mllm

Reviso is an agentic video captioning framework that uses hierarchical Context-Focus-Points planning and active perception to prioritize key visual evidence, trained with GRPO and specialized rewards, outperforming open-source baselines on four video captioning benchmarks.

PhotoGuider 提出一种 measure-then-frame 方法，利用共享的构图感知 token 进行 360° 全景图的观看方向选择与取景，并通过微调 MLLM 生成以测量结果为条件的构图解释，在两项 2D 裁剪基准上取得最高 mIoU，并在 120 人盲测中以 53.2% 的偏好率排名第一。

本文提出一个物理接地的多模态智能体框架，用于磁约束聚变装置中性束注入离子源的非平稳自适应调控，在EXL-50U平台455次放电的时序回放中验证了符号化物理量优于波形图像，并在监督运行下实现86.80%的束提取率。

本文研究多模态大模型中的检索头机制如何随输入模态扩展，发现跨模态共享的检索头虽不足2%，却对文档理解、空间感知等至关重要，并通过排序目标优化这些注意力头，将Qwen3-VL-8B在MMDocIR上的页面Recall@1提升了6.03个点。

SOMA is a new benchmark for fine-grained egocentric body-motion understanding in multimodal models, with 1,632 human-verified questions evaluating 37 MLLMs. Humans reach 68.5% accuracy while the best model, Qwen3.6-27B, achieves only 50.0% zero-shot, revealing significant gaps in body-motion reasoning.

OpenVLHarness 是一个多模态推理框架，通过能力层、持久化多模态记忆与编码接口实现异构工具组合，在 22 个基准上相较无工具基线平均提升 9.7-12.6 分，同时显著降低 token 成本。

The paper introduces SABER, a controlled benchmark exposing widespread attribute-object binding failures in multimodal retrievers, and shows that fine-tuning an MLLM-based embedder solely on synthetic attribute-binding text achieves SOTA on SABER while improving cross-modal and image-image retrieval without image supervision.

WinARDet 提出压缩窗口自回归（WinAR）机制，将长答案序列划分为受限自回归窗口并用压缩潜变量表示历史，缓解密集视觉语言定位中的注意力抢占与框重复问题，在 RSDense 上将 RexOmni F1 从 32.16 提升至 39.36。

该论文在统一可控的基准框架下系统评估了多模态嵌入学习中的各类对比训练策略，发现大多数策略相比标准对齐基线仅有边际收益甚至带来负面影响，并提出结合难负样本挖掘、假负样本缓解与难批次构建的协同训练方案，在MMEB上以Qwen2-VL-2B平均提升6.0%。

UniData 提出一个通用多模态指令生成流水线，将简单用户需求扩展为多轮多模态指令，并构建了涵盖九种模态的 20,000 条 UniDataset，以提升多模态大模型的数据质量与能力。

This paper introduces Humanity's Sixth Sense (HSS), a benchmark for intuitive visual reasoning in multimodal large language models, showing humans reach 93.1% accuracy while the strongest model, GPT-6-astra, achieves only 53.6%, highlighting a significant capability gap in implicit visual understanding.

OmniCapBench 提出一个深度结构化的音视频字幕评估框架，将预测目标从自由文本转为可验证的原子评估单元（实体引用、视觉镜头、音频事件），在 786 个密集标注视频上区分多模态大模型的感知错误，揭示前沿模型长时程音视觉推理能力薄弱。

TerraVis 提出了一个基于 MLLM 工作流的框架，用于评估文生图模型生成图像的世界一致性（如畸形结构、物理上不可能的交互、空间关系不一致等），并定义了涵盖对象、交互和场景层级的 18 类一致性违规分类体系。实验表明该框架与人类判断相关性最强，且常规指标表现优秀的模型仍可能存在大量世界一致性缺陷。

The paper proposes a causal decomposition framework that isolates whether MLLM failures on compositional tasks stem from intrinsic capability deficits or cascading errors from upstream prerequisites, and introduces CADET, a benchmark of 46 unit tasks with over 33,000 annotated questions, revealing that supplying critical prerequisites can eliminate much of the apparent reasoning deficit.

An EMNLP Findings 2026 paper introduces a multi-agent framework (story, image, and critic agents) that generates over 9,000 multimodal fake news posts and benchmarks 16 open- and closed-source MLLMs, finding they fall short of human-level detection accuracy, especially at judging image authenticity.

ThinkV2V introduces a reasoning-driven framework that activates MLLM thinking before visual generation for instruction-guided video editing, using an MLLM-to-DiT architecture with progressive curriculum training and inference-time thinking scaling. The authors also release the ThinkV2V-150K dataset and ThinkV2V-Bench, showing their 5B DiT model outperforms larger 10B baselines.

MaLiang-Harness introduces a unified framework to address the Program-to-Visual gap in executable programs for image and video generation, evaluating multiple MLLMs on benchmarks and revealing insights into visual generation performance.

The paper proposes DEEPO, a dual-stage reinforcement learning optimization method to reduce hallucination in multimodal large language models by addressing weaknesses in the correction chain from reward to parameter update.

This paper introduces Modl, a technique for creating cross-lingual remote-sensing multimodal large language models by composing domain and language LoRAs with mutual orthogonality, achieving superior performance without paired multilingual data.

VideoMM introduces an adaptive macro-micro inference framework that reduces visual token overhead in video MLLMs, achieving a 6.13× speedup and 7.4% accuracy gain for efficient long-form video understanding.