Elorian releases DecoyCount benchmark and Glimpse visual reasoning model
DecoyCount shows frontier models score below 50% on visual reasoning; Elorian’s Glimpse exceeds 50% in internal tests.
1 source · single source
- Elorian launched Glimpse, a visual reasoning model that reasons about the physical world from images.
- A new benchmark, DecoyCount, tests whether models genuinely count objects or rely on visual shortcuts.
- Frontier models scored below 50% on DecoyCount, with GPT-6 Astra at 43.17% and Gemini-3.8-flash at 32.71%.
- Glimpse scored 50.48% on DecoyCount in an internal test, exceeding all reported frontier model results.
- Elorian was co-founded by Andrew Dai, a former Google DeepMind researcher, and raised $55 million in seed funding.
Elorian, a startup co-founded by former Google DeepMind researcher Andrew Dai, has released Glimpse, a model described as a new category of visual thinking systems that reason about the physical world from visual inputs. The company also introduced DecoyCount, a benchmark designed to test whether models can genuinely count objects or rely on visual shortcuts and substitution heuristics.
DecoyCount evaluates models on seven interpretable counting failure mechanisms, aiming to isolate vision and reasoning errors that existing benchmarks often conflate. According to a paper accompanying the benchmark, frontier models performed below 50% accuracy on DecoyCount, with GPT-6 Astra achieving the highest score at 43.17%. Other models reported in the paper included Gemini-3.8-flash at 32.71%, Qwen3.8-max at 28.31%, and Claude-Fable-5.1 at 25.22%.
In internal testing, Glimpse scored 50.48% on DecoyCount, surpassing all reported frontier model results. The evaluation images and text were withheld from Glimpse’s training data, and contamination checks found no overlap, according to Dai. Glimpse’s results were obtained from a single run without an agent or tool harness, while the frontier models’ scores were averaged across three runs under their reported settings.
Elorian’s approach is grounded in the concept of visual thinking from human psychology, where reasoning involves generating and editing mental images. Dai described this as a process of visualizing and manipulating images in the mind to solve spatial or visual problems, such as mentally arranging furniture or counting objects around a table. Glimpse applies this idea by incorporating image generation and editing as part of its reasoning trace, alongside language.
The company positions Glimpse as distinct from world models like World Labs’ Atlas or Google’s Genie 4, which focus on direct action in simulated 3D environments. Instead, Elorian emphasizes indirect action—using visual reasoning to inform decisions or workflows, such as content moderation or compliance checks, without requiring a full 3D rendering.
Elorian emerged from stealth in April with $55 million in seed funding, co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital. The company is also partnering with AI data company Mercor on the DecoyCount benchmark release.
Beyond Glimpse, Elorian highlights potential applications in robotics, engineering, and defense, where accurate visual reasoning could improve task performance. The company claims its model rates equal or higher on two robotics benchmarks, WGO Bench and ERQA, though specific scores were not provided in the source.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74