Selected work

Representative Papers

World-To-Image: Grounding Text-to-Image Generation with Agent-Driven World Knowledge

Oct 05, 2025

Text-to-image (T2I) models suffer significant performance degradation when generating novel or out-of-distribution (OOD) entities due to knowledge cutoff and insufficient semantic alignment. To address this, we propose World-To-Image—a framework that employs a web-search agent to dynamically retrieve relevant online images and integrates multimodal prompt optimization for real-time external knowledge injection and prompt enhancement. Our method requires no model fine-tuning and achieves knowledge-guided generation in only 2.7 iterations on average. On the NICE benchmark, it improves semantic accuracy by 8.1% over state-of-the-art approaches; it also achieves superior semantic consistency and visual aesthetic quality under LLMGrader and ImageReward evaluations. The core contribution is a scalable, low-overhead “retrieve–optimize” closed loop—marking the first integration of embodied agents into T2I generation to tackle OOD challenges.

0 citationsRead paper

Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting

Jul 23, 2025

Dynamic 4D Gaussian Splatting (4DGS) suffers from prohibitive storage overhead due to its massive number of Gaussians, severe temporal redundancy, and lack of entropy-aware compression—hindering deployment on resource-constrained edge devices. To address this, we propose the first end-to-end rate-distortion optimized compression framework for 4DGS. Our method employs explicit dynamic Gaussian representations, models motion trajectories via wavelet transforms to enforce temporal smoothness priors, and introduces an entropy-aware coding module that jointly optimizes bitrate and rendering fidelity. The framework enables user-controllable accuracy–bitrate trade-offs. Implemented on Ex4DGS, our approach achieves up to 91× compression while preserving high visual fidelity, significantly improving cross-device real-time rendering efficiency.

0 citationsRead paper

Understanding Differential Transformer Unchains Pretrained Self-Attentions

May 22, 2025

Differential Transformers exhibit strong performance but suffer from opaque mechanisms and reliance on costly from-scratch pretraining, hindering reuse of existing pretrained weights. This work presents the first systematic dissection of differential attention’s effectiveness, identifying three key mechanisms: (i) negative attention enhances representational capacity, (ii) it reduces multi-head redundancy, and (iii) it improves learning dynamics. Building on these insights, we propose DEX—a plug-and-play, lightweight adaptation framework that integrates differential attention into arbitrary pretrained language models without full retraining. DEX achieves efficient synergy via softmax-score reuse and value-matrix differentiation. Theoretical analysis and extensive experiments demonstrate that DEX consistently boosts performance across diverse benchmarks, requiring less than 0.01% task-specific adaptation data and incurring negligible training and inference overhead.

0 citationsRead paper

Parameter-Efficient Checkpoint Merging via Metrics-Weighted Averaging

Apr 23, 2025

To address the low efficiency of merging multiple LoRA checkpoints in parameter-efficient fine-tuning (PEFT), this paper proposes Metric-Weighted Averaging (MWA): the first method to explicitly model training loss or step count as dynamic weighting criteria, incorporating an adjustable penalty factor to optimize weight distribution—requiring only a single hyperparameter and no additional training. Evaluated on mathematical reasoning, preference alignment, and instruction-following tasks, MWA consistently outperforms the best individual checkpoint, with loss-weighted fusion achieving up to a 5% absolute accuracy gain. Computational overhead is negligible. The core innovation lies in directly leveraging performance metrics to drive weight design, breaking from conventional uniform or empirically tuned weighting schemes. Within the PEFT paradigm, MWA enables efficient, lightweight, and plug-and-play model ensembling.

0 citationsRead paper

Online Difficulty Filtering for Reasoning Oriented Reinforcement Learning

Apr 04, 2025

In reasoning-oriented reinforcement learning (RORL), inefficient training arises under sparse rewards due to mismatched problem difficulty. Method: This paper proposes the first online dynamic difficulty selection framework with theoretical guarantees. Leveraging a KL-divergence lower bound, we rigorously prove that batch-wise filtering of intermediate-accuracy problems maximizes policy update efficiency. Our method integrates accuracy-variance–aware problem selection, dynamic batch resampling, GRPO-extended policy gradient optimization, and math-task–adaptive difficulty modeling. Results: Evaluated on five mathematical reasoning benchmarks, our approach achieves +10% on AIME score and +4% average accuracy, while surpassing the best baseline reward using only 60% of the training time—demonstrating substantial improvements in both sample and computational efficiency.

0 citationsRead paper

Recent publications

Latest Papers

World-To-Image: Grounding Text-to-Image Generation with Agent-Driven World Knowledge

Oct 05, 2025

Text-to-image (T2I) models suffer significant performance degradation when generating novel or out-of-distribution (OOD) entities due to knowledge cutoff and insufficient semantic alignment. To address this, we propose World-To-Image—a framework that employs a web-search agent to dynamically retrieve relevant online images and integrates multimodal prompt optimization for real-time external knowledge injection and prompt enhancement. Our method requires no model fine-tuning and achieves knowledge-guided generation in only 2.7 iterations on average. On the NICE benchmark, it improves semantic accuracy by 8.1% over state-of-the-art approaches; it also achieves superior semantic consistency and visual aesthetic quality under LLMGrader and ImageReward evaluations. The core contribution is a scalable, low-overhead “retrieve–optimize” closed loop—marking the first integration of embodied agents into T2I generation to tackle OOD challenges.

0 citationsRead paper

Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting

Jul 23, 2025

Dynamic 4D Gaussian Splatting (4DGS) suffers from prohibitive storage overhead due to its massive number of Gaussians, severe temporal redundancy, and lack of entropy-aware compression—hindering deployment on resource-constrained edge devices. To address this, we propose the first end-to-end rate-distortion optimized compression framework for 4DGS. Our method employs explicit dynamic Gaussian representations, models motion trajectories via wavelet transforms to enforce temporal smoothness priors, and introduces an entropy-aware coding module that jointly optimizes bitrate and rendering fidelity. The framework enables user-controllable accuracy–bitrate trade-offs. Implemented on Ex4DGS, our approach achieves up to 91× compression while preserving high visual fidelity, significantly improving cross-device real-time rendering efficiency.

0 citationsRead paper

Understanding Differential Transformer Unchains Pretrained Self-Attentions

May 22, 2025

Differential Transformers exhibit strong performance but suffer from opaque mechanisms and reliance on costly from-scratch pretraining, hindering reuse of existing pretrained weights. This work presents the first systematic dissection of differential attention’s effectiveness, identifying three key mechanisms: (i) negative attention enhances representational capacity, (ii) it reduces multi-head redundancy, and (iii) it improves learning dynamics. Building on these insights, we propose DEX—a plug-and-play, lightweight adaptation framework that integrates differential attention into arbitrary pretrained language models without full retraining. DEX achieves efficient synergy via softmax-score reuse and value-matrix differentiation. Theoretical analysis and extensive experiments demonstrate that DEX consistently boosts performance across diverse benchmarks, requiring less than 0.01% task-specific adaptation data and incurring negligible training and inference overhead.

0 citationsRead paper

Parameter-Efficient Checkpoint Merging via Metrics-Weighted Averaging

Apr 23, 2025

To address the low efficiency of merging multiple LoRA checkpoints in parameter-efficient fine-tuning (PEFT), this paper proposes Metric-Weighted Averaging (MWA): the first method to explicitly model training loss or step count as dynamic weighting criteria, incorporating an adjustable penalty factor to optimize weight distribution—requiring only a single hyperparameter and no additional training. Evaluated on mathematical reasoning, preference alignment, and instruction-following tasks, MWA consistently outperforms the best individual checkpoint, with loss-weighted fusion achieving up to a 5% absolute accuracy gain. Computational overhead is negligible. The core innovation lies in directly leveraging performance metrics to drive weight design, breaking from conventional uniform or empirically tuned weighting schemes. Within the PEFT paradigm, MWA enables efficient, lightweight, and plug-and-play model ensembling.

0 citationsRead paper

Online Difficulty Filtering for Reasoning Oriented Reinforcement Learning

Apr 04, 2025

In reasoning-oriented reinforcement learning (RORL), inefficient training arises under sparse rewards due to mismatched problem difficulty. Method: This paper proposes the first online dynamic difficulty selection framework with theoretical guarantees. Leveraging a KL-divergence lower bound, we rigorously prove that batch-wise filtering of intermediate-accuracy problems maximizes policy update efficiency. Our method integrates accuracy-variance–aware problem selection, dynamic batch resampling, GRPO-extended policy gradient optimization, and math-task–adaptive difficulty modeling. Results: Evaluated on five mathematical reasoning benchmarks, our approach achieves +10% on AIME score and +4% average accuracy, while surpassing the best baseline reward using only 60% of the training time—demonstrating substantial improvements in both sample and computational efficiency.

0 citationsRead paper