pushing their performance typically requires vast data, large model sizes, and full-parameter fine-tuning. While parameter-efficient fine-tuning (PEFT) helps reduce cost, most existing approaches primarily address domain adaptation or layer-wise allocation rather than explicitly tailoring data and parameters to different response demands. Inspired by “Thinking, Fast and Slow,” which characterizes two distinct modes of thought—System 1 (fast, intuitive, often automatic) and System 2 (slower, more deliberative and analytic)—we draw an analogy that different “subregions” of an LLM’s parameters might similarly specialize for tasks that demand quick, intuitive responses versus those requiring multi-step logical reasoning. Therefore, we propose LoRA-PAR, a dual-system LoRA framework that partitions both data and parameters by System 1 or System 2 demands, using fewer yet more focused parameters for each task. Specifically, we classify task data via multi-model role-playing and voting, and partition parameters based on importance scoring, then adopt a two-stage fine-tuning strategy of training System 1 tasks with supervised fine-tuning (SFT) to enhance knowledge and intuition and refine System 2 tasks with reinforcement learning (RL) to reinforce deeper logical deliberation next. Extensive experiments show that the two-stage fine-tuning strategy, SFT and RL, lowers active parameter usage while matching or surpassing SOTA PEFT baselines.

This paper describes the LingJing team’s method to the Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis (WASSA) 2022 shared task on Personality Prediction (PER) and Reactivity Index Prediction (IRI). In this paper, we adopt the prompt-based method with the pre-trained language model to accomplish these tasks. Specifically, the prompt is designed to provide knowledge of the extra personalized information for enhancing the pre-trained model. Data augmentation and model ensemble are adopted for obtaining better results. Extensive experiments are performed, which shows the effectiveness of the proposed method. On the final submission, our system achieves a Pearson Correlation Coefficient of 0.2301 and 0.2546 on Track 3 and Track 4 respectively. We ranked 1-st on both sub-tasks.

Emotion is the essential attribute of human beings. Perceiving and understanding emotions in a human-like manner is the most central part of developing emotional intelligence. This paper describes the contribution of the LingJing team’s method to the Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis (WASSA) 2022 shared task on Emotion Classification. The participants are required to predict seven emotions from empathic responses to news or stories that caused harm to individuals, groups, or others. This paper describes the continual pre-training method for the masked language model (MLM) to enhance the DeBERTa pre-trained language model. Several training strategies are designed to further improve the final downstream performance including the data augmentation with the supervised transfer, child-tuning training, and the late fusion method. Extensive experiments on the emotional classification dataset show that the proposed method outperforms other state-of-the-art methods, demonstrating our method’s effectiveness. Moreover, our submission ranked Top-1 with all metrics in the evaluation phase for the Emotion Classification task.

This paper introduces the approach of Team LingJing’s experiments on SemEval-2022 Task 1 Comparing Dictionaries and Word Embeddings (CODWOE). This task aims at comparing two types of semantic descriptions and including two sub-tasks: the definition modeling and reverse dictionary track. Our team focuses on the reverse dictionary track and adopts the multi-task self-supervised pre-training for multilingual reverse dictionaries. Specifically, the randomly initialized mDeBERTa-base model is used to perform multi-task pre-training on the multilingual training datasets. The pre-training step is divided into two stages, namely the MLM pre-training stage and the contrastive pre-training stage. The experimental results show that the proposed method has achieved good performance in the reverse dictionary track, where we rank the 1-st in the Sgns targets of the EN and RU languages. All the experimental codes are open-sourced at https://github.com/WENGSYX/Semeval.

This paper presents the results and main findings of our system on SemEval-2022 Task 3 Presupposed Taxonomies: Evaluating Neural Network Semantics (PreTENS). This task aims at semantic competence with specific attention on the evaluation of language models, which is a task with respect to the recognition of appropriate taxonomic relations between two nominal arguments. Two sub-tasks including binary classification and regression are designed for the evaluation. For the classification sub-task, we adopt the DeBERTa-v3 pre-trained model for fine-tuning datasets of different languages. Due to the small size of the training datasets of the regression sub-task, we transfer the knowledge of classification model (i.e., model parameters) to the regression task. The experimental results show that the proposed method achieves the best results on both sub-tasks. Meanwhile, we also report negative results of multiple training strategies for further discussion. All the experimental codes are open-sourced at https://github.com/WENGSYX/Semeval.

Question Answering (QA) is a Natural Language Processing (NLP) task that can measure language and semantics understanding ability, it requires a system not only to retrieve relevant documents from a large number of articles but also to answer corresponding questions according to documents. However, various language styles and sources of human questions and evidence documents form the different embedding semantic spaces, which may bring some errors to the downstream QA task. To alleviate these problems, we propose a framework for enhancing downstream evidence retrieval by generating evidence, aiming at improving the performance of response generation. Specifically, we take the pre-training language model as a knowledge base, storing documents’ information and knowledge into model parameters. With the Child-Tuning approach being designed, the knowledge storage and evidence generation avoid catastrophic forgetting for response generation. Extensive experiments carried out on the multi-documents dataset show that the proposed method can improve the final performance, which demonstrates the effectiveness of the proposed framework.