Selected work

Representative Papers

M$^{3}$T2IBench: A Large-Scale Multi-Category, Multi-Instance, Multi-Relation Text-to-Image Benchmark

Oct 27, 2025

Current text-to-image models exhibit limited capability in aligning generated images with complex textual descriptions involving multiple instances, diverse categories, and intricate semantic relationships; moreover, fine-grained evaluation benchmarks strongly correlated with human judgment remain scarce. To address these challenges, we introduce M3R-Bench—the first large-scale, multi-category, multi-instance, multi-relation image-text alignment benchmark—and propose Revise-Then-Enforce, a training-free post-editing method. We further design AlignScore, an automatic metric that jointly models object detection and semantic parsing to assess both visual entities and their relational structure, achieving strong correlation with human preferences (ρ > 0.85). Experiments reveal significant performance bottlenecks of mainstream open-source diffusion models on M3R-Bench, confirming its rigor; Revise-Then-Enforce consistently improves alignment quality across models—including Stable Diffusion—yielding an average AlignScore gain of 12.7%. This work establishes a new paradigm for evaluating and optimizing image-text alignment.

0 citationsRead paper

JointCQ: Improving Factual Hallucination Detection with Joint Claim and Query Generation

Oct 22, 2025

Large language models (LLMs) frequently generate factually inconsistent outputs—so-called hallucinations. Existing hallucination detection methods suffer from two critical limitations: (1) loss of contextual information during claim extraction, and (2) insufficient specificity in verification query generation. To address these issues, we propose the first end-to-end joint modeling framework that simultaneously optimizes claim extraction and verification query generation. Our approach employs fine-grained context-aware modeling to mitigate information decay and introduces a training strategy based on controllable synthetic data filtering to enhance query discriminability. Evaluated on multiple open-domain question-answering hallucination detection benchmarks—including HOVER and FEVER-Sym—our method achieves significant improvements over state-of-the-art approaches. Downstream retrieval-based verification accuracy increases by 4.2–7.8 percentage points on average, demonstrating enhanced robustness and interpretability in hallucination detection.

0 citationsRead paper

Beyond Isolated Facts: Synthesizing Narrative and Grounded Supervision for VideoQA

Sep 29, 2025

Existing VideoQA models rely on shallow supervision signals from isolated question-answer pairs, limiting their ability to model the narrative logic and causal structure of video events. To address this, we propose a question-driven narrativized supervision paradigm: leveraging Question-Based Paraphrasing (QBP) and Question-Based Captioning (QBC), we reconstruct discrete QA pairs into coherent narrative paragraphs grounded in fine-grained visual evidence. The resulting narratives are trained end-to-end within a unified next-token prediction framework. This approach elevates video understanding supervision from a “collection of facts” to a “structured narrative” for the first time, substantially enhancing models’ capacity to capture deep event semantics. Our method achieves new state-of-the-art results on STAR and NExT-QA: a 3B-parameter model improves accuracy on STAR by 4.9 points to 72.5%, while a 7B model attains 80.8% on NExT-QA. It also demonstrates improved cross-dataset generalization and faster training convergence.

0 citationsRead paper

Identity-Preserving Text-to-Video Generation via Training-Free Prompt, Image, and Guidance Enhancement

Sep 01, 2025

Addressing the challenges of identity preservation, reliance on fine-tuning, and data scarcity in text-to-video generation, this paper proposes a training-free triple-enhancement framework. First, GPT-4o–driven face-aware prompt enhancement bridges the semantic gap between textual descriptions and visual content. Second, a prompt-aware reference image optimization mechanism improves input consistency. Third, a unified gradient-guided strategy jointly optimizes identity fidelity and spatiotemporal coherence during diffusion model sampling—enabling inference-time refinement without architectural modification. The method requires no model training or fine-tuning. Extensive evaluation on a thousand-video benchmark demonstrates significant improvements in character identity consistency and video quality, outperforming state-of-the-art approaches in both automated metrics and human assessment. It ranked first in the ACM Multimedia 2025 Challenge, validating its strong generalizability and practical applicability.

0 citationsRead paper

AR-VRM: Imitating Human Motions for Visual Robot Manipulation with Analogical Reasoning

Aug 11, 2025

Visual robotic manipulation (VRM) suffers from scarce robot interaction data and high costs of multimodal annotation. Existing vision-language pretraining approaches either rely on non-task-specific web data or employ implicit modeling (e.g., frame prediction), resulting in poor generalization under few-shot settings. To address this, we propose an analogy-based cross-modal action transfer framework. Our method explicitly extracts action knowledge from human hand keypoints—establishing a structured analogical mapping between human motion and robot actuator dynamics for the first time. It integrates keypoint-driven vision-language pretraining, human action video retrieval, historical observation alignment, and an analogy reasoning network. Evaluated on the CALVIN benchmark and real-robot experiments, our approach significantly outperforms state-of-the-art methods in few-shot scenarios, demonstrating that human motion priors effectively enhance robotic generalization across tasks and environments.

0 citationsRead paper

Recent publications

Latest Papers

M$^{3}$T2IBench: A Large-Scale Multi-Category, Multi-Instance, Multi-Relation Text-to-Image Benchmark

Oct 27, 2025

Current text-to-image models exhibit limited capability in aligning generated images with complex textual descriptions involving multiple instances, diverse categories, and intricate semantic relationships; moreover, fine-grained evaluation benchmarks strongly correlated with human judgment remain scarce. To address these challenges, we introduce M3R-Bench—the first large-scale, multi-category, multi-instance, multi-relation image-text alignment benchmark—and propose Revise-Then-Enforce, a training-free post-editing method. We further design AlignScore, an automatic metric that jointly models object detection and semantic parsing to assess both visual entities and their relational structure, achieving strong correlation with human preferences (ρ > 0.85). Experiments reveal significant performance bottlenecks of mainstream open-source diffusion models on M3R-Bench, confirming its rigor; Revise-Then-Enforce consistently improves alignment quality across models—including Stable Diffusion—yielding an average AlignScore gain of 12.7%. This work establishes a new paradigm for evaluating and optimizing image-text alignment.

0 citationsRead paper

JointCQ: Improving Factual Hallucination Detection with Joint Claim and Query Generation

Oct 22, 2025

Large language models (LLMs) frequently generate factually inconsistent outputs—so-called hallucinations. Existing hallucination detection methods suffer from two critical limitations: (1) loss of contextual information during claim extraction, and (2) insufficient specificity in verification query generation. To address these issues, we propose the first end-to-end joint modeling framework that simultaneously optimizes claim extraction and verification query generation. Our approach employs fine-grained context-aware modeling to mitigate information decay and introduces a training strategy based on controllable synthetic data filtering to enhance query discriminability. Evaluated on multiple open-domain question-answering hallucination detection benchmarks—including HOVER and FEVER-Sym—our method achieves significant improvements over state-of-the-art approaches. Downstream retrieval-based verification accuracy increases by 4.2–7.8 percentage points on average, demonstrating enhanced robustness and interpretability in hallucination detection.

0 citationsRead paper

Beyond Isolated Facts: Synthesizing Narrative and Grounded Supervision for VideoQA

Sep 29, 2025

Existing VideoQA models rely on shallow supervision signals from isolated question-answer pairs, limiting their ability to model the narrative logic and causal structure of video events. To address this, we propose a question-driven narrativized supervision paradigm: leveraging Question-Based Paraphrasing (QBP) and Question-Based Captioning (QBC), we reconstruct discrete QA pairs into coherent narrative paragraphs grounded in fine-grained visual evidence. The resulting narratives are trained end-to-end within a unified next-token prediction framework. This approach elevates video understanding supervision from a “collection of facts” to a “structured narrative” for the first time, substantially enhancing models’ capacity to capture deep event semantics. Our method achieves new state-of-the-art results on STAR and NExT-QA: a 3B-parameter model improves accuracy on STAR by 4.9 points to 72.5%, while a 7B model attains 80.8% on NExT-QA. It also demonstrates improved cross-dataset generalization and faster training convergence.

0 citationsRead paper

Identity-Preserving Text-to-Video Generation via Training-Free Prompt, Image, and Guidance Enhancement

Sep 01, 2025

Addressing the challenges of identity preservation, reliance on fine-tuning, and data scarcity in text-to-video generation, this paper proposes a training-free triple-enhancement framework. First, GPT-4o–driven face-aware prompt enhancement bridges the semantic gap between textual descriptions and visual content. Second, a prompt-aware reference image optimization mechanism improves input consistency. Third, a unified gradient-guided strategy jointly optimizes identity fidelity and spatiotemporal coherence during diffusion model sampling—enabling inference-time refinement without architectural modification. The method requires no model training or fine-tuning. Extensive evaluation on a thousand-video benchmark demonstrates significant improvements in character identity consistency and video quality, outperforming state-of-the-art approaches in both automated metrics and human assessment. It ranked first in the ACM Multimedia 2025 Challenge, validating its strong generalizability and practical applicability.

0 citationsRead paper

AR-VRM: Imitating Human Motions for Visual Robot Manipulation with Analogical Reasoning

Aug 11, 2025

Visual robotic manipulation (VRM) suffers from scarce robot interaction data and high costs of multimodal annotation. Existing vision-language pretraining approaches either rely on non-task-specific web data or employ implicit modeling (e.g., frame prediction), resulting in poor generalization under few-shot settings. To address this, we propose an analogy-based cross-modal action transfer framework. Our method explicitly extracts action knowledge from human hand keypoints—establishing a structured analogical mapping between human motion and robot actuator dynamics for the first time. It integrates keypoint-driven vision-language pretraining, human action video retrieval, historical observation alignment, and an analogy reasoning network. Evaluated on the CALVIN benchmark and real-robot experiments, our approach significantly outperforms state-of-the-art methods in few-shot scenarios, demonstrating that human motion priors effectively enhance robotic generalization across tasks and environments.

0 citationsRead paper