Selected work

Representative Papers

One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs

Jun 12, 2026

This work addresses the performance degradation in existing large vision-language models caused by static visual token pruning, which permanently discards information inaccessible to subsequent layers. To overcome this limitation, the authors propose an adaptive, hierarchical visual token selection mechanism that dynamically determines, via lightweight selectors, which tokens to process or skip at each layer, followed by cross-layer token fusion—enabling full-model adaptive compression without retraining. Key innovations include layer-wise differentiated token retention and reuse, a low-rank attention approximation guided by importance consistency constraints, and a multi-stream token routing and fusion strategy. Experiments on LLaVA-1.5, LLaVA-NeXT, and Qwen2.5-VL demonstrate that retaining only 11% of the original visual tokens achieves 96.7% of the baseline model’s accuracy.

0 citationsRead paper

COREY: A Prototype Study of Entropy-Guided Operator Fusion with Hadamard Reparameterization for Selective State Space Models

Apr 12, 2026

This work addresses the memory bandwidth bottleneck of selective state space models during deployment, which arises from fragmented operator decomposition and redundant intermediate tensor generation in state updates. The authors propose a memory-aware operator fusion strategy that leverages activation entropy as a runtime scheduling metric to dynamically optimize fusion boundaries and tiling sizes. Additionally, they introduce a reparameterized linear projection based on normalized Hadamard transforms to mitigate the heavy-tailed distribution of activations, thereby reducing peak coordinate concentration while preserving functional equivalence. Experimental results demonstrate that the proposed approach significantly reduces DRAM traffic, lowers proxy latency, and improves throughput, outperforming both unfused and fixed-depth baselines.

0 citationsRead paper

Diffusion Epistemic Uncertainty with Asymmetric Learning for Diffusion-Generated Image Detection

Jan 21, 2026

Existing methods for detecting images generated by diffusion models fail to distinguish between aleatoric and epistemic uncertainty, limiting their discriminative performance and generalization capability. This work addresses this limitation by explicitly leveraging epistemic uncertainty for detection—a first in the field—and proposes a Laplace approximation–based approach to estimate epistemic uncertainty in diffusion models. Furthermore, an asymmetric loss function with a large margin is introduced to emphasize the most discriminative components of reconstruction error. The proposed method achieves state-of-the-art performance across multiple large-scale benchmarks and demonstrates significantly improved generalization in detecting images synthesized by previously unseen diffusion models.

0 citationsRead paper

MHA2MLA-VLM: Enabling DeepSeek's Economical Multi-Head Latent Attention across Vision-Language Models

Jan 16, 2026

This work addresses the significant memory and computational bottlenecks caused by the rapid growth of KV cache during inference in vision-language models. To mitigate this, the authors propose a parameter-efficient, multimodal-aware framework that transforms off-the-shelf models into a multi-head latent attention architecture to compress the KV cache and accelerate inference. Key innovations include a modality-adaptive partial RoPE mechanism, modality-disentangled low-rank approximation, and an efficient fine-tuning strategy that minimizes output activation error. With minimal supervised data, the method effectively recovers the original model performance across three mainstream vision-language models while substantially reducing KV cache memory footprint and maintaining natural compatibility with KV quantization techniques.

0 citationsRead paper

The Role of Entropy in Visual Grounding: Analysis and Optimization

Dec 07, 2025

The role of entropy in multimodal large language models (MLLMs) for visual grounding remains poorly understood, and existing entropy regulation strategies lack interpretability and task adaptability. Method: This paper introduces the Entropy-Controlled Visual Grounding Policy Optimization (ECVGPO) algorithm—a reinforcement learning–based approach featuring a dynamic entropy regularization mechanism that adaptively balances exploration and exploitation. Unlike conventional fixed-entropy or black-box entropy control methods, ECVGPO provides explicit, interpretable, and task-aware entropy modulation. Contribution/Results: Extensive experiments demonstrate that ECVGPO significantly improves both performance and training stability across multiple visual grounding benchmarks and mainstream MLLMs. It also exhibits superior generalization capability, establishing a novel paradigm for perception-decision co-optimization in MLLMs.

0 citationsRead paper

Recent publications

Latest Papers

One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs

Jun 12, 2026

This work addresses the performance degradation in existing large vision-language models caused by static visual token pruning, which permanently discards information inaccessible to subsequent layers. To overcome this limitation, the authors propose an adaptive, hierarchical visual token selection mechanism that dynamically determines, via lightweight selectors, which tokens to process or skip at each layer, followed by cross-layer token fusion—enabling full-model adaptive compression without retraining. Key innovations include layer-wise differentiated token retention and reuse, a low-rank attention approximation guided by importance consistency constraints, and a multi-stream token routing and fusion strategy. Experiments on LLaVA-1.5, LLaVA-NeXT, and Qwen2.5-VL demonstrate that retaining only 11% of the original visual tokens achieves 96.7% of the baseline model’s accuracy.

0 citationsRead paper

COREY: A Prototype Study of Entropy-Guided Operator Fusion with Hadamard Reparameterization for Selective State Space Models

Apr 12, 2026

This work addresses the memory bandwidth bottleneck of selective state space models during deployment, which arises from fragmented operator decomposition and redundant intermediate tensor generation in state updates. The authors propose a memory-aware operator fusion strategy that leverages activation entropy as a runtime scheduling metric to dynamically optimize fusion boundaries and tiling sizes. Additionally, they introduce a reparameterized linear projection based on normalized Hadamard transforms to mitigate the heavy-tailed distribution of activations, thereby reducing peak coordinate concentration while preserving functional equivalence. Experimental results demonstrate that the proposed approach significantly reduces DRAM traffic, lowers proxy latency, and improves throughput, outperforming both unfused and fixed-depth baselines.

0 citationsRead paper

Diffusion Epistemic Uncertainty with Asymmetric Learning for Diffusion-Generated Image Detection

Jan 21, 2026

Existing methods for detecting images generated by diffusion models fail to distinguish between aleatoric and epistemic uncertainty, limiting their discriminative performance and generalization capability. This work addresses this limitation by explicitly leveraging epistemic uncertainty for detection—a first in the field—and proposes a Laplace approximation–based approach to estimate epistemic uncertainty in diffusion models. Furthermore, an asymmetric loss function with a large margin is introduced to emphasize the most discriminative components of reconstruction error. The proposed method achieves state-of-the-art performance across multiple large-scale benchmarks and demonstrates significantly improved generalization in detecting images synthesized by previously unseen diffusion models.

0 citationsRead paper

MHA2MLA-VLM: Enabling DeepSeek's Economical Multi-Head Latent Attention across Vision-Language Models

Jan 16, 2026

This work addresses the significant memory and computational bottlenecks caused by the rapid growth of KV cache during inference in vision-language models. To mitigate this, the authors propose a parameter-efficient, multimodal-aware framework that transforms off-the-shelf models into a multi-head latent attention architecture to compress the KV cache and accelerate inference. Key innovations include a modality-adaptive partial RoPE mechanism, modality-disentangled low-rank approximation, and an efficient fine-tuning strategy that minimizes output activation error. With minimal supervised data, the method effectively recovers the original model performance across three mainstream vision-language models while substantially reducing KV cache memory footprint and maintaining natural compatibility with KV quantization techniques.

0 citationsRead paper

The Role of Entropy in Visual Grounding: Analysis and Optimization

Dec 07, 2025

The role of entropy in multimodal large language models (MLLMs) for visual grounding remains poorly understood, and existing entropy regulation strategies lack interpretability and task adaptability. Method: This paper introduces the Entropy-Controlled Visual Grounding Policy Optimization (ECVGPO) algorithm—a reinforcement learning–based approach featuring a dynamic entropy regularization mechanism that adaptively balances exploration and exploitation. Unlike conventional fixed-entropy or black-box entropy control methods, ECVGPO provides explicit, interpretable, and task-aware entropy modulation. Contribution/Results: Extensive experiments demonstrate that ECVGPO significantly improves both performance and training stability across multiple visual grounding benchmarks and mainstream MLLMs. It also exhibits superior generalization capability, establishing a novel paradigm for perception-decision co-optimization in MLLMs.

0 citationsRead paper