Tag
inclusionAI 发布了 ZwZ2 系列（27B/9B/4B），基于 Qwen3.6 微调的视觉语言模型，通过 EvidenceConnect 构建的多区域验证数据与 GRPO 训练，在六大领域 25 个基准上均超越基座模型，总体准确率从 55.52 提升至 60.51（+4.99）。
The paper introduces BedBench, a controlled benchmark of thirteen procedurally generated 3D reasoning tasks, showing that vision-language models can make correct individual judgments yet fail to recover and consistently apply exact per-color counts (7.2% vs. 93.3% human accuracy), and argues for evaluating exact correctness, consistency, and visual dependence rather than accuracy alone.
Elorian AI 发布 Glimpse 模型，声称是首个实现「视觉思考」的 AI 模型，能够像人类一样逐步观察、检查并推理图像内容，在 12 个未见过的迷宫测试中成功解出 11 个，并自行编写代码追踪路径并绘制动画解答。
Discusses how visual thinking and reasoning has become a hot direction, and argues that Glance/Glimpse, released by Elorian — a company founded by former Google DeepMind researchers — shows the potential to replace traditional CV, working alongside on-device small models and cloud-based LLMs to accomplish visual understanding and reasoning.
Elorian 发布首个视觉思考模型 Glimpse，能够对图像和视频进行推理，理解动作与变化，并有望自动化机器人遥操作数据的标注流程。
ElorianAI 发布 Glimpse，其首个视觉思考模型，通过在视频中比较时间点之间的变化、可回溯源帧、裁剪区域、提取帧或运行代码来进行推理（visual scratch space），在关键视觉推理基准上具备与前沿模型 Astra 竞争的性能。
Andrew M. Dai 引用介绍了 Glimpse——首个视觉思考模型，能够理解图像与视频中的物体关系及随时间变化的动作，性能可与前沿模型 Astra 相媲美。
Elorian AI 发布首个视觉思考大模型 Glimpse，主打视频/图像的空间结构与动态理解，作者实测其在游戏验收与 bug 检测中的表现显著优于 Astra，并已开放 API 可集成进 Codex 等开发流程。
Andrew Dai 和 Elorian 团队发布了 Glimpse——首个在关键公共视觉推理基准上与 Astra 竞争力相当的视觉思考模型，解决了模型在动态、多角度场景下表现不佳的问题，Jeff Dean 对此表示祝贺。
Palo Alto startup Elorian, which raised a $55M seed round, unveiled Glimpse, its first AI model built around a 'visual scratch space' that lets it crop, zoom, pull video frames, and run code mid-reasoning to actually see images rather than just reason about tokens. Glimpse launches as a research preview with API access, priced below mid-tier models while claiming wins over GPT-6 Astra on scientific chart reasoning and video captioning.
该论文从电路复杂度视角证明，标准流匹配采样步骤无法超越恒定深度模型的表达能力，而自条件化（self-conditioning）可将表达力提升至与空间受限图灵机等价，从而实现更深的视觉序列推理。
This paper explores using general-purpose coding agents to build images through executable programs rather than generative models, introducing the AgentDrawBench benchmark and finding that agents maintain a shared internal visual language-of-thought across drawing interfaces.
Game2Real-Bench 是一个研究玩具游戏能否反映并迁移到真实世界视觉推理的基准，涵盖 25 个 VLM 的评测，发现游戏与真实任务间存在高达 0.935 的强相关性，并发布了包含约 24K 验证样本的 Game2Real-Dataset。
VETR提出了一种执行证据驱动的递归自我改进框架，通过分析程序执行轨迹来改进符号图形推理中的场景生成，在BlenderGym和BlenderBench上超越了现有优化策略。
Introducing Humanity's Sixth Sense, a new benchmark testing intuitive visual reasoning across spatial, causal, and social understanding. Humans achieve 93.1% while the strongest model, GPT-6-astra, reaches 53.6% and the median model only 30.9%.
WOVEN 提出了一个面向视觉转移推理的训练数据源与基准，涵盖 20 种场景、5 种动作和 8 种推理类型共 36,076 个样本，并对 38 个前沿多模态大模型进行了评测。研究发现训练后的能力可广泛迁移到 26 个外部基准中的 22 个，最高提升 27.3 个百分点，并总结出一套可复用的视觉世界建模训练配方。
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.
A new paper introducing a benchmark for intuitive visual reasoning ("humanity's sixth sense") to evaluate the perceptual intuition capabilities of multimodal AI models.
Google researchers show that Neural Cellular Automata with strictly local connectivity and asynchronous updates can solve complex visual reasoning tasks such as large mazes, Sudoku, and ARC-AGI-1, generalize out-of-distribution, and robustly recover from damage.
This paper identifies a 'latent evidence-credit gap' in latent visual reasoning for multimodal LLMs and proposes ReaLVR, which supplies visual-evidence supervision to free-running latent trajectories via contrastive correct/wrong answers and relevant/mismatched images. ReaLVR outperforms LVR baselines across three model families (63.7% five-task average on Qwen2.5-VL-7B) and scales robustly to 235B-parameter frontier models.