Ex-Google Researcher's Elorian Launches Glimpse: Frontier Models Fail 50% on Visual Counting Benchmark
Latent Space · rss · 2026-10-29
Latent Space interviews Elorian CEO Andrew Dai (ex-Google; co-author of the 2015 semi-supervised sequence learning paper). Backed by a $55M seed round, Elorian launched the visual reasoning model Glimpse and the DecoyCount benchmark with Mercor.
- DecoyCount tests whether models genuinely count objects or rely on visual shortcuts, exposing seven interpretable failure mechanisms
- All frontier models scored below 50%: best was GPT-6 Astra at 43.17%, Gemini-3.8-flash 32.71%, Qwen3.8-max 28.31%, Claude-Fable-5.1 22.22%; Dai says Glimpse scored 50.48% on 1,660 items without agents or tools
- Paper conclusion: stronger general reasoning doesn't automatically generalize to aligned visual reasoning
- Glimpse's approach uses image generation/editing inside the reasoning trace, mimicking human visual thinking
- Unlike world models (World Labs' Atlas, Google Genie 4) that enable direct action in simulated 3D environments, Elorian focuses on "indirect action": visual reasoning to inform decisions and workflows (content moderation, building-code compliance, warehouse robotics)
- Dai argues visual reasoning is a foundational requirement for AGI; Elorian also claims parity or better than frontier models on robotics benchmarks WGO Bench and ERQA at lower cost
More from Models
- Sentdex: DeepSeek v4.1 Flash now his top local token generator ever — Sentdex · 2026-10-29
- LoopVL: 1B-parameter VLM reuses 32 layers across 128 calls, beats 4B baselines on 5/8 benchmarks — jiqizhixin · 2026-10-29
- Fable 5 praised for blending deep scientific knowledge with storytelling delight — repligate · 2026-10-29
- ChatGPT's unified UI confuses users: Pro model hides under GPT-6 Sol, not Astra — Angaisb_ · 2026-10-29
- User burns through OpenAI's $100 weekly quota in under 10 hours — teortaxesTex · 2026-10-29
- Hy-Nomas interactive video world model demo launches on Hugging Face — _akhaliq · 2026-10-29