Selected work

Representative Papers

Assessing Reliability of BERT-Based Models on Question Answering Tasks

Aug 11, 2026

This study addresses the lack of systematic evaluation of reliability in question-answering models by proposing a novel framework that integrates Monte Carlo Dropout (MCD) with input semantic paraphrasing perturbations, requiring no modification to the standard inference pipeline. The authors conduct a comprehensive analysis of BERT, RoBERTa, ALBERT, and DistilBERT on the SQuAD and QuAC datasets. Experimental results demonstrate that MCD effectively captures prediction stability, with RoBERTa exhibiting the highest reliability, while ALBERT and DistilBERT show notably lower stability. This work not only validates MCD as a robust metric for model reliability but also establishes a new paradigm for evaluating the robustness of question-answering systems.

0 citationsRead paper

Improving Reasoning in Vision-Language Models via Perception Verified Self-Training

Jun 20, 2026

This work addresses the susceptibility of current vision-language models to visual hallucinations and linguistic shortcuts, which often result in reasoning without reliable perceptual grounding. To mitigate these issues, the authors propose a perception-verification-based self-training framework that decouples perception from reasoning. The approach introduces an unsupervised caption evaluation mechanism (PerceptEval) and a “caption–reasoning–conclusion” template to triage training data according to perceptual quality and answer correctness. A two-stage curriculum learning strategy is designed to progressively refine model behavior. By integrating chain-of-thought reasoning, vision–text alignment assessment, and conditional generation—without requiring additional annotations—the method significantly enhances visual grounding and complex reasoning accuracy, effectively reducing both hallucination and shortcut reliance.

0 citationsRead paper

See, Think, Learn: A Self-Taught Multimodal Reasoner

Dec 02, 2025

Current vision-language models (VLMs) suffer from two interrelated bottlenecks in multimodal reasoning: inaccurate perception and brittle inference. Mainstream enhancement approaches rely on costly human annotations, proprietary models, or perception-agnostic self-training, limiting generalizability and practicality. To address this, we propose See-Think-Learn, a novel self-training framework that enables *co-evolution* of perception and reasoning for the first time. It leverages visual attribute extraction to guide structured chain-of-thought generation, integrates negative rationale mining, and employs template-based reasoning distillation—all without any human supervision. Experiments across diverse multimodal reasoning benchmarks demonstrate substantial improvements over strong baselines, achieving superior discriminability, robustness, and interpretability. Our approach establishes a new paradigm for low-cost, scalable multimodal reasoning.

0 citationsRead paper

D-HUMOR: Dark Humor Understanding via Multimodal Open-ended Reasoning

Sep 08, 2025

Detecting black humor in internet memes is challenging due to its reliance on implicit, sensitive, and highly culture-dependent multimodal cues. To address this, we introduce the first large-scale Chinese meme dataset for black humor analysis (4,379 samples), supporting three tasks: black humor detection, target category identification, and intensity grading. Methodologically, we propose a Tri-stream Cross-Reasoning Network that jointly fuses OCR-extracted text, ViT-derived visual features, and structured reasoning sequences generated by a large vision-language model. We further innovate with a role-reversal self-cycling mechanism to better model cultural context and ironic logic. Experiments demonstrate significant improvements over strong baselines across all three tasks. Both the dataset and source code are publicly released to advance research in content safety and multimodal humor understanding.

0 citationsRead paper

Dac-Fake: A Divide and Conquer Framework for Detecting Fake News on Social Media

Aug 22, 2025

To address the rapid propagation of fake news on social media and the latency inherent in traditional fact-checking, this paper proposes an automated detection framework based on a “divide-and-conquer” strategy. The method innovatively fuses over 80 linguistic features with CBOW and Skip-gram word embeddings to jointly model news content and contextual information; ten-fold cross-validation is employed to enhance generalizability and enable early detection. Evaluated on three benchmark datasets—Kaggle, McIntire+PolitiFact, and Reuters—the framework achieves accuracies of 97.88%, 96.05%, and 97.32%, respectively, outperforming state-of-the-art approaches. Key contributions include: (1) a highly efficient and scalable multi-feature fusion architecture; (2) empirical validation of cross-dataset robustness; and (3) a practical, deployable technical pathway for real-time misinformation mitigation.

0 citationsRead paper

Recent publications

Latest Papers

Assessing Reliability of BERT-Based Models on Question Answering Tasks

Aug 11, 2026

This study addresses the lack of systematic evaluation of reliability in question-answering models by proposing a novel framework that integrates Monte Carlo Dropout (MCD) with input semantic paraphrasing perturbations, requiring no modification to the standard inference pipeline. The authors conduct a comprehensive analysis of BERT, RoBERTa, ALBERT, and DistilBERT on the SQuAD and QuAC datasets. Experimental results demonstrate that MCD effectively captures prediction stability, with RoBERTa exhibiting the highest reliability, while ALBERT and DistilBERT show notably lower stability. This work not only validates MCD as a robust metric for model reliability but also establishes a new paradigm for evaluating the robustness of question-answering systems.

0 citationsRead paper

Improving Reasoning in Vision-Language Models via Perception Verified Self-Training

Jun 20, 2026

This work addresses the susceptibility of current vision-language models to visual hallucinations and linguistic shortcuts, which often result in reasoning without reliable perceptual grounding. To mitigate these issues, the authors propose a perception-verification-based self-training framework that decouples perception from reasoning. The approach introduces an unsupervised caption evaluation mechanism (PerceptEval) and a “caption–reasoning–conclusion” template to triage training data according to perceptual quality and answer correctness. A two-stage curriculum learning strategy is designed to progressively refine model behavior. By integrating chain-of-thought reasoning, vision–text alignment assessment, and conditional generation—without requiring additional annotations—the method significantly enhances visual grounding and complex reasoning accuracy, effectively reducing both hallucination and shortcut reliance.

0 citationsRead paper

See, Think, Learn: A Self-Taught Multimodal Reasoner

Dec 02, 2025

Current vision-language models (VLMs) suffer from two interrelated bottlenecks in multimodal reasoning: inaccurate perception and brittle inference. Mainstream enhancement approaches rely on costly human annotations, proprietary models, or perception-agnostic self-training, limiting generalizability and practicality. To address this, we propose See-Think-Learn, a novel self-training framework that enables *co-evolution* of perception and reasoning for the first time. It leverages visual attribute extraction to guide structured chain-of-thought generation, integrates negative rationale mining, and employs template-based reasoning distillation—all without any human supervision. Experiments across diverse multimodal reasoning benchmarks demonstrate substantial improvements over strong baselines, achieving superior discriminability, robustness, and interpretability. Our approach establishes a new paradigm for low-cost, scalable multimodal reasoning.

0 citationsRead paper

D-HUMOR: Dark Humor Understanding via Multimodal Open-ended Reasoning

Sep 08, 2025

Detecting black humor in internet memes is challenging due to its reliance on implicit, sensitive, and highly culture-dependent multimodal cues. To address this, we introduce the first large-scale Chinese meme dataset for black humor analysis (4,379 samples), supporting three tasks: black humor detection, target category identification, and intensity grading. Methodologically, we propose a Tri-stream Cross-Reasoning Network that jointly fuses OCR-extracted text, ViT-derived visual features, and structured reasoning sequences generated by a large vision-language model. We further innovate with a role-reversal self-cycling mechanism to better model cultural context and ironic logic. Experiments demonstrate significant improvements over strong baselines across all three tasks. Both the dataset and source code are publicly released to advance research in content safety and multimodal humor understanding.

0 citationsRead paper

Dac-Fake: A Divide and Conquer Framework for Detecting Fake News on Social Media

Aug 22, 2025

To address the rapid propagation of fake news on social media and the latency inherent in traditional fact-checking, this paper proposes an automated detection framework based on a “divide-and-conquer” strategy. The method innovatively fuses over 80 linguistic features with CBOW and Skip-gram word embeddings to jointly model news content and contextual information; ten-fold cross-validation is employed to enhance generalizability and enable early detection. Evaluated on three benchmark datasets—Kaggle, McIntire+PolitiFact, and Reuters—the framework achieves accuracies of 97.88%, 96.05%, and 97.32%, respectively, outperforming state-of-the-art approaches. Key contributions include: (1) a highly efficient and scalable multi-feature fusion architecture; (2) empirical validation of cross-dataset robustness; and (3) a practical, deployable technical pathway for real-time misinformation mitigation.

0 citationsRead paper