Xinzhi Wang

Weiwen Liu, Xu Huang, Xingshan Zeng et al.

Function calling significantly extends the application boundary of large language models, where high-quality and diverse training data is critical for unlocking this capability. However, real function-calling data is quite challenging to collect and annotate, while synthetic data generated by existing pipelines tends to lack coverage and accuracy. In this paper, we present ToolACE, an automatic agentic pipeline designed to generate accurate, complex, and diverse tool-learning data. ToolACE leverages a novel self-evolution synthesis process to curate a comprehensive API pool of 26,507 diverse APIs. Dialogs are further generated through the interplay among multiple agents, guided by a formalized thinking process. To ensure data accuracy, we implement a dual-layer verification system combining rule-based and model-based checks. We demonstrate that models trained on our synthesized data, even with only 8B parameters, achieve state-of-the-art performance on the Berkeley Function-Calling Leaderboard, rivaling the latest GPT-4 models. Our model and a subset of the data are publicly available at https://huggingface.co/Team-ACE.

Krishna Sri Ipsit Mantri, Xinzhi Wang, Carola-Bibiane Schönlieb et al.

In this paper, we propose a novel activation function tailored specifically for graph data in Graph Neural Networks (GNNs). Motivated by the need for graph-adaptive and flexible activation functions, we introduce DiGRAF, leveraging Continuous Piecewise-Affine Based (CPAB) transformations, which we augment with an additional GNN to learn a graph-adaptive diffeomorphic activation function in an end-to-end manner. In addition to its graph-adaptivity and flexibility, DiGRAF also possesses properties that are widely recognized as desirable for activation functions, such as differentiability, boundness within the domain, and computational efficiency. We conduct an extensive set of experiments across diverse datasets and tasks, demonstrating a consistent and superior performance of DiGRAF compared to traditional and graph-specific activation functions, highlighting its effectiveness as an activation function for GNNs. Our code is available at https://github.com/ipsitmantri/DiGRAF.

Xinzhi Wang, Peter Baile Chen, Gerardo Vitagliano et al.

Large language models with long context windows can answer complex questions directly from full-length academic, technical, and policy documents, but passing entire documents is often costly, slow, and can degrade answer quality while increasing the risk of unnecessary data leakage. This paper targets the common setting of answering many heterogeneous questions over long document(s), where fixed position heuristics and standard retrieval-augmented generation (RAG) can fail due to document structure variability and weak query-chunk semantic similarity, which often requires task- and domain-specific tuning of embedding retrievers. We propose {Selective Attention-Guided Extraction} (\ourmethod), a training-free, plug-and-play context reduction framework that uses a lightweight local LLM to perform a single prefilling pass and convert language model attention signals into a query-specific relevance heatmap at configurable granularities. \ourmethod\ further introduces \emph{differential attention} strategies to better isolate question-relevant evidence, then selects the top-scoring units under a user-defined token budget and forwards only this reduced context to a downstream LLM for answer generation. \ourmethod\ surpasses traditional reduction techniques across multiple long-document QA benchmarks, notably securing a top-4 rank on QuALITY-hard while constrained to a 10\% context budget. This enables a 90\% reduction in tokens with competitive accuracy, without the need for model fine-tuning or complex calibration.

Zhihong Zhang, Xiaojian Huang, Jin Xu et al.

Multimodal reward models (MRMs) play a crucial role in the training, inference, and evaluation of Large Vision Language Models (LVLMs) by assessing response quality. However, existing benchmarks for evaluating MRMs in the video domain suffer from a limited number and diversity of questions, a lack of comprehensive evaluation dimensions, and inadequate evaluation of diverse types of MRMs. To address these gaps, we introduce VideoRewardBench, the first comprehensive benchmark covering four core aspects of video understanding: perception, knowledge, reasoning, and safety. Through our AI-assisted data pipeline, we curate a high-quality preference dataset of 1,563 annotated samples, including 1,482 unique videos and 1,559 distinct questions--15 times the number found in the most question-rich prior benchmark. Each sample is a triplet consisting of a video-text prompt, a chosen response, and a rejected response. We also conduct a comprehensive evaluation across 28 multimodal reward models spanning three categories: generative, discriminative, and semi-scalar. Results show that even the top-performing model GPT-4o achieves only 57.0% overall accuracy, and the state-of-the-art open-source model Qwen2.5-VL-72B reaches merely 53.3%. Our analysis further reveals three key insights: (i) MRMs trained with reinforcement learning (RL) do not necessarily exhibit stronger cross-modal generalization than those trained without RL; (ii) except for discriminative MRMs, other types of MRMs across varying model capacities can benefit from inference-time scaling; and (iii) variations in input video frame count have different effects on different types of MRMs. We believe VideoRewardBench offers a challenging and valuable benchmark for advancing the evaluation and development of MRMs in the video domain.

Chen Chen, Xinlong Hao, Weiwen Liu et al.

Large Language Models (LLMs) have demonstrated significant potential in decision-making and reasoning, particularly when integrated with various tools to effectively solve complex problems. However, existing benchmarks for evaluating LLMs' tool usage face several limitations: (1) limited evaluation scenarios, often lacking assessments in real multi-turn dialogue contexts; (2) narrow evaluation dimensions, with insufficient detailed assessments of how LLMs use tools; and (3) reliance on LLMs or real API executions for evaluation, which introduces significant overhead. To address these challenges, we introduce ACEBench, a comprehensive benchmark for assessing tool usage in LLMs. ACEBench categorizes data into three primary types based on evaluation methodology: Normal, Special, and Agent. "Normal" evaluates tool usage in basic scenarios; "Special" evaluates tool usage in situations with ambiguous or incomplete instructions; "Agent" evaluates tool usage through multi-agent interactions to simulate real-world, multi-turn dialogues. We conducted extensive experiments using ACEBench, analyzing various LLMs in-depth and providing a more granular examination of error causes across different data types.

Xinzhi Wang, Huao Li, Hui Zhang et al.

Recently, there has been increasing interest in transparency and interpretability in Deep Reinforcement Learning (DRL) systems. Verbal explanations, as the most natural way of communication in our daily life, deserve more attention, since they allow users to gain a better understanding of the system which ultimately could lead to a high level of trust and smooth collaboration. This paper reports a novel work in generating verbal explanations for DRL behaviors agent. A rule-based model is designed to construct explanations using a series of rules which are predefined with prior knowledge. A learning model is then proposed to expand the implicit logic of generating verbal explanation to general situations by employing rule-based explanations as training data. The learning model is shown to have better flexibility and generalizability than the static rule-based model. The performance of both models is evaluated quantitatively through objective metrics. The results show that verbal explanation generated by both models improve subjective satisfaction of users towards the interpretability of DRL systems. Additionally, seven variants of the learning model are designed to illustrate the contribution of input channels, attention mechanism, and proposed encoder in improving the quality of verbal explanation.

Xinzhi Wang, Luyao Kou, Vijayan Sugumaran et al.

Emotion analysis has been attracting researchers' attention. Most previous works in the artificial intelligence field focus on recognizing emotion rather than mining the reason why emotions are not or wrongly recognized. Correlation among emotions contributes to the failure of emotion recognition. In this paper, we try to fill the gap between emotion recognition and emotion correlation mining through natural language text from web news. Correlation among emotions, expressed as the confusion and evolution of emotion, is primarily caused by human emotion cognitive bias. To mine emotion correlation from emotion recognition through text, three kinds of features and two deep neural network models are presented. The emotion confusion law is extracted through orthogonal basis. The emotion evolution law is evaluated from three perspectives, one-step shift, limited-step shifts, and shortest path transfer. The method is validated using three datasets-the titles, the bodies, and the comments of news articles, covering both objective and subjective texts in varying lengths (long and short). The experimental results show that, in subjective comments, emotions are easily mistaken as anger. Comments tend to arouse emotion circulations of love-anger and sadness-anger. In objective news, it is easy to recognize text emotion as love and cause fear-joy circulation. That means, journalists may try to attract attention using fear and joy words but arouse the emotion love instead; After news release, netizens generate emotional comments to express their intense emotions, i.e., anger, sadness, and love. These findings could provide insights for applications regarding affective interaction such as network public sentiment, social media communication, and human-computer interaction.

Xinzhi Wang, Shengcheng Yuan, Hui Zhang et al.

This paper focuses on sentiment mining and sentiment correlation analysis of web events. Although neural network models have contributed a lot to mining text information, little attention is paid to analysis of the inter-sentiment correlations. This paper fills the gap between sentiment calculation and inter-sentiment correlations. In this paper, the social emotion is divided into six categories: love, joy, anger, sadness, fear, and surprise. Two deep neural network models are presented for sentiment calculation. Three datasets - the titles, the bodies, the comments of news articles - are collected, covering both objective and subjective texts in varying lengths (long and short). From each dataset, three kinds of features are extracted: explicit expression, implicit expression, and alphabet characters. The performance of the two models are analyzed, with respect to each of the three kinds of the features. There is controversial phenomenon on the interpretation of anger (fn) and love (gd). In subjective text, other emotions are easily to be considered as anger. By contrast, in objective news bodies and titles, it is easy to regard text as caused love (gd). It means, journalist may want to arouse emotion love by writing news, but cause anger after the news is published. This result reflects the sentiment complexity and unpredictability.