Blog for our ECCV paper Demystifying Video Reasoning In 1935, Erwin Schrödinger put a cat in a box with a Geiger counter. Until you open the box, you do not know if the cat is alive or dead. According to Schrödinger, the cat is alive and dead, at the same time. 90 years later, it is the best picture I have for how a video model solves a problem. Link: hokindeng.com/schrodingers-c…

We hope to launch the Verification Age for video models: 50 scientists, 6 months, 100 rule-based verifiers, 300 new tasks, and a worked RL recipe. Coding models are advancing crazily thanks to the synthetic RL envs paradigms. Verification via codes or Lean gives models super-human math/coding intelligence like that in Astra and Fable. However, video models are still in the GPT-3 stage, where we just dump million hours video data into the model training and hope for magic. In this work, 1) we built 100 rule-based verifiers each paired with a synthetic data generators which give 10k+ diverse samples within the task domain 2) we made additional 300 tasks generators, including classic games i.e. Snake, Gold Miner, Mario, to further scale training data distribution with million samples rendered ready available on HuggingFace 3) ~60 experiments to find a worked RL recipe for video model post-training 4) comparison studies between image models, unified models, video models - which is the best model for visual thinking? 5) failure modes of how VLM-as-Judge will not satisfy video models RL in long term 6) interpretability studies show all image models, unified models, video models emerged "Schrödinger Effect", CoS reasoning through diffusion steps 7) many more insights A detailed thread [1/n]

Accepted at ECCV 2026: "Demystifying Video Reasoning" On Friday Aug 28, first author Ruisi Wang @Rui147000038622 the work at our journal club. The finding: video models don't reason across frames (Chain-of-Frames) — they reason along diffusion denoising steps (Chain-of-Steps). Paper: arxiv.org/abs/2603.16870 Fri Aug 28 · 7:30 PM PT RSVP: lu.ma/video-t1jl

#ICML Two last-authored papers at ICML 2026 this week. A Very Big Video Reasoning Suite. Video models today are optimized for one thing: visual quality. Whether they can reason, track causality, follow constraints, execute multi-step plans, has been nearly impossible to study, because the field never had the data. VBVR changes that. Over 50 researchers and engineers across 30+ institutions built: 1) 150 reasoning tasks grounded in cognitive and developmental psychology 2) 2,015,000 samples and over one million videos, roughly 1,000× larger than the nine prior video-reasoning datasets combined 3) every task procedurally generated with deterministic ground truth, so evaluation is fully rule-based and verifiable, no LLM judges, and it agrees with human judgments Insights: 1) Fine-tuning an open-source 14B model on VBVR lifts overall performance by 84.6 percent, a new state of the art that surpasses every proprietary system we evaluated, including Sora 2 and Veo 3.1. 2) The learned skills generalize to out-of-domain task families the model never saw, early evidence of emergent generalization in video reasoning. 3) And the honest part: gains saturate with scale, on two different architectures. A persistent gap to human performance remains that data alone cannot close. The bottleneck is architectural. 4) One principle kept surfacing throughout: controllability, doing exactly what is asked — is the bedrock of verifiable video reasoning. Great thanks to my collaborators @MaijunxianWang @Rui147000038622 @juyi_lin @SaraJiran @thwiedemer @Kaiagaoqy @carrot0817_ @RubyFreax Lianyu Huang Zelong Hong Jiahui Ge Qianli Ma Hang He Yifan Zhou Lingzi Guo Lantao Mei Jiachen li Hanwen Xing Tianqi Zhao Fengyuan Yu Weihang Xiao Yizheng Jiao Jaden Hou Danyang Zhang Pengcheng Xu Boyang Zhong Zehong Zhao Gaoyun Fang John Kitaoka Yile Xu Hua Xu Kenton Blacutt @tin_ng_qn @siyuansong_ @HaoranSunMr Shaoyue Wen @LinyangNeuroAI @RunmingW @Yanzhi_Wang_ @Mengyue_Yang_ @ziqiao_ma @raphaelmilliere @fredahshi Nuno Vasconcelos @DanielKhashabi Alan Yuille @du_yilun @ZimingLiu11 @Brian_Bo_Li Dahua Lin @liuziwei7 @Vikashplus Yijiang Li @drowsyleilei @caizhongang The second paper — "Vision Language Models Cannot Reason About Physical Transformation" — asks a question Piaget asked children 70 years ago. Pour water into a taller, thinner glass: is it the same amount? Seven-year-olds pass this reliably. We built ConservationBench of 384 real videos of conservation tasks with matched non-conserving controls, and evaluated 112 VLMs across 23,040 trials. Humans score 98 percent; models sit near chance, with 82 of 112 under 10 percent on the strict paired test. The most telling result: models answer "better with no image at all" than with the actual video, and chain-of-thought prompting makes things worse, showing they're leaning on a textual prior, not tracking object state through change. #videomodel #AI #research #reasoning