Can video generation models actually reason about the visual world — or do they just make things look plausible? These sound similar but are fundamentally different. Placing objects correctly, respecting physical laws, and solving multi-step tasks with coherent intermediate states: this is what "Visual Grounded Intelligence" means. VGI-Bench quantifies how far current models are from achieving it. Across 27 tasks and 810 instances, the strongest model — Seedance 2.0 — reaches only 51.0% overall. Three failure modes appear repeatedly: physical collapse (unrealistic deformations and object penetrations), rule violations (reaching plausible end states while ignoring constraints), and object inconsistency (losing temporal identity across frames). Open-source models lag far behind commercial ones, with the strongest open-source model at 19.1%. What the denoising analysis reveals is even more telling. Self-correction during intermediate generation steps occurs in fewer than 1% of transitions. Wrong-to-wrong transitions at mid-denoising reach 23.1%, and later denoising steps primarily refine early hypotheses rather than fixing errors. Fine-tuning on 1M synthetic samples improved planning and spatial skills but degraded temporal and physical capabilities — a tradeoff that points to deeper architectural constraints. VGI-Bench makes clear that visual reasoning remains a genuine unsolved frontier for video generation. Title: VGI-Bench: Probing Visual Intelligence in Video Generation Models URL: #VideoGeneration# #VisualReasoning#

TL;DR A new suite trains and evaluates "native visual reasoning," where visual generation itself is the medium of reasoning, using large-scale data and verifiable rewards. VLM-judge scores swung by up to 92.8% on the same video. Title: VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning URL: Key points 📊 300 tasks, 1.25M training instances, ~3.47M images and 1.3M videos in a large-scale dataset 🎯 A deterministic scorer using classical CV (HSV color segmentation, OCR) hits 0.60+ human agreement, beating GPT-5.5 and Gemini-3.1-Pro 🌀 Coefficients-Preserving Sampling (CPS) keeps predicted and fresh noise coefficients balanced, stabilizing RL exploration 🚀 Transfer to V-ReasonBench jumps from 10.21 to 38.22 (+28.01 pts), the largest gain across seven external benchmarks 🔬 Counterfactual tests show removing input images drops scores by 90%, confirming visual trajectories matter more than text for reasoning ✅ Verifiable-reward RL beats VLM-reward RL by +7.9% in-domain What stands out is moving visual reasoning away from language dependence toward something trainable and verifiable in its own right. #VisualReasoning# #MultimodalAI#

Glimpse from @ElorianAI is on AI Gateway. Bringing frontier visual reasoning at low cost. Analyze images and video, interpret charts, and track objects through motion. 𝚎𝚕𝚘𝚛𝚒𝚊𝚗/𝚐𝚕𝚒𝚖𝚙𝚜𝚎

A tiny new model just beat GPT-6 Sol, Gemini, and Claude Opus 5.5 at visual reasoning. >⁠Glimpse is Elorian's first model >⁠State of the art on visual thinking, counting & labeling, and short video understanding

Spent years watching models ace a single image and completely whiff on anything that moves or changes angle. Today we’re fixing that. Try Glimpse now, our first visual thinking model with competitive frontier performance compared to Astra on key public visual reasoning benchmarks.

according to Elorian's evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks.

GPT-6 Astra really looks like openai went hard on pre-training world modeling, math/abstraction, spatial + visual reasoning, interacting with environments- it pretty much leapfrogs every other model but coding is the interesting part. vs. Fable 5.1, the gap really isn’t that big

Step 3.7 Flash is now in the Baseten Model Library! This is a 198B-parameter sparse MoE model (11B active per token) with native image and video input, and a 256K context window. It's a strong option for visual reasoning, agentic coding, and long-context tasks.

Kimi K2.5 tech report just dropped! Quick hits: - Joint text–vision training: pretrained with 15T vision-text tokens, zero-vision SFT (text-only) to activate visual reasoning - Agent Swarm + PARL: dynamically orchestrated parallel sub-agents, up to 4.5× lower latency, 78.4% on BrowseComp - MoonViT-3D: a unified image–video encoder with 4× temporal compression, enabling 4× longer videos in the same context - Toggle: token-efficient RL, 25–30% fewer tokens with no accuracy drop Here's our work toward scalable, real-world agentic intelligence. More details in the report 👉

Claude Haiku 5.5 just dropped, and the benchmark jump is HUGE. @AnthropicAI's smallest model beats GPT-6 Luna on every benchmark where both have a reported score in the release chart. 🔥 Compared with Haiku 4.5: Computer use: 15.7% → 72.4% on OSWorld 2.1, versus Luna's 48.9%. Reasoning: 10.2% → 45.9% on Humanity's Last Exam without tools. With tools, it hits 57.4%. Agentic coding: 0% → 39.2% on Terminal-Bench 4.0, versus Luna's 16.4%. FrontierCode: 46.4%, beating Luna's 42.4% and approaching Sonnet 5.5's 52.1%. Visual reasoning: 6.4% → 46.4% on Chartography, versus Luna's 29.1%. Knowledge work: 1620 on GDPval-AA and 1578 on AA-Briefcase, ahead of Luna's 1437 and 1336. Sonnet 5.5 still leads across the chart, especially on Terminal-Bench. But the leap from the previous Haiku is enormous, particularly for computer use and reasoning.