Tag
论文提出 AC-JEPA（Anchored Cross-Modal JEPA），一种可与 SFT 兼容的即插即用模块，通过在双向跨模态预测掩码中间表示来提升视觉到符号序列的翻译能力。在光学化学结构识别（OCSR）任务上，AC-JEPA 在 WildMol Simple 上达到 76.07% 准确率（基线 SFT 为 67.40%），并在使用含噪声伪标签数据自训练时提升至 80.23%，而直接 SFT 反而性能下降。
AVC-Reward 提出一个事件中心的音视频一致性评估框架，通过可训练奖励模型在全局和谐、时间同步和属性对齐三个层面评估复杂多事件场景，并在与人类判断对齐上达到最先进水平，还可作为在线强化学习信号提升联合生成模型的一致性。
提出一种通过蒸馏最优传输在仅有100-500个配对样本下对齐两个独立预训练单模态编码器的方法，在Flickr30k、MS-COCO和ChEBI-20上以相同锚点预算超越现有方法。
提出 Adaptive Reward Routing，在联合音频-视频扩散模型的前向过程强化学习（DiffusionNFT）中动态调整更新位置与奖励权重：利用双向跨注意力响应定位跨模态影响，并以偏好先验加残差修正的方式协调多个竞争奖励，从而提升模态质量、语义一致性与音视频同步。
MME-Safety is a rigorously verified benchmark for evaluating the safety of Multimodal Large Language Models, featuring a four-dimensional annotation schema and a hierarchical framework to assess risk scenarios, harm severity, and modality-specific stealth levels.
This paper proposes a multimodal anomaly detection framework for fault detection in mechanical systems using self-supervised cross-modal reconstruction and adaptive thresholding to improve robustness under distribution shifts.
Omni-Streaming Thinking improves streaming omni-modal reasoning by deferring claims until cross-modal verification, reducing premature commitment and auditory hallucinations.
OmniHallu introduces a unified hallucination detection framework for multimodal large language models, covering comprehension and generation tasks across image, video, and audio modalities, with a benchmark and multi-agent architecture.
The paper introduces Tri-PvP, a benchmark that exposes visual bias and asymmetric evidence-form preferences in omni-modal large language models, revealing deep-seated modality biases that are linearly decodable from early layers and resistant to surface mitigation.
This paper introduces HalluPrism, a behavioral diagnostic method for multimodal large language models that uses visual perturbation probes to identify hallucination failure modes, improving failure-family classification over confidence-only methods.
VA-Judger is the first reward model for joint video-audio generation that uses human preference feedback to evaluate holistic quality, including a dataset and benchmark, and demonstrates significant improvements in human preference rates when applied to models like LTX-2.
This paper introduces The Unwritten Benchmark, a new challenge to evaluate abstract perceptual reasoning in multimodal AI models, revealing a significant performance gap between humans and current models like GPT-4o and Gemini 2.5-Pro.
This paper introduces HC-RAG, a hierarchical cross-modal retrieval-augmented generation framework for evidence-centric financial question answering over 10-K filings, along with a new benchmark Multi-Doc-2025. It outperforms RAPTOR and GraphRAG on financial QA benchmarks, especially for long-document and table-related queries.
This paper introduces a method to predict middle-layer attention in multimodal LLMs to prune visual tokens efficiently, using question-contrastive teacher selection and cross-modal attention distillation.
This paper introduces ConCor-1, a grounding model that treats vision-language grounding as bidirectional concept correspondence, jointly recovering text spans, image segments, and cross-modal matches without prespecified phrases. It unifies phrase grounding, referring expression grounding, and open-vocabulary detection, achieving significant F1 improvements on long-caption and zero-shot LVIS benchmarks.
Introduces C3PO, a benchmark of 3,404 samples for evaluating cross-modal composition and counterfactual reasoning in multimodal LLMs. It finds modality dominance causes most failures, with even the best model (Gemini-3.1-Pro) far below human accuracy.
This paper introduces Kontrast, a framework for automatically detecting knowledge inconsistencies across Wikipedia text, tables, and Wikidata knowledge graphs using Text-to-SPARQL and LLM reasoning.
Proposes a unified post-hoc detection framework for copyright infringement in AI models, using conditional sensitivity and differential privacy to measure memorization across modalities.
OmniVAE is a jointly trained audio-video VAE that uses segment-level contrastive learning and feature distillation to align latent spaces, improving joint generation quality and synchronization in text-to-audio-video generation.
This paper proposes UMMT, a token-level cross-modal transformer with contrastive multi-task learning for breast cancer subtype classification and survival prediction, achieving state-of-the-art results on METABRIC and TCGA-BRCA datasets.