多模态大模型

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 License at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3. The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3.

Current automated pipelines for audio-visual Question Answering (QA) generally adopt a ``video-caption-QA'' paradigm. However, these methods typically segment videos into short clips and generate separate descriptions for audio and visual modalities. This decoupled processing severs inherent associations between sounds and their visual sources, while independent clip processing often causes inconsistent descriptions of the same entity across segments. Furthermore, coupling long-text comprehension and QA synthesis into a single step often restricts models to localized events, yielding questions lacking long-term temporal connections and deep cross-modal reasoning. To address these issues, we propose an automated data engine featuring two mechanisms: (1) \textbf{Entity-Anchored Video Scripting} transforms videos into structured scripts, comprising summaries, main entity lists, and segment-wise audio-visual descriptions. The entity list serves as a global prior to ensure cross-segment referential consistency and reconstruct audio-visual associations. (2) \textbf{Clue-Guided QA Generation} prompts models to first mine cross-segment, multimodal clues from the script, and subsequently generate QA pairs based on these high-value clues. Leveraging this pipeline, we construct the instruction-tuning dataset \textbf{OmniVideo-100K} and a human-verified test set, \textbf{OmniVideo-Test}. Fine-tuning VITA-1.5, Qwen2.5-Omni-7B and Qwen3-Omni-30B on OmniVideo-100K yields performance gains of up to 20.59% on OmniVideo-Test, demonstrating strong generalization (up to 12.64% improvements) across established benchmarks like Daily-Omni and JointAVBench.

Evaluation of socially unsafe content in spoken dialogues remains text-centric, missing prosody and transcription failures. We present LALM-as-a-Judge, which includes an open benchmark of 24,000 multi-turn spoken dialogues with one localized unsafe turn, generated out of 8 socially unsafe categories and 5 severity levels. We evaluate 6 large audio-language models (LALMs) as judges, open and closed-source, in text-only, audio-only, and multimodal setups by their sensitivity, severity-order specificity, and turn-position bias for socially harmful content in the dialogue. Results show that audio contributes non-lexical evidence beyond transcript semantics and that multimodal gains are not universal but can be text-anchored, balanced, conservative, and interfering, which we link to the audio pathway bottlenecks and fusion limits. We position the benchmark as diagnostic and derive practitioner guidance for model, modality, and prompts choices.

Although Multimodal Large Language Models (MLLMs) demonstrate strong omni-modal perception, their ability to forecast future events from audio-visual cues remains largely unexplored, as existing benchmarks focus mainly on retrospective understanding. To bridge this gap, we introduce FutureOmni, the first benchmark designed to evaluate omni-modal future forecasting from audio-visual environments. The evaluated models are required to perform cross-modal causal and temporal reasoning, as well as effectively leverage internal knowledge to predict future events. FutureOmni is constructed via a scalable LLM-assisted, human-in-the-loop pipeline and contains 919 videos and 1,034 multiple-choice QA pairs across 8 primary domains. Evaluations on 13 omni-modal and 7 video-only models show that current systems struggle with audio-visual future prediction, particularly in speech-heavy scenarios, with the best accuracy of 64.8% achieved by Gemini 3 Flash. To mitigate this limitation, we curate a 7K-sample instruction-tuning dataset and propose an Omni-Modal Future Forecasting (OFF) training strategy. Evaluations on FutureOmni and popular audio-visual and video-only benchmarks demonstrate that OFF enhances future forecasting and generalization. We publicly release all code (https://github.com/OpenMOSS/FutureOmni) and datasets (https://huggingface.co/datasets/OpenMOSS-Team/FutureOmni).

AudioLLMs enable speech recognition conditioned on textual prompts such as domain descriptions or entity lists. However, it remains unclear whether these models genuinely utilise such context or rely on parametric knowledge learned during pretraining. Existing benchmarks cannot answer this question because they evaluate transcription under fixed prompting conditions and rarely include explicit contextual inputs. We introduce IndicContextEval, a 56-hour multilingual benchmark of natural speech from 555 speakers across 8 Indian languages and 23 professional domains. We design a 7-level prompting framework that progressively introduces contextual signals, including metadata, natural-language descriptions, entity lists in English and native script, and adversarial prompts with incorrect entities. Evaluating five models reveals substantial differences in context utilisation behaviour, highlighting the need for explicit evaluation of contextual grounding in AudioLLMs.

While Audio Large Language Models (Audio LLMs) excel at multimodal understanding, they suffer from text dominance, a bias where models blindly favor text over acoustic evidence, causing hallucinations. However, the internal mechanisms underlying how these models behave when audio and textual inputs contradict each other remain unexplored. In this work, we present the first mechanistic analysis of this phenomenon by tracing the propagation of internal representations across layers. Our investigation reveals three key findings: (i) text dominance is systematically and empirically across models; (ii) while text and audio rely on functionally distinct pathways, they ultimately converge into a shared semantic space in late layers; and (iii) the text pathway does not erase audio information, but rather actively suppresses intact audio representations. Building on these insights, we leverage back-patching, a training-free intervention that routes late-layer audio activations back into earlier layers. This amplifies the audio representations, enabling them to overcome textual suppression. Our evaluation shows that back-patching consistently reduces text dominance, paving the way for mechanistic multimodal alignment under conflict.

Large audio language models (LALMs) have shown impressive capabilities on diverse audio understanding tasks, ranging from speech transcription to music analysis. However, because LALMs are typically trained to produce text-aligned responses, their hidden states are progressively shaped for text generation rather than for preserving acoustic information. As a result, the diverse acoustic content that audio carries, such as phonetic detail, prosody, sound events, affect, and pitch, is lost along the way and difficult to leverage in the response. We introduce Continuous Audio Thinking (CoAT), a framework that equips audio language models with a continuous latent workspace for organizing acoustic information prior to response generation, grounded by distillation from audio experts. Within the thinking space, the model can utilize the rich acoustic information provided by expert distillation when generating its response. Furthermore, the proposed continuous thinking block can be processed in a single prefill, so CoAT does not require additional autoregressive decoding cost over the baseline. Across three LALMs, Qwen2-Audio, Qwen2.5-Omni-7B, and Audio Flamingo~3, performance gains on a broad benchmark suite spanning audio reasoning, audio understanding, music classification, speech emotion, and speech transcription demonstrate the effectiveness of CoAT. Further analysis confirms that the auxiliary supervision propagates from the thinking positions to the model's textual responses.

Speech-based Alzheimer's Disease (AD) detection is constrained by scarce pathological speech data. To address this, we propose CoSTA, a Text-to-Speech (TTS)-based data augmentation framework. Specifically, we first develop two Cognitive-State-Conditioned (CS-Cond) TTS models by adapting CosyVoice2 and F5-TTS to synthesize speech with distinct AD and Healthy Control characteristics. Furthermore, by constructing a transcript pool comprising Manual Transcripts (MT) and 36 Automatic Speech Recognition (ASR) transcripts, we investigate the impact of text sources on TTS-based augmentation. We also perform augmentation-factor analysis and test-time augmentation. Experiments on the ADReSS dataset show that CS-Cond TTS significantly improves synthetic speech utility, and ASR-driven augmentation frequently outperforms MT-driven augmentation. Finally, CoSTA yields a 4.16% gain over the baseline, achieving an audio-only accuracy of 85.83% on the ADReSS test set and outperforming prior methods.

Traditional RGB-based speech generation faces Temporal Granularity Mismatch since fixed camera exposure times inevitably blur the high-frequency articulatory transients essential for rendering emotional speech. To break this ceiling, we propose EventSpeech as a novel text-conditioned framework pioneering the use of neuromorphic events for expressive speech generation, since these microsecond-precise events naturally align with acoustic waveform dynamics. Our architecture integrates a dedicated Event Encoder to model sparse neuromorphic events alongside a multi-scale Audio Encoder featuring a Hierarchical Wavelet Contextualizer (HWC). A bidirectional alignment mechanism seamlessly synchronizes linguistic content and visual dynamics with dense acoustic features. Furthermore, we construct EVT-SPK as the first benchmark comprising large-scale synthetic data and real-world recordings from specialized neuromorphic hardware. Extensive evaluations demonstrate that EventSpeech significantly outperforms current baselines by preserving fine-grained emotions and resisting motion blur to establish a new paradigm for multimodal speech generation. Code and demo are available at https://xrfang-0102.github.io/EventSpeechWeb/.

Large Audio Language Models (LALMs) integrate audio encoders with pretrained Large Language Models to perform complex multimodal reasoning tasks. While these models can generate Chain-of-Thought (CoT) explanations, the faithfulness of these reasoning chains remains unclear. In this work, we propose a systematic framework to evaluate CoT faithfulness in LALMs with respect to both the input audio and the final model prediction. We define three criteria for audio faithfulness: hallucination-free, holistic, and attentive listening. We also introduce a benchmark based on both audio and CoT interventions to assess faithfulness\footnote{The benchmarking interface and evaluation results are available at https://poonehmousavi.github.io/faithfulness/. Experiments on Audio Flamingo 3 and Qwen2.5-Omni suggest a potential multimodal disconnect: reasoning often aligns with the final prediction but is not always strongly grounded in the audio and can be vulnerable to hallucinations or adversarial perturbations.

Speech-aware large language models (LLMs) can accept speech inputs, yet their training objectives largely emphasize linguistic content or specific fields such as emotions or the speaker's gender, leaving it unclear whether they encode speaker identity. First, we propose a model-agnostic scoring protocol that produces continuous verification scores for both API-only and open-weight models, using confidence scores or log-likelihood ratios from the Yes/No token probabilities. Using this protocol, we benchmark recent speech-aware LLMs and observe weak speaker discrimination (EERs above 20% on VoxCeleb1). Second, we introduce a lightweight augmentation that equips an LLM with ASV capability by injecting frozen ECAPA-TDNN speaker embeddings through a learned projection and training only LoRA adapters. On TinyLLaMA-1.1B, the resulting ECAPA-LLM achieves 1.03% EER on VoxCeleb1-E, approaching a dedicated speaker verification system while preserving a natural-language interface.

Large Audio Language Models (LALMs) are increasingly capable of reasoning over audio, yet existing benchmarks offer limited coverage of reasoning in polyphonic audio, where multiple sound events co-occur and induce compositional structure. To address this gap, we introduce PolyBench, a benchmark designed to evaluate compositional reasoning in polyphonic audio, comprising five evaluation subsets that cover counting, classification, detection, concurrency, and duration estimation, all of which require reasoning over multiple concurrent events and their relations. Our evaluation of state-of-the-art LALMs reveals consistent performance degradation in polyphonic settings, indicating a fundamental bottleneck in current LALMs.

Automatic Speech Recognition (ASR) often degrades in real-world noisy environments, making noise robustness essential for deployment. Supervised noise-augmented fine-tuning is a common remedy, but it can introduce a robustness-clean trade-off and overfit to specific corruptions, degrading recognition in clean conditions. We propose DASH, a self-distillation framework that improves robustness by learning clean--noisy consistency from paired views. DASH distills hidden representations from multiple encoder layers to capture features from low-level acoustics to high-level semantics, and stabilizes training by minimizing KL divergence between prototype assignment distributions of clean and noisy views. Experiments on LibriSpeech show that DASH consistently improves recognition under diverse noisy conditions while preserving clean accuracy, achieved by a label-free pre-training stage with minimal additional overhead (about 4% of fine-tuning time) beyond standard fine-tuning.

This study presents a comparative analysis between the speaker embeddings of speech foundation models and human subjective perception of speaker similarity. Human listeners have the ability to judge speaker similarity on a continuous scale discerning how similar two voices are. In contrast, speech foundation models embed speaker characteristics into numerical representation. However, a question remains: does the numerical distance between speaker embeddings in these models truly align with the similarity perceived by humans? To address this, we conduct a comprehensive investigation using more than 40 models to compare model-derived distances with human-perceived similarity scores. Furthermore, we identify which factors in model configuration contribute most to a speaker embedding that mirrors human perception. Our findings provide insights for the development of more perceptually grounded speech foundation models.

Recent advancements in Multimodal Large Language Models (MLLMs) have achieved remarkable success in vision-language tasks, yet the quadratic computational complexity arising from the vast number of visual tokens incurs significant memory and latency bottlenecks. While visual token reduction (VTR) strategies have been explored to mitigate this burden, existing methods overlook the positional and attentional consistency between the full and reduced sequences, resulting in a distorted representation. To this end, we propose RESTORE, a novel VTR framework that rectifies the positional and attentional distortions while maintaining efficiency. Specifically, we present a simple yet effective calibration method that restores lost visual attention by augmenting attention weights based on relative distances. We also introduce a distinctive anchor selection for token merging to mitigate information loss during feature averaging. Experimental results on multiple benchmarks demonstrate that our method consistently improves the accuracy of various reduction methods, achieving state-of-the-art performance while maintaining computational efficiency. Project page is available at https://cvlab.yonsei.ac.kr/projects/RESTORE

On-policy self-distillation (OPSD) trains a model on its own rollouts and uses a frozen copy to provide dense token-level targets conditioned on a reference target. This works well for LLM reasoning, but a direct extension to multimodal large language models (MLLMs) can create a shortcut: the privileged target may guide tokens mainly based on the text reference target rather than the image. We propose ViGOS, a visually grounded OPSD framework for MLLM post-training. The student first writes a visual description and then reasons toward the final answer. For valid rollouts, an image-only perception teacher supervises the description, while a privileged reasoning teacher supervises the reasoning and final answer on the same student prefix. A reference teacher is used only for invalid rollouts to recover the output format. Across general vision-language, expert reasoning, visual math, spatial grounding, and visual-language-prior benchmarks, ViGOS keeps the main benefits of OPSD and improves image-grounded behavior in shortcut-prone settings.

Multimodal deception detection is critical for identifying fraudulent intentions, yet existing approaches predominantly rely on end to end black--box paradigms. These methods suffer from a severe lack of interpretability failing to provide transparent reasoning trajectories and struggling to explicitly capture the subtle, cross modal inconsistencies inherent in deceptive behaviors. To transcend these limitations, we propose ThinkDeception, a novel and interpretable multimodal deception detection framework. As a pioneering effort, it introduces Multimodal Large Language Models (MLLMs) into this domain, transforming deception detection from a traditional binary classification task into an explicit cognitive reasoning process. Facilitated by the first meticulously annotated step--by--step multimodal Chain of Thought (CoT) dataset, we develop a foundational model, ThinkDeception Base, empirically validating the critical role of modal inconsistency in decoding deception. Building upon this foundation, our core innovation lies in proposing Visual-Audio Consistency Group Relative Policy Optimization(VAC--GRPO) equipped with a progressive training strategy. Distinct from standard GRPO, we stratify the training data into four progressive difficulty tiers, guiding the model through a psychologically grounded easy--to--hard cognitive transition. By innovatively coupling this dynamic curriculum scheduler with a multi dimensional, process aware reward mechanism and a reflective learning paradigm, we significantly elevate the model's overall reasoning quality. Extensive experiments on mainstream benchmarks demonstrate that ThinkDeception establishes a new SOTA, significantly outperforming existing methods in both detection accuracy and rationale quality. Ultimately, this work successfully drives the field of deception detection toward interpretable, multimodal cognitive reasoning.

Multimodal Information Extraction (MIE)-covering tasks such as Multimodal Named Entity Recognition (MNER), Relation Extraction (MRE), and Event Extraction (MEE)-is essential for understanding multimedia content but remains constrained by severe data scarcity. Although data augmentation is a promising remedy, existing approaches are impeded by coarse cross-modal alignment and fragmented, task-specific designs that fail to exploit shared semantic knowledge. To overcome these limitations, we introduce Semantic Anchor-aligned Multimodal Augmentation (SAMA), a unified framework for generating high-fidelity, task-aware synthetic data. SAMA constructs structured semantic anchors from ground-truth labels to guide a Collaborative Multi-Experts Multimodal Large Language Model (CME-MLLM), which integrates a Universal Adapter for shared semantics with Task-Specific Adapters to produce diverse yet constraint-compliant textual samples. For image synthesis, SAMA employs an Anchor-Preserving Diffusion mechanism that uses anchor-weighted prompts and latent conditioning to maintain critical semantic anchors while diversifying visual contexts. To eliminate the need for manual verification, SAMA further introduces a Dual-Constraint Filtering module that selects synthetic samples based on both cross-modal consistency and anchor fidelity. Extensive experiments across benchmark datasets for MNER, MRE, and MEE demonstrate that SAMA consistently outperforms state-of-the-art augmentation baselines under both fully supervised and low-resource settings, underscoring its versatility, robustness, and effectiveness.

We introduce Qwen-RobotWorld, a language-conditioned video world model for embodied intelligence. With natural language as a unified action interface, it predicts physically grounded future visual trajectories from current observations across robotic manipulation, autonomous driving, indoor navigation, and human-to-robot transfer. This unified formulation provides three promising application directions: synthetic data generation for policy training augmentation, scalable virtual environments for policy evaluation, and language-guided planning signals for downstream robot control. This is achieved through a three-part design: a) Double-Stream MMDiT with MLLM Action Encoding, where a 60-layer double-stream diffusion transformer couples frozen Qwen2.5-VL semantics with video-VAE latents through layer-wise joint attention; b) Embodied World Knowledge (EWK), an 8.6M video-text corpus (200M+ frames) with action-language mapping over 20+ embodiments and 500+ action categories; and c) General+Expert Progressive Curriculum, a two-stage training strategy that first learns general visual priors and then injects embodied specialization under a shared language interface. Extensive results show strong competitiveness: ranks 1st overall on EWMBench and DreamGen Bench, outperforms all open-source models on WorldModelBench and PBench. Additional zero-shot analyses on RoboTwin-IF benchmark further support robust generalization and multi-view consistency.

A model can learn that the piano piece Für Elise is calm and reflective by listening to the audio or by reading a text description, but does it matter which route that knowledge took when it is later at risk of being forgotten? Forgetting research in multimodal models measures what knowledge is lost under adaptation, yet has not asked whether acquisition route affects how easily that knowledge is forgotten. We call this untested premise the Pathway-Invariant Assumption. Music understanding enables a clean test because a music clip and a canonical text description can be aligned to the same perceptual content, allowing the same knowledge unit to enter a model through listening or reading while the target remains fixed. Across multiple architecturally distinct audio-language models, we observe a consistent asymmetry: text-pathway knowledge is forgotten more than matched audio-pathway knowledge under identical adaptation pressure. To attribute this effect to route rather than confounds, we introduce the Paired Pathway Controlled Protocol (PPCP), a three-phase design that establishes matched pathway baselines, activates both pathways under symmetric supervision on the same knowledge pool, and applies identical forgetting pressure to both pathways. The gap is stable across models and gain-controlled analyses, persists when contradictory overwrite is replaced by correct-label cross-domain learning, remains under single-modality pressure, and is not removed by lightweight replay. Two independent routing-depth controls confirm that the effect is not explained by architectural depth, pointing to input representation as the dominant factor. Under PPCP, our results demonstrate that forgetting is highly route-dependent, establishing acquisition route as a new analytical dimension for forgetting research and multimodal system design.

Unified multimodal models (UMMs) interleave generated ''visual thoughts'' (VTs) with text reasoning to improve spatial tasks. This incurs roughly an order-of-magnitude inference cost from multi-step diffusion. We find this cost yields limited direct benefit. On ThinkMorph, removing or noising VTs barely changes accuracy across nine benchmarks. Once rendered, attention concentrates on the VT regardless of content. Yet a KL diagnostic shows that conditioning on a privileged VT trace shifts the model's completion distribution. This suggests the generation pathway encodes useful reasoning beyond the rendered pixels. Motivated by this gap, we propose Visual On-Policy Self-Distillation(Visual-OPSD). Teacher and student share identical weights but differ in context: the teacher sees privileged VTs while the student sees only the question. Token-level JSD distillation on on-policy student trajectories transfers the teacher's reasoning to a text-only student. Across nine benchmarks, Visual-OPSD improves over its generative teacher by $+3.40$pp with $14.3\times$ speedup (10.0s vs. 142.8s per sample) and outperforms same-scale VLMs by $+63.83$pp on VSP. A Gaussian-noise control ($+0.40$pp vs. $+10.28$pp for real VTs) and $58.4\%$ closure of the KL gap confirm that gains come from the semantic content of the generation pathway.

Multimodal emotion-cause pair extraction (MECPE) requires reliable pair confidence over candidate pairs. Existing pair scorers commonly use pair-level cross entropy over valid candidates, which treats links mostly independently. This leaves the relative confidence geometry among competing causes under-constrained, allowing gold pairs to stay close to hard negatives or rely on incidental non-gold context. We study this vulnerability as pair-confidence brittleness and propose RPCL (Robust Pair Confidence Learning), a training-only framework for pair-confidence learning. RPCL encourages pair confidence to be both discriminative and stable: gold pairs are separated from row-wise hard negatives through a confidence-difference margin constraint, and clean pair predictions are aligned with predictions from a corrupted view where non-gold contextual utterance representations are partially corrupted. The original clean pair scorer and decoding pipeline are used unchanged at inference time. On ECF, MECAD, and MEC4, RPCL improves the three-seed mean Pair F1 over a matched base model by 2.58 to 2.83 percentage points in the full text-audio-video setting, and improves mean Pair AUPRC on all three datasets. Diagnostic analysis further shows larger gold-negative confidence gaps and lower margin-violation severity. These results suggest that explicitly shaping pair confidence is an effective training strategy for MECPE.

Distributed large language model (LLM) inference frameworks connect isolated consumer-grade devices for large-scale model inference, substantially reducing hardware constraints. However, recent studies show that intermediate embeddings transmitted among participants can leak private prompts. As LLMs evolve into multimodal LLMs (MLLMs), this risk extends beyond text: image prompts contain rich visual and semantic information, making their intermediate embeddings highly privacy-sensitive. Yet, image-prompt leakage in distributed MLLM inference remains largely unexplored. In this paper, we investigate privacy risks to input images caused by intermediate embeddings in distributed MLLM frameworks. We first analyze the information flow from image pixels to intermediate representations. Since image and text embeddings are often intertwined across MLLM layers, we design an image embedding extraction algorithm as a prerequisite for reconstruction attacks, achieving 100% extraction accuracy across almost all MLLM layers in our experiments. Building on this, we develop two passive black-box image reconstruction attacks, MPAA and IEDA, reflecting realistic threats from normal participants with limited knowledge and capability. MPAA performs fine-grained pixel-level reconstruction via patch-wise information extraction and assembly, while IEDA performs coarse-grained semantic reconstruction through embedding-guided diffusion generation. We evaluate our attacks on four representative MLLM families: Gemma 3, Phi 4 Multimodal, Qwen 2.5 VL, and Llama 4 Scout. Results show consistently superior reconstruction performance in various settings. We further analyze the effects of MoE architecture, image preprocessing, model size, and text-image dependency on attack performance. To our knowledge, this is the first study of image reconstruction attacks on MLLMs.

Multimodal large language models (MLLMs) are increasingly being evaluated for medical applications, where computational constraints often make prompting strategies the only practical alternative to fine-tuning. Such strategies are generally assumed to support diagnostic reasoning, yet their potential failure modes in medical MLLMs remain poorly characterized. We analyze FundusExpert-1B, an open-source ophthalmology MLLM, on a hemorrhage versus drusen discrimination task using the public BRSET dataset, adopted here as a controlled testbed for our analysis. (i) A controlled probe with artificially injected markers confirms that the model retains coarse, region-level spatial grounding. (ii) Compared with zero-shot inference, one-shot textual prompts bias predictions toward the prompted finding. (iii) When an overlaid lesion contour is paired with an inconsistent textual claim, the textual prompt overrides the correct visual cue: overall accuracy drops from 75% to 46% relative to the visual-only condition, and Chain-of-Thought (CoT) reasoning is associated with further degradation rather than self-correction. Although limited to a single model and dataset, our findings suggest that prompting strategies alone may be insufficient for the safe clinical deployment of medical MLLMs.

Intelligent landslide hazard interpretation is critical for disaster prevention, yet current paradigms struggle to simultaneously extract visual features and high-level geoscientific semantics, while general-purpose vision-language models (VLMs) suffer from perceptual limitations and domain hallucinations in complex geological scenarios. To address these challenges, we propose an instruction-driven agentic framework comprising three components. First, LandslideBench, a multimodal fine-grained dataset with seven subtype labels, high-resolution imagery, pixel-level masks, and high-quality textual descriptions, is constructed via multi-VLM cross-validation and interactive annotation. Then, LandslideVLM, a landslide-oriented VLM, is fine-tuned via LoRA on LandslideBench to enhance geological semantic understanding. Finally, LandslideAgent, a domain rule-enhanced agent taking LandslideVLM as its cognitive backbone, employs a dual-rule controller incorporating structured report metadata constraints and cross-validation identification constraints to regulate automated tool invocation. Experiments demonstrate that LandslideBench provides effective baselines across five mainstream models on fine-grained classification and semantic segmentation. LandslideVLM achieves accuracy improvements of 10.96%, 32.87%, and 15.91% on landslide discrimination, fine-grained classification, and semantic description quality, respectively. LandslideAgent further enables autonomous multi-source spatial data inference, realizing full-process intelligence for landslide identification and analysis.

Reinforcement learning has improved the reasoning ability of large language models, but applying outcome-only rewards to video multimodal large language models (Video-MLLMs) provides limited guidance on which visual evidence should support the answer. Inspired by multisensory integration, where consistent cues can enhance the salience and reliability of perceptual estimates, we introduce Consensus Frame GRPO (CF-GRPO), a temporal-annotation-free process-level reward framework for evidence-aware video reasoning. CF-GRPO constructs a consensus frame prior from intrinsic video cues, including temporal coverage, scene-transition cues, and query-conditioned visual relevance. It then computes a model-side frame-use score from visual and response representations and optimizes their agreement through the Consensus Frame Reward (CFR). With salience-aware sparse aggregation and distribution sharpening, CFR provides a high-contrast reward signal without requiring human temporal annotations. Experiments show that VideoCFR achieves competitive performance across complex video reasoning benchmarks and improves several metrics over representative Video-MLLM and RL baselines, while the consensus prior provides an interpretable view of the evidence frames emphasized during training. The implementation is available at https://github.com/1Pansy/VideoCFR.

This paper presents factorized linear projection (FLiP) models for understanding pretrained sentence embedding spaces. We train FLiP models to recover the lexical content from multilingual (LaBSE), multimodal (SONAR) and API-based (Gemini) sentence embedding spaces in several high- and mid-resource languages. We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines. Using this as a diagnostic tool, we uncover the modality and language biases across the selected sentence encoders and provide practitioners with intrinsic insights about the encoders without relying on conventional downstream evaluation tasks. Our implementation is public https://github.com/BUTSpeechFIT/FLiP.

Deploying multimodal foundation models as closed-loop policies increasingly requires conditioning actions on observations that are no longer visible. However, existing benchmarks either expose the full state, conflate hidden-state reconstruction with other agent skills, or test recall only after an episode has ended. We introduce RNG-Bench (Reconstructive Non-Markov Games), a benchmark suite designed to isolate a base model's ability to reconstruct past observations and act on them during multi-step interaction. RNG-Bench includes two complementary games: Matching Pairs, where card identities briefly revealed at specific locations must later be recalled, and 3D Maze, where egocentric views must be integrated into a spatial map. Both games are evaluated under a unified harness with three controlled difficulty axes: grid size, visual pattern, and observation modality. The benchmark further introduces a head-to-head duel protocol to control for instance-level variance and a Memory Gap metric that disentangles forgetting from poor action selection. The hardest configurations require contexts of roughly 128K tokens and 350 image inputs per episode, and remain far from saturated by frontier MLLMs. Memory Gap analysis shows that most residual errors stem from forgetting earlier observations rather than from suboptimal decision making. Finally, fine-tuning Qwen3.5-9B on optimal-policy rollouts and filtered model demonstrations improves performance on RNG-Bench and transfers to existing benchmarks without degrading general multimodal capability.

In today's media-driven world, the exponential growth of video content across domains such as surveillance, education, and entertainment has made retrieving semantically relevant videos via natural language queries increasingly critical. Early video retrieval systems relied on handcrafted features or shallow cross-modal mappings, limiting their ability to capture complex semantics and temporal dynamics. While large-scale vision-language models have improved cross-modal alignment, challenges remain in modeling fine-grained temporal dependencies and nuanced linguistic structures. In this paper, we introduce DREAM: Dual-path Representation Enhancement and Alignment Model, a novel multimodal framework that addresses these limitations through enhanced visual and textual encoding. DREAM incorporates a hybrid language modeling strategy that combines masked and permuted language modeling objectives to capture both local and global linguistic semantics. On the visual side, we design a hierarchical vision encoder with cascaded group attention, which integrates spatial and temporal information through multi-stage token interaction and coarse-to-fine attention refinement. We validate DREAM through comprehensive evaluations on the widely-used MSRVTT, MSVD and LSMDC benchmark datasets, where it achieves new state-of-the-art R1 scores of 49.4%, 49.7% and 27.3%, respectively. Qualitative analyses further show the model's ability to maintain coherent attention across frames and align complex queries with dynamic video content. These findings underscore the effectiveness of hierarchical attention and dual-objective textual modeling in enabling robust, context-aware video retrieval, and pave the way for future research in advancing cross-modal representation learning.

Image retrieval in crowded scenes is particularly challenging due to the salience bias of conventional visual encoders, which tend to focus on dominant objects while neglecting low-attention regions that are often crucial for fine-grained retrieval. We propose LARE (Low-Attention Region Encoding), a framework that explicitly models these overlooked regions. LARE adopts a dual-encoding strategy that encodes low-attention regions of an image and the full image in parallel, leading to more diverse and informative image embeddings. To evaluate image retrieval performance in challenging crowded scenes, we introduce Dense-Set, a challenging subset derived from COCO and Flickr30K. In this subset, images are re-captioned to provide richer descriptions of low-attention or previously overlooked regions. This dataset highlights the limitations of existing retrieval models and enables a more rigorous evaluation under densely crowded scene conditions. Experimental results demonstrate that the proposed framework improves retrieval performance by preserving subtle, non-dominant visual cues within the shared latent space.