Haojun Fei

Yehui Yang, Dalu Yang, Wenshuo Zhou et al.

As multimodal AI becomes widely used for credit risk assessment and document review, a domain-specific benchmark is urgently needed that (1) reflects documents and workflows specific to financial credit applications, (2) includes credit-specific understanding and real-world robustness, and (3) preserves privacy compliance without sacrificing practical utility. Here, we introduce FCMBench-V1.0 -- a large-scale financial credit multimodal benchmark for real-world applications, covering 18 core certificate types, with 4,043 privacy-compliant images and 8,446 QA samples. The FCMBench evaluation framework consists of three dimensions: Perception, Reasoning, and Robustness, including 3 foundational perception tasks, 4 credit-specific reasoning tasks that require decision-oriented understanding of visual evidence, and 10 real-world acquisition artifact types for robustness stress testing. To reconcile compliance with realism, we construct all samples via a closed synthesis-capture pipeline: we manually synthesize document templates with virtual content and capture scenario-aware images in-house. This design also mitigates pre-training data leakage by avoiding web-sourced or publicly released images. FCMBench can effectively discriminate performance disparities and robustness across modern vision-language models. Extensive experiments were conducted on 23 state-of-the-art vision-language models (VLMs) from 14 top AI companies and research institutes. Among them, Gemini 3 Pro achieves the best F1(\%) score as a commercial model (64.61), Qwen3-VL-235B achieves the best score as an open-source baseline (57.27), and our financial credit-specific model, Qfin-VL-Instruct, achieves the top overall score (64.92). Robustness evaluations show that even top-performing models suffer noticeable performance drops under acquisition artifacts.

Che Zhang, Ping'an Liu, Zhenyang Xiao et al.

The study of human values is essential in both practical and theoretical domains. With the development of computational linguistics, the creation of large-scale datasets has made it possible to automatically recognize human values accurately. SemEval 2023 Task 4\cite{kiesel:2023} provides a set of arguments and 20 types of human values that are implicitly expressed in each argument. In this paper, we present our team's solution. We use the Roberta\cite{liu_roberta_2019} model to obtain the word vector encoding of the document and propose a multi-head attention mechanism to establish connections between specific labels and semantic components. Furthermore, we use a contrastive learning-enhanced K-nearest neighbor mechanism\cite{su_contrastive_2022} to leverage existing instance information for prediction. Our approach achieved an F1 score of 0.533 on the test set and ranked fourth on the leaderboard.

Jinming Chen, Jingyi Fang, Yuanzhong Zheng et al.

Currently, end-to-end (E2E) speech recognition methods have achieved promising performance. However, auto speech recognition (ASR) models still face challenges in recognizing multi-accent speech accurately. We propose a layer-adapted fusion (LAF) model, called Qifusion-Net, which does not require any prior knowledge about the target accent. Based on dynamic chunk strategy, our approach enables streaming decoding and can extract frame-level acoustic feature, facilitating fine-grained information fusion. Experiment results demonstrate that our proposed methods outperform the baseline with relative reductions of 22.1$\%$ and 17.2$\%$ in character error rate (CER) across multi accent test datasets on KeSpeech and MagicData-RMAC.

Ziming Dai, Dabiao Ma, Jinle Tong et al.

Although the Gradient Boosted Decision Trees (GBDTs) dominate industrial tabular applications, upgrading legacy models in high-concurrency production environments still faces prohibitive retraining costs and systemic risks. To address this problem, we present NSR-Boost, a neuro-symbolic residual boosting framework designed specifically for industrial scenarios. Its core advantage lies in being "non-intrusive". It treats the legacy model as a frozen model and performs targeted repairs on "hard regions" where predictions fail. The framework comprises three key stages: First, finding hard regions through residuals, then generating interpretable experts by generating symbolic code structures using Large Language Model (LLM) and fine-tuning parameters using Bayesian optimization, and finally dynamically integrating experts with legacy model output through a lightweight aggregator. Experimental results demonstrate that the framework not only significantly outperforms state-of-the-art (SOTA) baselines across six public datasets and one private dataset. More importantly, we report the successful deployment of NSR-Boost within the core financial risk control system of Qfin Holdings, where empirical results on real-world online traffic exhibit superior performance improvements and a significant reduction in the bad rate. In conclusion, it effectively captures long-tail risks missed by traditional models and offers a safe, low-cost evolutionary paradigm for industry.

Dabiao Ma, Zhiba Su, Jian Yang et al.

This paper introduces a novel approach to securing machine learning model deployments against potential distribution shifts in practical applications, the Total Variation Out-of-Distribution (TV-OOD) detection method. Existing methods have produced satisfactory results, but TV-OOD improves upon these by leveraging the Total Variation Network Estimator to calculate each input's contribution to the overall total variation. By defining this as the total variation score, TV-OOD discriminates between in- and out-of-distribution data. The method's efficacy was tested across a range of models and datasets, consistently yielding results in image classification tasks that were either comparable or superior to those achieved by leading-edge out-of-distribution detection techniques across all evaluation metrics.

Xudong Yan, Songhe Feng, Yang Zhang et al.

Compositional zero-shot learning (CZSL) aims to recognize novel compositions of attributes and objects learned from seen compositions. Previous works disentangle attributes and objects by extracting shared and exclusive parts between the image pair sharing the same attribute (object), as well as aligning them with pretrained word embeddings to improve unseen attribute-object recognition. Despite the significant achievements of existing efforts, they are hampered by three limitations: (1) The efficacy of disentanglement is compromised due to the influence of the background and the intricate entanglement of attributes with objects in the same parts. (2) Existing word embeddings fail to capture complex multimodal semantic information. (3) Overconfidence exhibited by existing models in seen compositions hinders their generalization to novel compositions. Being aware of these, we propose a novel framework named multimodal large language model (MLLM) embeddings and attribute smoothing guided disentanglement for CZSL. First, we leverage feature adaptive aggregation modules to mitigate the impact of background, and utilize learnable condition masks to capture multi-granularity features for disentanglement. Moreover, the last hidden states of MLLM are employed as word embeddings for their superior representation capabilities. Furthermore, we propose attribute smoothing with auxiliary attributes generated by the large language model (LLM) for seen compositions to address the overconfidence challenge. Extensive experiments demonstrate that our method achieves state-of-the-art performance on three challenging datasets. The source code will be available at https://github.com/xud-yan/Trident .

Yuqi Li, Yuanzhong Zheng, Zhongtian Guo et al.

This paper presents SpecWav-Attack, an adversarial model for detecting speakers in anonymized speech. It leverages Wav2Vec2 for feature extraction and incorporates spectrogram resizing and incremental training for improved performance. Evaluated on librispeech-dev and librispeech-test, SpecWav-Attack outperforms conventional attacks, revealing vulnerabilities in anonymized speech systems and emphasizing the need for stronger defenses, benchmarked against the ICASSP 2025 Attacker Challenge.

Jinming Chen, Jingyi Fang, Yuanzhong Zheng et al.

Emotion recognition plays a pivotal role in intelligent human-machine interaction systems. Multimodal approaches benefit from the fusion of diverse modalities, thereby improving the recognition accuracy. However, the lack of high-quality multimodal data and the challenge of achieving optimal alignment between different modalities significantly limit the potential for improvement in multimodal approaches. In this paper, the proposed Qieemo framework effectively utilizes the pretrained automatic speech recognition (ASR) model backbone which contains naturally frame aligned textual and emotional features, to achieve precise emotion classification solely based on the audio modality. Furthermore, we design the multimodal fusion (MMF) module and cross-modal attention (CMA) module in order to fuse the phonetic posteriorgram (PPG) and emotional features extracted by the ASR encoder for improving recognition accuracy. The experimental results on the IEMOCAP dataset demonstrate that Qieemo outperforms the benchmark unimodal, multimodal, and self-supervised models with absolute improvements of 3.0%, 1.2%, and 1.9% respectively.

Dabiao Ma, Ziming Dai, Zhimin Xin et al.

In the field of large models (LMs) for natural language processing (NLP) and computer vision (CV), Parameter-Efficient Fine-Tuning (PEFT) has emerged as a resource-efficient method that modifies a limited number of parameters while keeping the pretrained weights fixed. This paper investigates the traditional PEFT approach, which applies modifications to all position indices, and questions its necessity. We introduce a new paradigm called Token-Selective PEFT (TS-PEFT), in which a function S selectively applies PEFT modifications to a subset of position indices, potentially enhancing performance on downstream tasks. Our experimental results reveal that the indiscriminate application of PEFT to all indices is not only superfluous, but may also be counterproductive. This study offers a fresh perspective on PEFT, advocating for a more targeted approach to modifications and providing a framework for future research to optimize the fine-tuning process for large models.

Xiaochuan Zhang, Mengran Li, Ye Wang et al.

Attribute graphs are ubiquitous in multimedia applications, and graph representation learning (GRL) has been successful in analyzing attribute graph data. However, incomplete graph data and missing node attributes can have a negative impact on media knowledge discovery. Existing methods for handling attribute missing graph have limited assumptions or fail to capture complex attribute-graph dependencies. To address these challenges, we propose Attribute missing Graph Contrastive Learning (AmGCL), a framework for handling missing node attributes in attribute graph data. AmGCL leverages Dirichlet energy minimization-based feature precoding to encode in missing attributes and a self-supervised Graph Augmentation Contrastive Learning Structure (GACLS) to learn latent variables from the encoded-in data. Specifically, AmGCL utilizies feature reconstruction based on structure-attribute energy minimization while maximizes the lower bound of evidence for latent representation mutual information. Our experimental results on multiple real-world datasets demonstrate that AmGCL outperforms state-of-the-art methods in both feature imputation and node classification tasks, indicating the effectiveness of our proposed method in real-world attribute graph analysis tasks.