Multimodal Large Language Models

Spatial reasoning is a cornerstone capability for intelligent systems to perceive and interact with the physical world. However, multimodal large language models (MLLMs) frequently suffer from hallucinations and imprecision when parsing complex geometric layouts. As data-driven scaling struggles to internalize structured geometric priors and spatial constraints, integrating mature, specialized vision models presents a compelling alternative. Despite its promise, applying this paradigm to spatial reasoning is hindered by two key challenges: The difficulty of invoking heterogeneous, parameter-rich tools, as well as the challenge of understanding and effectively leveraging their diverse low-level outputs (e.g., segmentation masks, depth maps) in high-level reasoning. To address these challenges, we propose LAST, a unified framework for tool-augmented spatial reasoning. LAST features an extensible interactive sandbox, termed LAST-Box, which abstracts heterogeneous tool invocations into atomic instructions and reusable spatial skills, returning multimodal hints (e.g., annotated images and textual descriptions) that can be directly consumed by LLMs. We further design a three-stage progressive training strategy that guides models from understanding tool outputs to proficient and adaptive tool invocation. Experiments on four datasets show that LAST-7B achieves around 20% performance gains over its backbone and outperforms strong proprietary closed-source LLMs, substantially enhancing reasoning on complex spatial tasks.

AI applications driven by multimodal large language models (MLLMs) are prone to hallucinations and pose considerable risks to human users. Crucially, such hallucinations are not equally problematic: some hallucination contents could be detected by human users(i.e., obvious hallucinations), while others are often missed or require more verification effort(i.e., elusive hallucinations). This indicates that multimodal AI hallucinations vary significantly in their verifiability. Yet, little research has explored how to control this property for AI applications with diverse security and usability demands. To address this gap, we construct a dataset from 4,470 human responses to AI-generated hallucinations and categorize these hallucinations into obvious and elusive types based on their verifiability by human users. Further, we propose an activation-space intervention method that learns separate probes for obvious and elusive hallucinations. We reveal that obvious and elusive hallucinations elicit different intervention probes, allowing for fine-grained control over the model's verifiability. Empirical results demonstrate the efficacy of this approach and show that targeted interventions yield superior performance in regulating corresponding verifiability. Moreover, simply mixing these interventions enables flexible control over the verifiability required for different scenarios.

Multimodal news often loses its image or text as it spreads across Web platforms. Existing methods treat this as information loss, assuming that the original veracity label remains valid. This assumption fails for fake news reusing authentic material, whose remainder may be genuine. Detectors should therefore decouple modality cues and adapt their contributions to modality availability. However, coarse news labels provide little guidance for modality decoupling, while training on complete inputs can induce modality bias. To address these challenges, we propose Head-wise Modality Decoupling and Regulation (HMDR) for multimodal large language models. First, Modality Head Decoupling decouples modality representations through soft head grouping by attention preference. Second, Modality Head Regulation regulates head contributions through compensation for imbalanced modality attention. Finally, Modality Invariance Learning promotes invariance by aligning representations of each modality under replacement of the other. Experiments on two benchmarks show that HMDR improves detection under missing modalities while maintaining performance on complete inputs.

Cultural values are expressed not only through language but also through visual scenes and everyday social practices. Yet existing evaluations of cultural values in language models are almost entirely text-only, leaving it unclear whether culture-conditioned judgments remain stable when response options are visualized. We introduce ValueGround, a benchmark for evaluating culture-conditioned visual value grounding in multimodal large language models (MLLMs). Built from World Values Survey questions, ValueGround uses minimally contrastive image pairs to represent opposing response options while controlling irrelevant variation. Given a country, a question, and an image pair, a model must choose the image that best matches the country's value tendency without access to the original response-option texts. Experiments across six MLLMs and 13 countries show that models perform substantially worse with visualized response options than with the original textual options, with average accuracy dropping from 72.8% to 62.6%. Our benchmark provides a controlled testbed for studying cross-modal transfer of culture-conditioned value judgments.

Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key-value (KV) caches. Each visual input expands into thousands of tokens, causing caches to scale linearly with context length and remain resident in GPU memory throughout decoding, which leads to prohibitive memory overhead and latency even on high-end GPUs. A common solution is to compress caches under a fixed allocated budget at different granularities: token-level uniformly discards less important tokens, layer-level varies retention across layers, and head-level redistributes budgets across heads. Yet these approaches stop at allocation and overlook the heterogeneous behaviors of attention heads that require distinct compression strategies. We propose HybridKV, a hybrid KV cache compression framework that integrates complementary strategies in three stages: heads are first classified into static or dynamic types using text-centric attention; then a top-down budget allocation scheme hierarchically assigns KV budgets; finally, static heads are compressed by text-prior pruning and dynamic heads by chunk-wise retrieval. Experiments on 11 multimodal benchmarks with Qwen2.5-VL-7B show that HybridKV reduces KV cache memory by up to $7.9\times$ and achieves $1.52\times$ faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in 2D visual tasks but still struggle to understand physical space in real-world visual streams. Recently, feed-forward geometric foundation models that implicitly extract geometric priors from visual inputs have offered a promising direction. However, many existing geometry-aware MLLMs extract features from a single deep encoder layer and perform one-shot fusion at the language-model input, potentially overlooking complementary information across geometric encoder layers, limiting access to fine-grained geometric cues, and hindering progressive cross-modal adaptation. To address these limitations, we propose GUIDE (Geometric Unrolling Inside MLLM Early-layers), a progressive framework for integrating geometric priors. GUIDE samples and aligns multi-level features from the geometric encoder and sequentially injects them into the early layers of the MLLM. This design enables the MLLM to continuously access multi-granularity geometric cues and progressively integrate them with visual representations during layer-wise semantic abstraction. GUIDE further introduces a dual context-aware gating mechanism at both the token and layer levels, selectively regulating geometric information to reduce redundant injection and interference with pretrained representations. Extensive experiments on spatial reasoning and 3D scene understanding benchmarks, including VSI-Bench, ScanRefer, and Scan2Cap, validate the effectiveness and cross-task applicability of GUIDE. Our 5B and 9B models achieve average scores of 71.5 and 72.1 on VSI-Bench, respectively.

Omni Large Language Models (Omni-LLMs) have demonstrated impressive capabilities in holistic multi-modal perception, yet they consistently falter in complex scenarios requiring synergistic omni-modal reasoning. Beyond understanding global multimodal context, effective reasoning also hinges on fine-grained cross-modal alignment, especially identifying shared referents across modalities, yet this aspect has been largely overlooked. To bridge this gap, we formalize the challenge as a cross-modal coreference problem, where a model must localize a referent in a source modality and re-identify it in a target modality. Building on this paradigm, we introduce CrossOmni, a dataset comprising nine tasks equipped with human-designed reasoning rationales to evaluate and enhance this capability. Experiments on 13 Omni-LLMs reveal systematic weaknesses in cross-modal coreference, which we attribute to the absence of coreference-aware thinking patterns. To address this, we enhance cross-modal alignment via two strategies: a training-free In-Context Learning method and a training-based SFT+GRPO framework designed to induce such thinking patterns. Both approaches yield substantial performance gains and generalize effectively to collaborative reasoning tasks. Overall, our findings highlight cross-modal coreference as a crucial missing piece for advancing robust omni-modal reasoning.

Multimodal large language models have become an important infrastructure for unified processing of visual and linguistic tasks. However, such models are highly susceptible to backdoor implantation during supervised fine-tuning and will steadily output the attacker's predefined harmful responses once a specific trigger pattern is activated. The core challenge of backdoor defense lies in suppressing attack success under low poisoning ratios while preserving the model's normal generation ability. These two objectives are inherently conflicting. Strong suppression often degrades benign performance, whereas weak regularization fails to mitigate backdoor behaviors. To this end, we propose a unified defense framework based on patch augmentation and cross-view regularity, which simultaneously constrains the model's anomalous behaviors in response to triggered patterns from both the feature representation and output distribution levels. Specifically, patch-level data augmentation is combined with cross-view output difference regularization to exploit the fact that backdoor responses are abnormally invariant to non-semantic perturbations and to proactively pull apart the output distributions of the original and perturbed views, thereby significantly suppressing the success rate of backdoor triggering. At the same time, we avoid over-suppression of the model during defense by imposing output entropy constraints, ensuring the quality of normal command generation. Experimental results across three models, two tasks, and six attacks show that our proposed defense method effectively reduces the attack success rate while maintaining a high level of normal text generation capability. Our work enables the secure, controlled deployment of large-scale multimodal models in realistic low-frequency poisoning and covert triggering scenarios.

Large Language Models (LLMs) have significantly advanced molecular discovery, but existing multimodal molecular architectures fundamentally rely on autoregressive (AR) backbones. This strict left-to-right inductive bias is sub-optimal for generating chemically valid molecules, as it struggles to account for non-local global constraints (e.g., ring closures) and often accumulates structural errors during sequential generation. To address these limitations, we propose MolDA (Molecular language model with masked Diffusion with mAsking), a novel multimodal framework that replaces the conventional AR backbone with a discrete Large Language Diffusion Model. MolDA extracts comprehensive structural representations using a hybrid graph encoder, which captures both local and global topologies, and aligns them into the language token space via a Q-Former. Furthermore, we mathematically reformulate Molecular Structure Preference Optimization specifically for the masked diffusion. Through bidirectional iterative denoising, MolDA ensures global structural coherence, chemical validity, and robust reasoning across molecule generation, captioning, and property prediction.

As audio-visual multi-modal large language models (MLLMs) are increasingly deployed in safety-critical applications, understanding their cross-modal vulnerabilities is essential. To this end, we introduce Multi-Modal Typography, a systematic study examining how perturbations added to speech and on-screen text impact audio-visual reasoning of closed- and open-source models. For instance, on the WorldSense benchmark, the attack success rate (ASR) against Gemini 3.7 Flash increased from 9.53% to 43.41%, highlighting its vulnerability. We analyze the cross-modal effects of unimodal and multimodal attacks and evaluate how specific properties of the attacks impact the overall multi-modal reasoning. By applying fine-tuning as a mitigation, we lower the ASR of Qwen2.5-Omni-7B on WorldSense from $60.6\%$ to $52.0\%$. Our findings across multi-modal reasoning tasks, content moderation, and offline robot task selection benchmarks establish multi-modal typography as a critical and underexplored attack strategy in multi-modal reasoning. Code and data will be made publicly available.

Understanding student engagement usually requires time-consuming manual observation or invasive recording that raises privacy concerns. We present a privacy-preserving pipeline that analyzes classroom videos to extract insights about student attention, without storing any identifiable footage. Our system runs on a single GPU, using OpenPose for skeletal extraction and Gaze-LLE for visual attention estimation. Original video frames are deleted immediately after pose extraction, thus only geometric coordinates (stored as JSON) are retained, ensuring compliance with FERPA. The extracted pose and gaze data is processed by QwQ-32B-Reasoning, which performs zero-shot analysis of student behavior across lecture segments. Instructors access results through a web dashboard featuring attention heatmaps and behavioral summaries. Our preliminary findings suggest that LLMs may show promise for multimodal behavior understanding, although they still struggle with spatial reasoning about classroom layouts. We discuss these limitations and outline directions for improving LLM spatial comprehension in educational analytics contexts.

For multimodal large language models (MLLMs), visual information is relatively sparse compared with text. As a result, research on visual pruning emerges for efficient inference. Current approaches typically measure token importance based on the attention scores in the visual encoder or in the LLM decoder, then select visual tokens with high attention scores while pruning others. In this paper, we pursue a different and more surgical approach. Instead of relying on mechanism-specific signals, we directly compute Mutual Information (MI) between visual and textual features themselves, prior to their interaction. This allows us to explicitly measure crossmodal dependency at the feature levels. Our MI-Pruner is simple, efficient and non-intrusive, requiring no access to internal attention maps or architectural modifications. Experimental results demonstrate that our approach outperforms previous attention-based pruning methods with minimal latency.

Multimodal Large Language Models (MLLMs) excel in general domains but struggle with complex, real-world science. We posit that polymer science, an interdisciplinary field spanning chemistry, physics, biology, and engineering, is an ideal high-stakes testbed due to its diverse multimodal data. Yet, existing benchmarks related to polymer science largely overlook real-world workflows, limiting their practical utility and failing to systematically evaluate MLLMs across the full, practice-grounded lifecycle of experimentation. We introduce PolyReal, a novel multimodal benchmark grounded in real-world scientific practices to evaluate MLLMs on the full lifecycle of polymer experimentation. It covers five critical capabilities: (1) foundational knowledge application; (2) lab safety analysis; (3) experiment mechanism reasoning; (4) raw data extraction; and (5) performance & application exploration. Our evaluation of leading MLLMs on PolyReal reveals a capability imbalance. While models perform well on knowledge-intensive reasoning (e.g., Experiment Mechanism Reasoning), they drop sharply on practice-based tasks (e.g., Lab Safety Analysis and Raw Data Extraction). This exposes a severe gap between abstract scientific knowledge and its practical, context-dependent application, showing that these real-world tasks remain challenging for MLLMs. Thus, PolyReal helps address this evaluation gap and provides a practical benchmark for assessing AI systems in real-world scientific workflows.

Can warping tokens, rather than pixels, help multimodal large language models (MLLMs) understand how a scene appears from a nearby viewpoint? While MLLMs perform well on visual reasoning, they remain fragile to viewpoint changes, as pixel-wise warping is highly sensitive to small depth errors and often introduces geometric distortions. Drawing on theories of mental imagery that posit part-level structural representations as the basis for human perspective transformation, we examine whether image tokens in ViT-based MLLMs serve as an effective substrate for viewpoint changes. We compare forward and backward warping, finding that backward token warping, which defines a dense grid on the target view and retrieves a corresponding source-view token for each grid point, achieves greater stability and better preserves semantic coherence under viewpoint shifts. Experiments on our proposed ViewBench benchmark demonstrate that token-level warping enables MLLMs to reason reliably from nearby viewpoints, consistently outperforming all baselines including pixel-wise warping approaches, spatially fine-tuned MLLMs, and a generative warping method.

Multimodal Large Language Models (MLLMs) have shown strong reasoning ability, but their high computational and memory costs hinder deployment in resource-constrained settings. While Post-Training Quantization (PTQ) and vision token pruning are standard compression techniques, they are usually treated as independent optimizations. In this paper, we show that these two techniques are strongly coupled: naively applying semantic-based token pruning to PTQ-optimized MLLMs can discard activation outliers that are important for numerical stability and thus worsen quantization errors in low-bit regimes (e.g., W4A4). To address this issue, we propose a quantization-aware vision token pruning framework. Our method introduces a lightweight hybrid sensitivity metric that combines simulated group-wise quantization error with outlier intensity. By combining this metric with standard semantic relevance scores, the method retains tokens that are both semantically informative and robust to quantization. Experiments on standard LLaVA architectures show that our method consistently outperforms naive integration baselines. At an aggressive pruning ratio that retains only 12.5% of visual tokens, our framework improves accuracy by 2.24% over the baseline and even surpasses dense quantization without pruning. To the best of our knowledge, this is the first method that explicitly co-optimizes vision token pruning and PTQ for accurate low-bit MLLM inference.

Domain reweighting can improve sample efficiency and downstream generalization, but data-mixture optimization for multimodal midtraining remains largely unexplored. Current multimodal training recipes tune mixtures along a single dimension, typically data format or task type. We introduce MixAtlas, a method that produces benchmark-targeted data recipes that can be inspected, adapted, and transferred to new corpora. MixAtlas decomposes the training corpus along two axes: image concepts (10 visual-domain clusters discovered via CLIP embeddings) and task supervision (5 objective types including captioning, OCR, grounding, detection, and VQA). Using small proxy models (Qwen2-0.5B) paired with a Gaussian-process surrogate and GP-UCB acquisition, MixAtlas searches the resulting mixture space with the same proxy budget as regression-based baselines but finds better-performing mixtures. We evaluate on 10 benchmarks spanning visual understanding, document reasoning, and multimodal reasoning. On Qwen2-7B, optimized mixtures improve average performance by 8.5%-17.6% over the strongest baseline; on Qwen2.5-7B, gains are 1.0%-3.3%. Both settings reach baseline-equivalent training loss in up to 2 times fewer steps. Recipes discovered on 0.5B proxies transfer to 7B-scale training across Qwen model families.

Multimodal large language models (MLLMs) have shown strong potential for autonomous driving, yet existing benchmarks remain largely ego-centric and therefore cannot systematically assess model performance in infrastructure-centric and cooperative driving conditions. In this work, we introduce V2X-QA, a real-world dataset and benchmark for evaluating MLLMs across vehicle-side, infrastructure-side, and cooperative viewpoints. V2X-QA is built around a view-decoupled evaluation protocol that enables controlled comparison under vehicle-only, infrastructure-only, and cooperative driving conditions within a unified multiple-choice question answering (MCQA) framework. The benchmark is organized into a twelve-task taxonomy spanning perception, prediction, and reasoning and planning, and is constructed through expert-verified MCQA annotation to enable fine-grained diagnosis of viewpoint-dependent capabilities. Benchmark results across ten representative state-of-the-art proprietary and open-source models show that viewpoint accessibility substantially affects performance, and infrastructure-side reasoning supports meaningful macroscopic traffic understanding. Results also indicate that cooperative reasoning remains challenging since it requires cross-view alignment and evidence integration rather than simply additional visual input. To address these challenges, we introduce V2X-MoE, a benchmark-aligned baseline with explicit view routing and viewpoint-specific LoRA experts. The strong performance of V2X-MoE further suggests that explicit viewpoint specialization is a promising direction for multi-view reasoning in autonomous driving. Overall, V2X-QA provides a foundation for studying multi-perspective reasoning, reliability, and cooperative physical intelligence in connected autonomous driving. The dataset and V2X-MoE resources are publicly available at: https://github.com/junwei0001/V2X-QA.

Audio-Visual Large Language Models (AVLLMs) are emerging as unified interfaces to multimodal perception. We present the first mechanistic interpretability study of AVLLMs, analyzing how audio and visual features evolve and fuse through different layers of an AVLLM to produce the final text outputs. We find that although AVLLMs encode rich audio semantics at intermediate layers, these capabilities largely fail to surface in the final text generation when audio conflicts with vision. Probing analyses show that useful latent audio information is present, but deeper fusion layers disproportionately privilege visual representations that tend to suppress audio cues. We further trace this imbalance to training: the AVLLM's audio behavior strongly matches its vision-language base model, indicating limited additional alignment to audio supervision. Our findings reveal a fundamental modality bias in AVLLMs and provide new mechanistic insights into how multimodal LLMs integrate audio and vision.

Multimodal Large Language Models (MLLMs) offer an opportunity to support multimedia learning through conversational systems grounded in educational content. However, while conversational AI is known to boost engagement, its impact on learning in visually-rich STEM domains remains under-explored. Moreover, there is limited understanding of how multimodality and conversationality jointly influence learning in generative AI systems. This work reports findings from a randomized controlled online study (N = 124) comparing three approaches to learning biology from textbook content: (1) a document-grounded conversational AI with interleaved text-and-image responses (MuDoC), (2) a document-grounded conversational AI with text-only responses (TexDoC), and (3) a textbook interface with semantic search and highlighting (DocSearch). Learners using MuDoC achieved the highest post-test scores and reported the most positive learning experience. Notably, while TexDoC was rated as significantly more engaging and easier to use than DocSearch, it led to the lowest post-test scores, revealing a disconnect between student perceptions and learning outcomes. Interpreted through the lens of the Cognitive Load Theory, these findings suggest that conversationality reduces extraneous load, while visual-verbal integration induced by multimodality increases germane load, leading to better learning outcomes. When conversationality is not complemented by multimodality, reduced cognitive effort may instead inflate perceived understanding without improving learning outcomes.

We present the first systematic analysis of multimodal large language models (MLLMs) in personalized question-answering requiring ego-grounding - the ability to understand the camera-wearer in egocentric videos. To this end, we introduce MyEgo, the first egocentric VideoQA dataset designed to evaluate MLLMs' ability to understand, remember, and reason about the camera wearer. MyEgo comprises 541 long videos and 5K personalized questions asking about "my things", "my activities", and "my past". Benchmarking reveals that competitive MLLMs across variants, including open-source vs. proprietary, thinking vs. non-thinking, small vs. large scales all struggle on MyEgo. Top closed- and open-source models (e.g., GPT-5 and Qwen3-VL) achieve only~46% and 36% accuracy, trailing human performance by near 40% and 50% respectively. Surprisingly, neither explicit reasoning nor model scaling yield consistent improvements. Models improve when relevant evidence is explicitly provided, but gains drop over time, indicating limitations in tracking and remembering "me" and "my past". These findings collectively highlight the crucial role of ego-grounding and long-range memory in enabling personalized QA in egocentric videos. We hope MyEgo and our analyses catalyze further progress in these areas for egocentric personalized assistance. Data and code are available at https://github.com/Ryougetsu3606/MyEgo

Smell's connections with food, memory, and social experience have motivated researchers to bring olfaction into interactive systems. Multimodal AI opens new possibilities for generating smells from natural language, yet it remains unclear whether pretrained models encode sufficient olfactory knowledge to translate language into perceptible compositions. We present AromaGen, an AI-powered 12-channel wearable olfactory system that maps free-form natural-language descriptions to 12 base odorants selected to cover a semantically derived olfactory space, while supporting iterative refinement through natural-language feedback. In a between-subjects study (N=60) across 50 real-world benchmark smells, participants distinguished AromaGen-generated target smells above chance in a three-alternative forced-choice task, with retrieval-augmented AI composition performing comparably to human-expert composition and zero-shot AI composition. Natural-language feedback significantly improved the perceived similarity of both AI- and human-expert compositions. Our findings demonstrate the feasibility of using pretrained multimodal AI, with human aroma composition data, to generate perceptible olfactory experiences from natural language.

Multimodal Large Language Models (MLLMs) have recently demonstrated promising capabilities in multimodal coding tasks such as chart-to-code generation. However, existing methods primarily rely on supervised fine-tuning (SFT), which requires the model to learn code patterns through chart-code pairs but does not expose the model to a code execution environment. Moreover, while self-correction through execution feedback offers a potential route to improve coding quality, even state-of-the-art MLLMs have been shown to struggle with effective self-correction. In this work, we introduce MM-ReCoder, a chart-to-code generation model trained with reinforcement learning (RL) and equipped with self-correction ability. We propose a two-stage multi-turn self-correction RL strategy based on Group Relative Policy Optimization (GRPO). The first stage enhances the model's self-correction ability via rolling out a shared first turn, while the second stage improves the coding capability with full-trajectory optimization. MM-ReCoder learns to produce more accurate and executable code through the interaction with the environment and by iteratively correcting its own outputs. Our results on three chart-to-code benchmarks demonstrate the state-of-the-art performance of MM-ReCoder.

Multimodal large language models (MLLMs) are increasingly being applied to spatial cognition tasks, where they are expected to understand and interact with complex environments. Most existing works improve spatial reasoning by introducing 3D priors or geometric supervision, which enhances performance but incurs substantial data preparation and alignment costs. In contrast, purely 2D approaches often struggle with multi-frame spatial reasoning due to their limited ability to capture cross-frame spatial relationships. To address these limitations, we propose EgoMind, a Chain-of-Thought framework that enables geometry-free spatial reasoning through Role-Play Caption, which jointly constructs a coherent linguistic scene graph across frames, and Progressive Spatial Analysis, which progressively reasons toward task-specific questions. With only 5K auto-generated SFT samples and 20K RL samples, EgoMind achieves competitive results on VSI-Bench, SPAR-Bench, SITE-Bench, and SPBench, demonstrating its effectiveness in strengthening the spatial reasoning capabilities of MLLMs and highlighting the potential of linguistic reasoning for spatial cognition. Code and data are released at https://github.com/Hyggge/EgoMind.

Spatial consistency is a fundamental property of the visual world and a key requirement for models that aim to understand physical reality. Despite recent advances, multimodal large language models (MLLMs) often struggle to reason about 3D geometry across multiple views. Rather than asking models to describe scene attributes, we introduce a more challenging task: given two views of the same scene, identify the object that violates 3D motion consistency. We propose a simple and scalable method for generating realistic, spatially inconsistent image pairs from multi-view scenes, enabling systematic evaluation of this capability. Our results show that state-of-the-art MLLMs significantly underperform human observers and exhibit substantial variability across different scene attributes, revealing a fragile and incomplete understanding of 3D structure. We hope our findings underscore the need for approaches that develop a more deeply grounded understanding of the physical world.

With the rapid growth of e-commerce, exploring general representations rather than task-specific ones has attracted increasing attention. Although recent multimodal large language models (MLLMs) have driven significant progress in product understanding, they are typically employed as feature extractors that implicitly encode product information into global embeddings, thereby limiting their ability to capture fine-grained attributes. Therefore, we argue that leveraging the reasoning capabilities of MLLMs to explicitly model fine-grained product attributes holds significant potential. Nevertheless, achieving this goal remains non-trivial due to several key challenges: (i) long-context reasoning tends to dilute the model's attention to salient information in the raw input; (ii) supervised fine-tuning (SFT) primarily encourages rigid imitation, limiting the exploration of effective reasoning strategies; and (iii) fine-grained details are progressively attenuated during forward propagation. To address these issues, we propose MOON3.0, the first reasoning-aware MLLM-based model for product representation learning. Our method (1) employs a multi-head modality fusion module to adaptively integrate raw signals; (2) incorporates a joint contrastive and reinforcement learning framework to autonomously explore more effective reasoning strategies; and (3) introduces a fine-grained residual enhancement module to progressively preserve local details throughout the network. Additionally, we release a large-scale multimodal e-commerce benchmark MBE3.0. Experimentally, our model demonstrates state-of-the-art zero-shot performance across various downstream tasks on both our benchmark and public datasets.

Multimodal large-language models (MLLMs) often experience degraded safety alignment when harmful queries exploit cross-modal interactions. Models aligned on text alone show a higher rate of successful attacks when extended to two or more modalities. We propose a simple conditional decoding strategy, CASA (Classification Augmented with Safety Attention) that uses internal representations of MLLMs to predict a binary safety token before response generation. The classification is guided by a safety attention mechanism that computes attention between a dedicated query prompt and the input to scale the classification logits, sharpening the model's ability to detect malicious queries. Our design ensures robust safety alignment without relying on any external classifier or auxiliary head, and without modality-specific safety fine-tuning. Unlike prior safety-decoding methods that attach an external classifier or moderation head, or rescale every generated token with a reward model, CASA makes a single, internal, attention-gated safety decision over the shared multimodal representation, which is precisely what lets safety learned from text-only training transfer to image and audio inputs. Compared against 4 baseline models across 7 attack types spanning text, image, and audio modalities, evaluated on benchmarks such as MM-SafetyBench, JailbreakV-28k, and adversarial audio tests, CASA lowers the average attack success rate by >90% relative when paired with a standard safety prompt (applied to all baselines for fair comparison), and by 54--61% using its internal safety mechanism alone. Our empirical evaluations also show that CASA~maintains strong utility on benign inputs, validated through both automated and human evaluations (via 13 trained annotators). Together, these results highlight CASA as a simple and generalizable framework to improve multimodal LLM safety.

Final-answer video QA can show whether a model predicts the right number, but not which instances it counted, when the supporting evidence occurs, or why it failed. We diagnose long-video quantitative reasoning in multimodal large language models (MLLMs) through three coupled abilities: enumerating query-relevant instances, temporally grounding supporting evidence, and aggregating the evidence into counts. To support this analysis, we build EC-Bench, an evidence-annotated evaluation suite with 152 untrimmed videos longer than 30 minutes, 1,699 open-ended queries across six reasoning categories, and human-verified evidence spans. We evaluate 22 open-source and proprietary MLLMs using timestamped visual frames and transcripts. The best average scores reach only 29.98% Enumeration F1 and 23.74% Counting accuracy, compared with human performance of 78.57% and 82.97%, respectively. Our analyses show that counting errors are rarely isolated arithmetic mistakes: Enumeration F1 is strongly associated with Counting accuracy, temporal grounding quality is associated with lower counting error, and Counting accuracy drops as supporting evidence becomes more distributed. These findings recast long-video counting as evidence retrieval, temporal grounding, deduplication, and aggregation across the video, rather than simple numerical prediction.

Although multimodal large language models (MLLMs) are increasingly deployed in real-world applications, their instruction-following behavior leaves them vulnerable to prompt injection attacks. Existing prompt injection methods predominantly rely on textual prompts or perceptible visual prompts that are observable by human users. In this work, we study imperceptible visual prompt injection against powerful closed-source MLLMs, where adversarial instructions are embedded in the visual modality. Our method adaptively embeds the malicious prompt into the input image via a bounded text overlay to provide semantic guidance. Meanwhile, the imperceptible visual perturbation is iteratively optimized to align the feature representation of the attacked image with those of the malicious visual and textual targets at both coarse- and fine-grained levels. Specifically, the visual target is instantiated as a text-rendered image and progressively refined during optimization to more faithfully represent the desired semantics and improve transferability. Extensive experiments on two multimodal understanding tasks across multiple closed-source MLLMs demonstrate the superior performance of our approach compared to existing methods.

Imagine advanced humanoid robots, powered by multimodal large language models (MLLMs), coordinating missions across industries like warehouse logistics, manufacturing, and safety rescue. While individual robots show local autonomy, realistic tasks demand coordination among multiple agents sharing vast streams of sensor data. Communication is indispensable, yet transmitting comprehensive data can overwhelm networks, especially when a system-level orchestrator or cloud-based MLLM fuses multimodal inputs for route planning or anomaly detection. These tasks are often initiated by high-level natural language instructions. This intent serves as a filter for resource optimization: by understanding the goal via MLLMs, the system can selectively activate relevant sensing modalities, dynamically allocate bandwidth, and determine computation placement. Thus, R2X is fundamentally an intent-to-resource orchestration problem where sensing, communication, and computation are jointly optimized to maximize task-level success under resource constraints. This survey examines how integrated design paves the way for multi-robot coordination under MLLM guidance. We review state-of-the-art sensing modalities, communication strategies, and computing approaches, highlighting how reasoning is split between on-device models and powerful edge/cloud servers. We present four end-to-end demonstrations (sense -> communicate -> compute -> act): (i) digital-twin warehouse navigation with predictive link context, (ii) mobility-driven proactive MCS control, (iii) a FollowMe robot with a semantic-sensing switch, and (iv) real-hardware open-vocabulary trash sorting via edge-assisted MLLM grounding. We emphasize system-level metrics -- payload, latency, and success -- to show why R2X orchestration outperforms purely on-device baselines.

Long video understanding is a key challenge that plagues the advancement of Multimodal Large language Models (MLLMs). In this paper, we study this problem from the perspective of visual memory mechanism, and proposed a novel and training-free approach, termed Flexible Memory (FlexMem). In principle, FlexMem aims to mimic human behavior of video watching, i.e., continually watching video content and recalling the most relevant memory fragments to answer the question. In this way, FlexMem can help MLLMs achieve video understanding of infinite lengths, unlike previous methods that process all video information at once and have input upper-limit. Concretely, FlexMem first consider the visual KV caches as the memory sources, and realize the effective memory transfer and writing via a dual-pathway compression design. Afterwards, FlexMem also explores different memory reading strategies for the diverse video understanding tasks, including the popular streaming one. To validate FlexMem, we apply it to two popular video-MLLMs, and conduct extensive experiments on five long video and one streaming video task. The experimental results show that on a single 3090 GPU, our FlexMem can achieve obvious improvements than existing efficient video understanding methods and process more than 1k frames, which also helps the base MLLMs achieve comparable or even better performance than SOTA MLLMs on some benchmarks, e.g. , GPT-4o and Gemini-1.5 Pro.