VGR: Visual Grounded Reasoning

In the field of multimodal chain-of-thought (CoT) reasoning, existing approaches predominantly rely on reasoning on pure language space, which inherently suffers from language bias and is largely confined to math or science domains. This narrow focus limits their ability to handle complex visual reasoning tasks that demand comprehensive understanding of image details. To address these limitations, this paper introduces VGR, a novel reasoning multimodal large language model (MLLM) with enhanced fine-grained visual perception capabilities. Unlike traditional MLLMs that answer the question or reasoning solely on the language space, our VGR first detects relevant regions that may help to solve problems, and then provides precise answers based on replayed image regions. To achieve this, we conduct a large-scale SFT dataset called VGR -SFT that contains reasoning data with mixed vision grounding and language deduction. The inference pipeline of VGR allows the model to choose bounding boxes for visual reference and a replay stage is introduced to integrates the corresponding regions into the reasoning process, enhancing multimodel comprehension. Experiments on the LLaVA-NeXT-7B baseline show that VGR achieves superior performance on multi-modal benchmarks requiring comprehensive image detail understanding. Compared to the baseline, VGR uses only 30\% of the image token count while delivering scores of +4.1 on MMStar, +7.1 on AI2D, and a +12.9 improvement on ChartQA.

Dynamic Modality Scheduling for Multimodal Large Models via Confidence, Uncertainty, and Semantic Consistency

Multimodal Large Models (MLLMs) have achieved remarkable progress in vision-language understanding and generation tasks. However, existing MLLMs typically rely on static modality fusion strategies, which treat all modalities equally regardless of their instance-level reliability or semantic contribution. This often leads to suboptimal performance, especially in scenarios with noisy, missing, or misaligned modalities. In this paper, we propose Dynamic Modality Scheduling (DMS), a novel framework that adaptively adjusts the contribution of each modality at a per-sample level. DMS evaluates each modality based on three key factors: (1) \textit{confidence}, estimated from predictive entropy; (2) \textit{uncertainty}, obtained via Monte Carlo dropout; and (3) \textit{semantic consistency}, computed through inter-modal similarity. These signals are combined through a learnable or rule-based scheduler to generate soft modality weights used in downstream fusion.To ensure stable training, we further introduce a \textit{Modality Weight Consistency Loss}, which regularizes the fused representation to stay close to unimodal embeddings proportionally to their assigned weights. Our method is model-agnostic and can be integrated into existing MLLMs such as BLIP-2 and LLaVA. Experimental results on VQA, image-text retrieval, and captioning tasks show that DMS significantly improves both clean and robust performance, especially under modality corruption or dropout conditions. This work provides a general and effective mechanism to enable instance-aware and robustness-enhanced multimodal modeling.

Native Visual Understanding: Resolving Resolution Dilemmas in Vision-Language Models

Vision-Language Models (VLMs) face significant challenges when dealing with the diverse resolutions and aspect ratios of real-world images, as most existing models rely on fixed, low-resolution inputs. While recent studies have explored integrating native resolution visual encoding to improve model performance, such efforts remain fragmented and lack a systematic framework within the open-source community. Moreover, existing benchmarks fall short in evaluating VLMs under varied visual conditions, often neglecting resolution as a critical factor. To address the "Resolution Dilemma" stemming from both model design and benchmark limitations, we introduce RC-Bench, a novel benchmark specifically designed to systematically evaluate VLM capabilities under extreme visual conditions, with an emphasis on resolution and aspect ratio variations. In conjunction, we propose NativeRes-LLaVA, an open-source training framework that empowers VLMs to effectively process images at their native resolutions and aspect ratios. Based on RC-Bench and NativeRes-LLaVA, we conduct comprehensive experiments on existing visual encoding strategies. The results show that Native Resolution Visual Encoding significantly improves the performance of VLMs on RC-Bench as well as other resolution-centric benchmarks. Code is available at https://github.com/Niujunbo2002/NativeRes-LLaVA.

Manager: Aggregating Insights from Unimodal Experts in Two-Tower VLMs and MLLMs

Two-Tower Vision--Language Models (VLMs) have demonstrated strong performance across various downstream VL tasks. While BridgeTower further enhances performance by building bridges between encoders, it \textit{(i)} suffers from ineffective layer-by-layer utilization of unimodal representations, \textit{(ii)} restricts the flexible exploitation of different levels of unimodal semantic knowledge, and \textit{(iii)} is limited to the evaluation on traditional low-resolution datasets only with the Two-Tower VLM architecture. In this work, we propose Manager, a lightweight, efficient and effective plugin that adaptively aggregates insights from different levels of pre-trained unimodal experts to facilitate more comprehensive VL alignment and fusion. First, under the Two-Tower VLM architecture, we introduce ManagerTower, a novel VLM that introduces the manager in each cross-modal layer. Whether with or without VL pre-training, ManagerTower outperforms previous strong baselines and achieves superior performance on 4 downstream VL tasks. Moreover, we extend our exploration to the latest Multimodal Large Language Model (MLLM) architecture. We demonstrate that LLaVA-OV-Manager significantly boosts the zero-shot performance of LLaVA-OV across different categories of capabilities, images, and resolutions on 20 downstream datasets, whether the multi-grid algorithm is enabled or not. In-depth analysis reveals that both our manager and the multi-grid algorithm can be viewed as a plugin that improves the visual representation by capturing more diverse visual details from two orthogonal perspectives (depth and width). Their synergy can mitigate the semantic ambiguity caused by the multi-grid algorithm and further improve performance. Code and models are available at https://github.com/LooperXX/ManagerTower.

Quizzard@INOVA Challenge 2025 -- Track A: Plug-and-Play Technique in Interleaved Multi-Image Model

This paper addresses two main objectives. Firstly, we demonstrate the impressive performance of the LLaVA-NeXT-interleave on 22 datasets across three different tasks: Multi-Image Reasoning, Documents and Knowledge-Based Understanding and Interactive Multi-Modal Communication. Secondly, we add the Dense Channel Integration (DCI) connector to the LLaVA-NeXT-Interleave and compare its performance against the standard model. We find that the standard model achieves the highest overall accuracy, excelling in vision-heavy tasks like VISION, NLVR2, and Fashion200K. Meanwhile, the DCI-enhanced version shows particular strength on datasets requiring deeper semantic coherence or structured change understanding such as MIT-States_PropertyCoherence and SlideVQA. Our results highlight the potential of combining powerful foundation models with plug-and-play techniques for Interleave tasks. The code is available at https://github.com/dinhvietcuong1996/icme25-inova.

Can Sound Replace Vision in LLaVA With Token Substitution?

While multimodal systems have achieved impressive advances, they typically rely on text-aligned representations rather than directly integrating audio and visual inputs. This reliance can limit the use of acoustic information in tasks requiring nuanced audio understanding. In response, SoundCLIP explores direct audio-visual integration within multimodal large language models (MLLMs) by substituting CLIP's visual tokens with audio representations and selecting sound-relevant patch tokens in models such as LLaVA. We investigate two configurations: (1) projecting audio features into CLIP's visual manifold via a multilayer perceptron trained with InfoNCE on paired audio-video segments, and (2) preserving raw audio embeddings with minimal dimensional adjustments. Experiments with five state-of-the-art audio encoders reveal a fundamental trade-off. While audio-to-video retrieval performance increases dramatically (up to 44 percentage points in Top-1 accuracy) when audio is projected into CLIP's space, text generation quality declines. Encoders pre-trained with text supervision (CLAP, Whisper, ImageBind) maintain stronger generative capabilities than those focused primarily on audiovisual alignment (Wav2CLIP, AudioCLIP), highlighting the value of language exposure for generation tasks. We introduce WhisperCLIP, an architecture that fuses intermediate representations from Whisper, as well as AudioVisual Event Evaluation (AVE-2), a dataset of 580,147 three-second audiovisual clips with fine-grained alignment annotations. Our findings challenge the assumption that stronger cross-modal alignment necessarily benefits all multimodal tasks; instead, a Pareto frontier emerges wherein optimal performance depends on balancing retrieval accuracy with text generation quality. Codes and datasets: https://github.com/ali-vosoughi/SoundCLIP.

Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs

In multimodal large language models (MLLMs), the length of input visual tokens is often significantly greater than that of their textual counterparts, leading to a high inference cost. Many works aim to address this issue by removing redundant visual tokens. However, current approaches either rely on attention-based pruning, which retains numerous duplicate tokens, or use similarity-based pruning, overlooking the instruction relevance, consequently causing suboptimal performance. In this paper, we go beyond attention or similarity by proposing a novel visual token pruning method named CDPruner, which maximizes the conditional diversity of retained tokens. We first define the conditional similarity between visual tokens conditioned on the instruction, and then reformulate the token pruning problem with determinantal point process (DPP) to maximize the conditional diversity of the selected subset. The proposed CDPruner is training-free and model-agnostic, allowing easy application to various MLLMs. Extensive experiments across diverse MLLMs show that CDPruner establishes new state-of-the-art on various vision-language benchmarks. By maximizing conditional diversity through DPP, the selected subset better represents the input images while closely adhering to user instructions, thereby preserving strong performance even with high reduction ratios. When applied to LLaVA, CDPruner reduces FLOPs by 95\% and CUDA latency by 78\%, while maintaining 94\% of the original accuracy. Our code is available at https://github.com/Theia-4869/CDPruner.

LLaVA-c: Continual Improved Visual Instruction Tuning

Multimodal models like LLaVA-1.5 achieve state-of-the-art visual understanding through visual instruction tuning on multitask datasets, enabling strong instruction-following and multimodal performance. However, multitask learning faces challenges such as task balancing, requiring careful adjustment of data proportions, and expansion costs, where new tasks risk catastrophic forgetting and need costly retraining. Continual learning provides a promising alternative to acquiring new knowledge incrementally while preserving existing capabilities. However, current methods prioritize task-specific performance, neglecting base model degradation from overfitting to specific instructions, which undermines general capabilities. In this work, we propose a simple but effective method with two modifications on LLaVA-1.5: spectral-aware consolidation for improved task balance and unsupervised inquiry regularization to prevent base model degradation. We evaluate both general and task-specific performance across continual pretraining and fine-tuning. Experiments demonstrate that LLaVA-c consistently enhances standard benchmark performance and preserves general capabilities. For the first time, we show that task-by-task continual learning can achieve results that match or surpass multitask joint learning. The code will be publicly released.

Autoregressive Semantic Visual Reconstruction Helps VLMs Understand Better

Typical large vision-language models (LVLMs) apply autoregressive supervision solely to textual sequences, without fully incorporating the visual modality into the learning process. This results in three key limitations: (1) an inability to utilize images without accompanying captions, (2) the risk that captions omit critical visual details, and (3) the challenge that certain vision-centric content cannot be adequately conveyed through text. As a result, current LVLMs often prioritize vision-to-language alignment while potentially overlooking fine-grained visual information. While some prior works have explored autoregressive image generation, effectively leveraging autoregressive visual supervision to enhance image understanding remains an open challenge. In this paper, we introduce Autoregressive Semantic Visual Reconstruction (ASVR), which enables joint learning of visual and textual modalities within a unified autoregressive framework. We show that autoregressively reconstructing the raw visual appearance of images does not enhance and may even impair multimodal understanding. In contrast, autoregressively reconstructing the semantic representation of images consistently improves comprehension. Notably, we find that even when models are given continuous image features as input, they can effectively reconstruct discrete semantic tokens, resulting in stable and consistent improvements across a wide range of multimodal understanding benchmarks. Our approach delivers significant performance gains across varying data scales (556k-2M) and types of LLM bacbones. Specifically, ASVR improves LLaVA-1.5 by 5% in average scores across 14 multimodal benchmarks. The code is available at https://github.com/AlenjandroWang/ASVR.

Learning Compact Vision Tokens for Efficient Large Multimodal Models

Large multimodal models (LMMs) suffer significant computational challenges due to the high cost of Large Language Models (LLMs) and the quadratic complexity of processing long vision token sequences. In this paper, we explore the spatial redundancy among vision tokens and shorten the length of vision token sequences for inference acceleration. Specifically, we propose a Spatial Token Fusion (STF) method to learn compact vision tokens for short vision token sequence, where spatial-adjacent tokens are fused into one. Meanwhile, weight-frozen vision encoder can not well adapt to the demand of extensive downstream vision-language tasks. To this end, we further introduce a Multi-Block Token Fusion (MBTF) module to supplement multi-granularity features for the reduced token sequence. Overall, we combine STF and MBTF module to balance token reduction and information preservation, thereby improving inference efficiency without sacrificing multimodal reasoning capabilities. Experimental results demonstrate that our method based on LLaVA-1.5 achieves comparable or even superior performance to the baseline on 8 popular vision-language benchmarks with only $25\%$ vision tokens of baseline. The source code and trained weights are available at https://github.com/visresearch/LLaVA-STF.