Yuchen Feng

Yuchen Feng, Zhenyu Zhang, Naibin Gu et al.

Multimodal large language models (MLLMs) have achieved remarkable progress on various vision-language tasks, yet their visual perception remains limited. Humans, in comparison, perceive complex scenes efficiently by dynamically scanning and focusing on salient regions in a sequential "blink-like" process. Motivated by this strategy, we first investigate whether MLLMs exhibit similar behavior. Our pilot analysis reveals that MLLMs naturally attend to different visual regions across layers and that selectively allocating more computation to salient tokens can enhance visual perception. Building on this insight, we propose Blink, a dynamic visual token resolution framework that emulates the human-inspired process within a single forward pass. Specifically, Blink includes two modules: saliency-guided scanning and dynamic token resolution. It first estimates the saliency of visual tokens in each layer based on the attention map, and extends important tokens through a plug-and-play token super-resolution (TokenSR) module. In the next layer, it drops the extended tokens when they lose focus. This dynamic mechanism balances broad exploration and fine-grained focus, thereby enhancing visual perception adaptively and efficiently. Extensive experiments validate Blink, demonstrating its effectiveness in enhancing visual perception and multimodal understanding.

Jiacheng Ruan, Yebin Yang, Zehao Lin et al.

Benefiting from the revolutionary advances in large language models (LLMs) and foundational vision models, large vision-language models (LVLMs) have also made significant progress. However, current benchmarks focus on tasks that evaluating only a single aspect of LVLM capabilities (e.g., recognition, detection, understanding). These tasks fail to fully demonstrate LVLMs' potential in complex application scenarios. To comprehensively assess the performance of existing LVLMs, we propose a more challenging task called the Flow Text with Image Insertion task (FTII). This task requires LVLMs to simultaneously possess outstanding abilities in image comprehension, instruction understanding, and long-text interpretation. Specifically, given several text paragraphs and a set of candidate images, as the text paragraphs accumulate, the LVLMs are required to select the most suitable image from the candidates to insert after the corresponding paragraph. Constructing a benchmark for such a task is highly challenging, particularly in determining the sequence of flowing text and images. To address this challenge, we turn to professional news reports, which naturally contain a gold standard for image-text sequences. Based on this, we introduce the Flow Text with Image Insertion Benchmark (FTII-Bench), which includes 318 high-quality Chinese image-text news articles and 307 high-quality English image-text news articles, covering 10 different news domains. Using these 625 high-quality articles, we construct problems of two different types with multiple levels of difficulty. Furthermore, we establish two different evaluation pipelines based on the CLIP model and existing LVLMs. We evaluate 9 open-source and 2 closed-source LVLMs as well as 2 CLIP-based models. Results indicate that even the most advanced models (e.g., GPT-4o) face significant challenges when tackling the FTII task.

Yuchen Feng, Bowen Shen, Naibin Gu et al.

Large language models (LLMs) with the Mixture-of-Experts (MoE) architecture achieve high cost-efficiency by selectively activating a subset of the parameters. Despite the inference efficiency of MoE LLMs, the training of extensive experts from scratch incurs substantial overhead, whereas reconstructing a dense LLM into an MoE LLM significantly reduces the training budget. However, existing reconstruction methods often overlook the diversity among experts, leading to potential redundancy. In this paper, we come up with the observation that a specific LLM exhibits notable diversity after being pruned on different calibration datasets, based on which we present a Diversity-Enhanced reconstruction method named DIVE. The recipe of DIVE includes domain affinity mining, pruning-based expert reconstruction, and efficient retraining. Specifically, the reconstruction includes pruning and reassembly of the feed-forward network (FFN) module. After reconstruction, we efficiently retrain the model on routers, experts and normalization modules. We implement DIVE on Llama-style LLMs with open-source training corpora. Experiments show that DIVE achieves training efficiency with minimal accuracy trade-offs, outperforming existing pruning and MoE reconstruction methods with the same number of activated parameters.

Yilong Chen, Naibin Gu, Junyuan Shang et al.

Mixture-of-Experts (MoE) decouples model capacity from per-token computation, yet their scalability remains limited by the physical dimensions of depth and width. To overcome this, we propose Mixture of Universal Experts (MOUE),a MoE generalization introducing a novel scaling dimension: Virtual Width. In general, MoUE aims to reuse a universal layer-agnostic expert pool across layers, converting depth into virtual width under a fixed per-token activation budget. However, two challenges remain: a routing path explosion from recursive expert reuse, and a mismatch between the exposure induced by reuse and the conventional load-balancing objectives. We address these with three core components: a Staggered Rotational Topology for structured expert sharing, a Universal Expert Load Balance for depth-aware exposure correction, and a Universal Router with lightweight trajectory state for coherent multi-step routing. Empirically, MoUE consistently outperforms matched MoE baselines by up to 1.3% across scaling regimes, enables progressive conversion of existing MoE checkpoints with up to 4.2% gains, and reveals a new scaling dimension for MoE architectures.

Jihao Zhao, Zhiyuan Ji, Yuchen Feng et al.

While Retrieval-Augmented Generation (RAG) has emerged as a promising paradigm for boosting large language models (LLMs) in knowledge-intensive tasks, it often overlooks the crucial aspect of text chunking within its workflow. This paper proposes the Meta-Chunking framework, which specifically enhances chunking quality through a dual strategy that identifies optimal segmentation points and preserves global information. Initially, breaking limitations of similarity-based chunking, we design two adaptive chunking techniques based on uncertainty, namely Perplexity Chunking and Margin Sampling Chunking, by utilizing the logical perception capabilities of LLMs. Given the inherent complexity across different texts, we integrate meta-chunk with dynamic merging, striking a balance between fine-grained and coarse-grained text chunking. Furthermore, we establish the global information compensation mechanism, encompassing a two-stage hierarchical summary generation process and a three-stage text chunk rewriting procedure focused on missing reflection, refinement, and completion. These components collectively strengthen the semantic integrity and contextual coherence of chunks. Extensive experiments demonstrate that Meta-Chunking effectively addresses challenges of the chunking task within the RAG system, providing LLMs with more logically coherent text chunks. Additionally, our methodology validates the feasibility of implementing high-quality chunking tasks with smaller-scale models, thereby eliminating the reliance on robust instruction-following capabilities.

Naibin Gu, Zhenyu Zhang, Yuchen Feng et al.

Mixture-of-Experts (MoE) models typically fix the number of activated experts $k$ at both training and inference. Intuitively, activating more experts at inference $k'$ (where $k'> k$) means engaging a larger set of model parameters for the computation and thus is expected to improve performance. However, contrary to this intuition, we find the scaling range to be so narrow that performance begins to degrade rapidly after only a slight increase in the number of experts. Further investigation reveals that this degradation stems from a lack of learned collaboration among experts. To address this, we introduce Elastic Mixture-of-Experts (EMoE), a novel training framework that enables MoE models to scale the number of activated experts at inference without incurring additional training overhead. By simultaneously training experts to collaborate in diverse combinations and encouraging the router for high-quality selections, EMoE ensures robust performance across computational budgets at inference. We conduct extensive experiments on various MoE settings. Our results show that EMoE significantly expands the effective performance-scaling range, extending it to as much as 2-3$\times$ the training-time $k$, while also pushing the model's peak performance to a higher level.

Jiaxuan Zhao, Naibin Gu, Yuchen Feng et al.

The high costs of customizing large language models (LLMs) fundamentally limit their adaptability to user-specific needs. Consequently, LLMs are increasingly offered as cloud-based services, a paradigm that introduces critical limitations: providers struggle to support personalized customization at scale, while users face privacy risks when exposing sensitive data. To address this dual challenge, we propose Customized Black-box Prompt Tuning (CBP-Tuning), a novel framework that facilitates efficient local customization while preserving bidirectional privacy. Specifically, we design a two-stage framework: (1) a prompt generator trained on the server-side to capture domain-specific and task-agnostic capabilities, and (2) user-side gradient-free optimization that tailors soft prompts for individual tasks. This approach eliminates the need for users to access model weights or upload private data, requiring only a single customized vector per task while achieving effective adaptation. Furthermore, the evaluation of CBP-Tuning in the commonsense reasoning, medical and financial domain settings demonstrates superior performance compared to baselines, showcasing its advantages in task-agnostic processing and privacy preservation.

Bo Tang, Junyi Zhu, Chenyang Xi et al.

Traditional search engines struggle to synthesize fragmented information for complex queries, while generative AI search engines face challenges in relevance, comprehensiveness, and presentation. To address these limitations, we introduce Xinyu AI Search, a novel system that incorporates a query-decomposition graph to dynamically break down complex queries into sub-queries, enabling stepwise retrieval and generation. Our retrieval pipeline enhances diversity through multi-source aggregation and query expansion, while filtering and re-ranking strategies optimize passage relevance. Additionally, Xinyu AI Search introduces a novel approach for fine-grained, precise built-in citation and innovates in result presentation by integrating timeline visualization and textual-visual choreography. Evaluated on recent real-world queries, Xinyu AI Search outperforms eight existing technologies in human assessments, excelling in relevance, comprehensiveness, and insightfulness. Ablation studies validate the necessity of its key sub-modules. Our work presents the first comprehensive framework for generative AI search engines, bridging retrieval, generation, and user-centric presentation.