Recent advances in enabling Large Language Models (LLMs) to use external tools have significantly extended their functional capabilities beyond internal knowledge. However, most existing approaches rely on retrieval-based mechanisms to select suitable tools from massive repositories, which often struggle to align user queries with tool documentation and exhibit limited generalization. To address these issues, we propose ToolBox-RL. The code is available at https://github.com/S-cavy/ToolBox-RL., a novel reinforcement learning framework that unifies query rewriting, intent understanding, and large-scale tool retrieval into an end-to-end optimization process. ToolBox-RL introduces a query rewriting stage to better capture user intent and ensure semantic alignment with tool descriptions, while reinforcement learning encourages autonomous discovery of generalized tool-use strategies through combined cold-start and policy optimization training. Our experiments demonstrate that ToolBox-RL not only achieves the best tool call accuracy on both white-box and black-box tools but also exhibits strong generalization capabilities on out-of-domain dataset. Ablation studies show that ToolBox-RL can adapt to different retrieval methods and maintains good performance even when incorporating a large number of distractor tools.

Query-service relevance prediction in e-commerce search systems faces strict latency requirements that prevent the direct application of Large Language Models (LLMs). To bridge this gap, we propose a two-stage reasoning distillation framework to transfer reasoning capabilities from a powerful teacher LLM to a lightweight, deployment-friendly student model. In the first stage, we address the limitations of general-purpose LLMs by constructing a domain-adapted teacher model. This is achieved through a three-step process: domain-adaptive pre-training to inject platform knowledge, supervised fine-tuning to elicit reasoning skills, and preference optimization with a multi-dimensional reward model to ensure the generation of reliable and preference-aligned reasoning paths. This teacher can then automatically annotate massive query-service pairs from search logs with both relevance labels and reasoning chains. In the second stage, to address the challenges of architectural heterogeneity in standard distillation, we introduce Contrastive Reasoning Self-Distillation (CRSD). By modeling the behavior of the same student model under ``standard'' and ``reasoning-augmented'' inputs as a teacher-student relationship, CRSD enables the lightweight model to internalize the teacher's complex decision-making mechanisms without needing the explicit reasoning path at inference. Offline evaluations and online A/B testing in the Meituan search advertising system demonstrate that our framework achieves significant improvements across multiple metrics, validating its effectiveness and practical value.

Constrained decoding approaches aim to control the meaning or style of text generated by a Pre-trained Language Model (PLM) for various task-specific objectives at inference time. However, these methods often guide plausible continuations by greedily and explicitly selecting targets, which, while fulfilling the task requirements, may overlook the natural patterns of human language generation. In this work, we propose a novel decoding framework,Decider, which enables us to program high-level rules on how we might effectively complete tasks to control a PLM. Differing from previous works, our framework transforms the encouragement of concrete target words into the encouragement of all words that satisfy the high-level rules. Specifically,Decideris a dual system in which a PLM is equipped and controlled by a First-Order Logic (FOL) reasoner to express and evaluate the rules, along with a decision function that merges the outputs from both systems to guide the generation. Experiments on CommonGen and PersonaChat demonstrate thatDecidercan effectively follow given rules to guide a PLM in achieving generation tasks in a more human-like manner.

The trade-off between privacy and accuracy presents a challenge for current federated learning (FL) frameworks, hindering their progress from theory to application. The main issues with existing FL frameworks stem from a lack of interpretability and targeted privacy protections. To cope with these, we proposed Disentangled Federated Learning for Privacy (DFLP) which employes disentanglement, one of interpretability techniques, in private FL frameworks. Since sensitive properties are client-specific in nature, our main idea is to turn this feature into a tool that strikes the balance between data privacy and FL model performance, enabling the sensitive attributes to be private. DFLP disentangles the client-specific and class-invariant attributes to mask the sensitive attributes precisely. To our knowledge, this is the first work that successfully integrates disentanglement and the nature of sensitive attributes to achieve privacy protection while ensuring high FL model performance. Extensive experiments validate that disentanglement is an effective method for accuracy-aware privacy protection in FL frameworks.

Dialogue structure discovery is essential in dialogue generation. Well-structured topic flow can leverage background information and predict future topics to help generate controllable and explainable responses. However, most previous work focused on dialogue structure learning in task-oriented dialogue other than open-domain dialogue which is more complicated and challenging. In this paper, we present a new framework CTRLStruct for dialogue structure learning to effectively explore topic-level dialogue clusters as well as their transitions with unlabelled information. Precisely, dialogue utterances encoded by bi-directional Transformer are further trained through a special designed contrastive learning task to improve representation. Then we perform clustering to utterance-level representations and form topic-level clusters that can be considered as vertices in dialogue structure graph. The edges in the graph indicating transition probability between vertices are calculated by mimicking expert behavior in datasets. Finally, dialogue structure graph is integrated into dialogue model to perform controlled response generation. Experiments on two popular open-domain dialogue datasets show our model can generate more coherent responses compared to some excellent dialogue models, as well as outperform some typical sentence embedding methods in dialogue utterance representation. Code is available in GitHub1.

Maintaining a consistent persona is essential for building a human-like conversational model. However, the lack of attention to the partner makes the model more egocentric: they tend to show their persona by all means such as twisting the topic stiffly, pulling the conversation to their own interests regardless, and rambling their persona with little curiosity to the partner. In this work, we propose COSPLAY(COncept Set guided PersonaLized dialogue generation Across both partY personas) that considers both parties as a "team": expressing self-persona while keeping curiosity toward the partner, leading responses around mutual personas, and finding the common ground. Specifically, we first represent self-persona, partner persona and mutual dialogue all in the concept sets. Then, we propose the Concept Set framework with a suite of knowledge-enhanced operations to process them such as set algebras, set expansion, and set distance. Based on these operations as medium, we train the model by utilizing 1) concepts of both party personas, 2) concept relationship between them, and 3) their relationship to the future dialogue. Extensive experiments on a large public dataset, Persona-Chat, demonstrate that our model outperforms state-of-the-art baselines for generating less egocentric, more human-like, and higher quality responses in both automatic and human evaluations.

In e-commerce, opinion tags refer to a ranked list of tags provided by the e-commerce platform that reflect characteristics of reviews of an item. To assist consumers to quickly grasp a large number of reviews about an item, opinion tags are increasingly being applied by e-commerce platforms. Current mechanisms for generating opinion tags rely on either manual labelling or heuristic methods, which is time-consuming and ineffective. In this paper, we propose the abstractive opinion tagging task, where systems have to automatically generate a ranked list of opinion tags that are based on, but need not occur in, a given set of user-generated reviews. The abstractive opinion tagging task comes with three main challenges: the noisy nature of reviews; the formal nature of opinion tags vs. the colloquial language usage in reviews; and the need to distinguish between different items with very similar aspects. To address these challenges, we propose an abstractive opinion tagging framework, named AOT-Net, to generate a ranked list of opinion tags given a large number of reviews. First, a sentence-level salience estimation component estimates each review's salience score. Next, a review clustering and ranking component ranks reviews in two steps: first, reviews are grouped into clusters and ranked by cluster size; then, reviews within each cluster are ranked by their distance to the cluster center. Finally, given the ranked reviews, a rank-aware opinion tagging component incorporates an alignment feature and alignment loss to generate a ranked list of opinion tags. To facilitate the study of this task, we create and release a large-scale dataset, called eComTag, crawled from real-world e-commerce websites. Extensive experiments conducted on the eComTag dataset verify the effectiveness of the proposed AOT-Net in terms of various evaluation metrics.

Tips, as a compacted and concise form of reviews, were paid less attention by researchers. In this paper, we investigate the task of tips generation by considering the “persona” information which captures the intrinsic language style of the users or the different characteristics of the product items. In order to exploit the persona information, we propose a framework based on adversarial variational auto-encoders (aVAE) for persona modeling from the historical tips and reviews of users and items. The latent variables from aVAE are regarded as persona embeddings. Besides representing persona using the latent embeddings, we design a persona memory for storing the persona related words for users and items. Pointer Network is used to retrieve persona wordings from the memory when generating tips. Moreover, the persona embeddings are used as latent factors by a rating prediction component to predict the sentiment of a user over an item. Finally, the persona embeddings and the sentiment information are incorporated into a recurrent neural networks based tips generation component. Extensive experimental results are reported and discussed to elaborate the peculiarities of our framework.

Recently, some E-commerce sites launch a new interaction box called Tips on their mobile apps. Users can express their experience and feelings or provide suggestions using short texts typically several words or one sentence. In essence, writing some tips and giving a numerical rating are two facets of a user's product assessment action, expressing the user experience and feelings. Jointly modeling these two facets is helpful for designing a better recommendation system. While some existing models integrate text information such as item specifications or user reviews into user and item latent factors for improving the rating prediction, no existing works consider tips for improving recommendation quality. We propose a deep learning based framework named NRT which can simultaneously predict precise ratings and generate abstractive tips with good linguistic quality simulating user experience and feelings. For abstractive tips generation, gated recurrent neural networks are employed to "translate'' user and item latent representations into a concise sentence. Extensive experiments on benchmark datasets from different domains show that NRT achieves significant improvements over the state-of-the-art methods. Moreover, the generated tips can vividly predict the user experience and feelings.

A recommendation is called explainable if it not only predicts a numerical rating for an item, but also generates explanations for users' preferences. Most existing methods for explainable recommendation apply topic models to analyze user reviews to provide descriptions along with the recommendations they produce. So far, such methods have neglected user opinions and influences from social relations as a source of information for recommendations, even though these are known to improve the rating prediction.

We aim at summarizing answers in community question-answering (CQA). While most previous work focuses on factoid question-answering, we focus on the non-factoid question-answering. Unlike factoid CQA, non-factoid question-answering usually requires passages as answers. The shortness, sparsity and diversity of answers form interesting challenges for summarization. To tackle these challenges, we propose a sparse coding-based summarization strategy that includes three core ingredients: short document expansion, sentence vectorization, and a sparse-coding optimization framework. Specifically, we extend each answer in a question-answering thread to a more comprehensive representation via entity linking and sentence ranking strategies. From answers extended in this manner, each sentence is represented as a feature vector trained from a short text convolutional neural network model. We then use these sentence representations to estimate the saliency of candidate sentences via a sparse-coding framework that jointly considers candidate sentences and Wikipedia sentences as reconstruction items. Given the saliency vectors for all candidate sentences, we extract sentences to generate an answer summary based on a maximal marginal relevance algorithm. Experimental results on a benchmark data collection confirm the effectiveness of our proposed method in answer summarization of non-factoid CQA, and moreover, its significant improvement compared to state-of-the-art baselines in terms of ROUGE metrics.

We consider the problem of learning to summarize images by text and visualize text utilizing images, which we call Mutual-Summarization. We divide the web image-text data space into three subspaces, namely pure image space (PIS), pure text space (PTS) and image-text joint space (ITJS). Naturally, we treat the ITJS as a knowledge base.

Most of the state-of-the-art systems for object localization rely on supervised machine learning techniques, and are thus limited by the lack of labeled training data. In this paper, our motivation is to provide training dataset for object localization effectively and efficiently. We argue that the notes data in Flickr can be exploited as a novel source for object modeling. At first, we apply a text mining method to gather semantically related images for a specific class. Then a handful of images are selected manually as seed images or initial training set. At last, the training set is expanded by an incremental active learning framework. Our approach requires significantly less manual supervision compared to standard methods. The experimental results on the PASCAL VOC 2007 and NUS-WIDE datasets show that the training data acquired by our approach can complement or even substitute conventional training data for object localization.