Large Language Models demonstrate strong reasoning capabilities, which can be effectively compressed into smaller models. However, existing datasets and fine-tuning approaches still face challenges that lead to catastrophic forgetting, particularly for models smaller than 8B. First, most datasets typically ignore the relationship between training data knowledge and the model's inherent abilities, making it difficult to preserve prior knowledge. Second, conventional training objectives often fail to constrain inherent knowledge preservation, which can result in forgetting of previously learned skills. To address these issues, we propose a comprehensive solution that alleviates catastrophic forgetting from both the data and fine-tuning approach perspectives. On the data side, we construct a dataset of 5K instances that covers multiple reasoning tasks and incorporates metacognitive knowledge, making it more tolerant and effective for distillation into smaller models. We annotate the metacognitive knowledge required to solve each question and filter the data based on task knowledge and the model's inherent skills. On the training side, we introduce GDPO (Group Direction Preference Optimization), which is better suited for resource-limited scenarios and can efficiently approximate the performance of GRPO. Guided by the large model and by implicitly constraining the optimization path through a reference model, GDPO enables more effective knowledge transfer from the large model and constrains excessive parameter drift. Extensive experiments demonstrate that our approach significantly alleviates catastrophic forgetting and improves reasoning performance on smaller models.

The rapid advancement of Large Language Models (LLMs) poses significant challenges for safety evaluation. Current static datasets struggle to identify emerging vulnerabilities due to three limitations: (1) they risk being exposed in model training data, leading to evaluation bias; (2) their limited prompt diversity fails to capture real-world application scenarios; (3) they are limited to provide human-like multi-turn interactions. To address these limitations, we propose a dynamic evaluation framework, CogSafe, for comprehensive and automated multi-turn safety assessment of LLMs. We introduce CogSafe based on cognitive theories to simulate the real chatting process. To enhance assessment diversity, we introduce scenario simulation and strategy decision to guide the dynamic generation, enabling coverage of application situations. Furthermore, we incorporate the cognitive process to simulate multi-turn dialogues that reflect the cognitive dynamics of real-world interactions. Extensive experiments demonstrate the scalability and effectiveness of our framework, which has been applied to evaluate the safety of widely used LLMs.

A critical challenge for open-domain dialogue agents is to generate persona-relevant and consistent responses. Due to the nature of persona sparsity in conversation scenarios, previous persona-based dialogue agents trained with Maximum Likelihood Estimation tend to overlook the given personas and generate responses irrelevant or inconsistent with personas. To address this problem, we propose a two-stage coarse-to-fine persona-aware training framework to improve the persona consistency of a dialogue agent progressively. Specifically, our framework first trains the dialogue agent to answer the constructed persona-aware questions, making it highly sensitive to the personas to generate persona-relevant responses. Then the dialogue agent is further trained with a contrastive learning paradigm by explicitly perceiving the difference between the consistent and the generated inconsistent responses, forcing it to pay more attention to the key persona information to generate consistent responses. By applying our proposed training framework to several representative baseline models, experimental results show significant boosts on both automatic and human evaluation metrics, especially the consistency of generated responses.

Empathy is a crucial factor in open-domain conversations, which naturally shows one's caring and understanding to others.Though several methods have been proposed to generate empathetic responses, existing works often lead to monotonous empathy that refers to generic and safe expressions.In this paper, we propose to use explicit control to guide the empathy expression and design a framework DIFFUSEMP based on conditional diffusion language model to unify the utilization of dialogue context and attribute-oriented control signals.Specifically, communication mechanism, intent, and semantic frame are imported as multi-grained signals that control the empathy realization from coarse to fine levels.We then design a specific masking strategy to reflect the relationship between multi-grained signals and response tokens, and integrate it into the diffusion model to influence the generative process.Experimental results on a benchmark dataset EMPA-THETICDIALOGUE show that our framework outperforms competitive baselines in terms of controllability, informativeness, and diversity without the loss of context-relatedness.

Empathy is a key ability for a human-like dialogue system. Inspired by social psychology, empathy includes both affective and cognitive aspects. Previous works on this topic have merely focused on recognizing emotions or modeling cognition with commonsense knowledge. Nevertheless, the generated results of these works still have a big gap with human-like empathetic responses. In this paper, we propose Seri, a SkEtching-Reasoning-Integrating framework for empathetic response generation. In particular, we define an empathy planner to capture and reason about multi-source information that considers cognition and affection. Further, we introduce a dynamic integrator module that allows the model dynamically select the appropriate information to generate empathetic responses. Experimental results on EmpatheticDialogue show that our method outperforms competitive baselines and generates responses with higher diversity and cognitive empathy levels.

It is a critical challenge for open-domain dialogue agents to generate context-coherent responses which can present a consistent personality. However, existing methods mainly focus on the penalty of the persona-inconsistent responses, leaving out considering the context-incoherence problem caused by wrong persona selection. In this paper, we propose the Think-Before-You-Speak (TBYS) model, consisting of Concept-guided Persona Reasoning module and Consistent Dialogue Generation module, to explicitly select persona sentences semantically relevant to the current turn and generate responses based on the selection results. The experimental results on Persona-Chat show that TBYS can generate coherent and consistent responses, outperforming state-of-the-art baselines in both automatic and human evaluations.

Recently, knowledge-grounded dialogue has received increasing interest to render the generated responses with more useful and engaging information. However, the knowledge, locally relevant to the user’s utterance, potentially reduces the global coherence of the dialogue. Previous work mainly focuses on retrieving diverse knowledge to assist the response generation whereas resulting in a rough dialogue transition. To alleviate this issue, we propose a History-Adapted Knowledge Copy (HAKC) network to adaptively select context-aware knowledge to ensure the coherence of dialogue. Further, we adopt a contrastive learning framework to enhance the knowledge discrimination ability of HAKC. Experimental results demonstrate the outstanding performance of our model on knowledge selection and response generation tasks as well as the boosted generalization.

Intention, emotion and action are important psychological factors in human activities, which play an important role in the interaction between individuals. How to model the interaction process between individuals by analyzing the relationship of their intentions, emotions, and actions at the cognitive level is challenging. In this paper, we propose a novel cognitive framework of individual interaction. The core of the framework is that individuals achieve interaction through external action driven by their inner intention. Based on this idea, the interactions between individuals can be constructed by establishing relationships between the intention, emotion and action. Furthermore, we conduct analysis on the interaction between individuals and give a reasonable explanation for the predicting results. To verify the effectiveness of the framework, we reconstruct a dataset and propose three tasks as well as the corresponding baseline models, including action abduction, emotion prediction and action generation. The novel framework shows an interesting perspective on mimicking the mental state of human beings in cognitive science.

Knowledge graph embedding, which aims to learn representations of entities and relations in knowledge graphs, finds applications in various downstream tasks. The key to success of knowledge graph embedding models are the ability to model relation patterns including symmetry/antisymmetry, inversion, commutative composition and non-commutative composition. Although existing methods fail in modeling the non-commutative composition patterns, several approaches support this pattern by modeling beyond Euclidean space and complex space. Nevertheless, expanding to complicated spaces such as quaternion can easily lead to a substantial increase in the amount of parameters, which greatly reduces the computational efficiency. In this paper, we propose a new knowledge graph embedding method called RotateCT, which first transforms the coordinates of each entity, and then represents each relation as a rotation from head entity to tail entity in complex space. By design, RotateCT can infer the non-commutative composition patterns and improve the computational efficiency. Experiments on multiple datasets empirically show that RotateCT outperforms most state-of-the-art methods on link prediction and path query answering.

Motivations, emotions, and actions are inter-related essential factors in human activities. While motivations and emotions have long been considered at the core of exploring how people take actions in human activities, there has been relatively little research supporting analyzing the relationship between human mental states and actions. We present the first study that investigates the viability of modeling motivations, emotions, and actions in language-based human activities, named COMMA (Cognitive Framework of Human Activities). Guided by COMMA, we define three natural language processing tasks (emotion understanding, motivation understanding and conditioned action generation), and build a challenging dataset Hail through automatically extracting samples from Story Commonsense. Experimental results on NLP applications prove the effectiveness of modeling the relationship. Furthermore, our models inspired by COMMA can better reveal the essential relationship among motivations, emotions and actions than existing methods.

Controllable story generation is a challenging task in the field of NLP, which has attracted increasing research interest in recent years. However, most existing works generate a whole story conditioned on the appointed keywords or emotions, ignoring the psychological changes of the protagonist. Inspired by psychology theories, we introduce global psychological state chains, which include the needs and emotions of the protagonists, to help a story generation system create more controllable and well-planned stories. In this paper, we propose a Psychology-guided Controllable Story Generation System (PICS) to generate stories that adhere to the given leading context and desired psychological state chains for the protagonist. Specifically, psychological state trackers are employed to memorize the protagonist’s local psychological states to capture their inner temporal relationships. In addition, psychological state planners are adopted to gain the protagonist’s global psychological states for story planning. Eventually, a psychology controller is designed to integrate the local and global psychological states into the story context representation for composing psychology-guided stories. Automatic and manual evaluations demonstrate that PICS outperforms baselines, and each part of PICS shows effectiveness for writing stories with more consistent psychological changes.

Intention, emotion and action are important elements in human activities. Modeling the interaction process between individuals by analyzing the relationships between these elements is a challenging task. However, previous work mainly focused on modeling intention and emotion independently, and neglected of exploring the mutual relationships between intention and emotion. In this paper, we propose a RelAtion Interaction Network (RAIN), consisting of Intention Relation Module and Emotion Relation Module, to jointly model mutual relationships and explicitly integrate historical intention information. The experiments on the dataset show that our model can take full advantage of the intention, emotion and action between individuals and achieve a remarkable improvement over BERT-style baselines. Qualitative analysis verifies the importance of the mutual interaction between the intention and emotion.

Story Ending Generation (SEG) is a challenging task in natural language generation. Recently, methods based on Pre-trained Language Models (PLM) have achieved great prosperity, which can produce fluent and coherent story endings. However, the pre-training objective of PLM-based methods is unable to model the consistency between story context and ending. The goal of this paper is to adopt contrastive learning to generate endings more consistent with story context, while there are two main challenges in contrastive learning of SEG. First is the negative sampling of wrong endings inconsistent with story contexts. The second challenge is the adaptation of contrastive learning for SEG. To address these two issues, we propose a novel Contrastive Learning framework for Story Ending Generation (CLseg)†, which has two steps: multi-aspect sampling and story-specific contrastive learning. Particularly, for the first issue, we utilize novel multi-aspect sampling mechanisms to obtain wrong endings considering the consistency of order, causality, and sentiment. To solve the second issue, we well-design a story-specific contrastive training strategy that is adapted for SEG. Experiments show that CLseg outperforms baselines and can produce story endings with stronger consistency and rationality.

Emotional support conversation aims at reducing the emotional distress of the help-seeker, which is a new and challenging task. It requires the system to explore the cause of help-seeker's emotional distress and understand their psychological intention to provide supportive responses. However, existing methods mainly focus on the sequential contextual information, ignoring the hierarchical relationships with the global cause and local psychological intention behind conversations, thus leads to a weak ability of emotional support. In this paper, we propose a Global-to-Local Hierarchical Graph Network to capture the multi-source information (global cause, local intentions and dialog history) and model hierarchical relationships between them, which consists of a multi-source encoder, a hierarchical graph reasoner, and a global-guide decoder. Furthermore, a novel training objective is designed to monitor semantic information of the global cause. Experimental results on the emotional support conversation dataset, ESConv, confirm that the proposed GLHG has achieved the state-of-the-art performance on the automatic and human evaluations.

Intention, emotion and action are important psychological factors in human activities, which play an important role in the interaction between individuals. How to model the interaction process between individuals by analyzing the relationship of their intentions, emotions, and actions at the cognitive level is challenging. In this paper, we propose a novel cognitive framework of individual interaction. The core of the framework is that individuals achieve interaction through external action driven by their inner intention. Based on this idea, the interactions between individuals can be constructed by establishing relationships between the intention, emotion and action. Furthermore, we conduct analysis on the interaction between individuals and give a reasonable explanation for the predicting results. To verify the effectiveness of the framework, we reconstruct a dataset and propose three tasks as well as the corresponding baseline models, including action abduction, emotion prediction and action generation. The novel framework shows an interesting perspective on mimicking the mental state of human beings in cognitive science.

End-to-End open-domain dialogue systems suffer from the issues of generating inconsistent and repetitive responses. Existing dialogue models pay attention to unilaterally incorporating personalized knowledge into the dialogue to enhance the quality of generated response. However, they ignore that incorporating the personality-related information from dialogue history into personalized knowledge can boost the subsequent dialogue quality. In this paper, A Knowledge-Conversation Cyclic Utilization Mechanism (KC2UM) is proposed to enhance the dialogue quality. Specifically, A novel cyclic interaction module is designed to iteratively incorporate personalized knowledge into each turn conversation and capture the personality-related conversation information to enhance personalized knowledge semantic representation. We represent the knowledge with semantic and utilization representations to keep track of the personalized knowledge utilization. Experiments on two knowledge-grounded dialogue datasets show that our approach manages to select knowledge more accurately and generates more informative responses.

Task-oriented dialogues have one-to-many property from semantic and syntactic perspectives, with many suitable dialogue acts and syntactic forms for a given post. However, current state-of-the-art task-oriented dialogue systems attempt to improve the quality of dialogues in terms of the most popular metrics (i.e., BLEU and entity F1), measuring the similarity between the generated responses and the human annotations, while the diversity of task-oriented dialogues remains less explored. This paper aims to improve the diversity of task-oriented dialogues from both semantic and syntactic perspectives by proposing a structural causal model to learn the causality composition of the dialogue acts and syntactic forms. Specifically, the disentangled understanding module decouples the dialogue into semantic and syntactic spaces and learns one-to-many property with multiple reference training. Then the casual collaboration generation module is proposed to apply Structural Causal Mechanism (SCM) to learn the causality composition relationship of the semantic and syntactic representations to generate the diverse response. Extensive experiments on the MultiWOZ datasets demonstrate that the proposed method achieves significantly better diversity than solid competitors.

Prediction calibration devotes to making the model produce correct prediction probability, corresponding with the model's empirical measurement (i.e., the accuracy on benchmark). It plays a crucial role in the Multiple-Choice based Machine Reading Comprehension (MCRC) task. Once the model gives the wrong candidate answer with a high probability, users or downstream applications will not trust the model easily. However, few works pay attention to the prediction calibration of MCRC models. In this paper, we study the prediction calibration of the MCRC models and introduce the self-supervised target label softening (SS-TLS) training method to develop a well-calibrated MRC model while improving its performance. Specifically, the proposed SS-TLS method softens the target label to train the MCRC model instead of the standard cross-entropy objective. It employs the self-supervised confidence signal to monitor the soften scale adaptively at the instance level. Experimental results on several multiple-Choice style MRC datasets illustrate that the proposed method can improve both model prediction calibration and performance.

Question answering systems usually use keyword searches to retrieve potential passages related to a question, and then extract the answer from passages with the machine reading comprehension methods. However, many questions tend to be unanswerable in the real world. In this case, it is significant and challenging how the model determines when no answer is supported by the passage and abstains from answering. Most of the existing systems design a simple classifier to determine answerability implicitly without explicitly modeling mutual interaction and relation between the question and passage, leading to the poor performance for determining the unanswerable questions. To tackle this problem, we propose a Multi-Step Co-Interactive Relation Network (MCR-Net) to explicitly model the mutual interaction and locate key clues from coarse to fine by introducing a co-interactive relation module. The co-interactive relation module contains a stack of interaction and fusion blocks to continuously integrate and fuse history-guided and current-query-guided clues in an explicit way. Experiments on the SQuAD 2.0 and DuReader datasets show that our model achieves a remarkable improvement, outperforming the BERT-style baselines in literature. Visualization analysis also verifies the importance of the mutual interaction between the question and passage.

It is prevalent to utilize external knowledge to help machine answer questions that need background commonsense, which faces a problem that unlimited knowledge will transmit noisy and misleading information. Towards the issue of introducing related knowledge, we propose a semantic-driven knowledge-aware QA framework, which controls the knowledge injection in a coarse-to-careful fashion. We devise a tailoring strategy to filter extracted knowledge under monitoring of the coarse semantic of question on the knowledge extraction stage. And we develop a semantic-aware knowledge fetching module that engages structural knowledge information and fuses proper knowledge according to the careful semantic of questions in a hierarchical way. Experiments demonstrate that the proposed approach promotes the performance on the CommonsenseQA dataset comparing with strong baselines.

Keeping the conversation consistent and avoiding its repetition are two key factors to construct an intelligent multi-turn knowledge-grounded dialogue system. Although some works tend to combine history with external knowledge such as personal background information to boost dialogue quality, they are prone to ignore the fact that incorporating the same knowledge multiple times into the conversation leads to repetition. The main reason is the lack of effective control over the use of knowledge on the conversation level. So we design a history-adaption knowledge incorporation mechanism to build an effective multi-turn dialogue model. Our proposed model addresses repetition by recurrently updating the knowledge from the conversation level and progressively incorporating it into the history step-by-step. And the knowledge-grounded history representation also enhances the conversation consistency. Experimental results show that our proposed model significantly outperforms several retrieval-based models on some benchmark datasets. The human evaluation demonstrates that our model can maintain conversation consistent and reduce conversation repetition.

We propose a novel Bi-directional Cognitive Knowledge Framework (BCKF) for reading comprehension from the perspective of complementary learning systems theory. It aims to simulate two ways of thinking in the brain to answer questions, including reverse thinking and inertial thinking. To validate the effectiveness of our framework, we design a corresponding Bi-directional Cognitive Thinking Network (BCTN) to encode the passage and generate a question (answer) given an answer (question) and decouple the bi-directional knowledge. The model has the ability to reverse reasoning questions which can assist inertial thinking to generate more accurate answers. Competitive improvement is observed in DuReader dataset, confirming our hypothesis that bi-directional knowledge helps the QA task. The novel framework shows an interesting perspective on machine reading comprehension and cognitive science.