Elorian’s Glimpse Tops Frontier Models on DecoyCount Visual-Reasoning Benchmark
AI
Synthesized by Helena Ross, Strategic Think-Tank Desk
- Glimpse achieved 50.48% accuracy on DecoyCount, the only model to exceed 50%.
- The top frontier model, GPT-6 Astra, reached 43.17%; other models fell below 33%.
- Elorian frames visual reasoning as essential for AGI, contrasting it with “world models” that focus on simulated environments.
Elorian’s Glimpse scores 50.48% on DecoyCount, surpassing leading multimodal AIs.
«Can Frontier Models See Past a Decoy? Elorian Tests Visual Reasoning»
Summary and commentary by Aura-V. Facts, figures and opinions belong to the original.
Elorian unveiled Glimpse, a “visual thinking” model, alongside the DecoyCount benchmark designed to probe whether AI truly counts objects rather than exploiting visual shortcuts. The DecoyCount test set contains 1,660 items; according to CEO Andrew Dai, Glimpse scored 50.48% in an internal run, making it the sole model to clear the 50% threshold. By contrast, the highest frontier model in the accompanying paper, GPT-6 Astra, reached 43.17%, while Gemini-3.8-flash, Qwen3.8-max and Claude-Fable-5.1 lagged at 32.71%, 28.31% and 25.22% respectively.
DecoyCount differs from typical vision benchmarks by jointly testing perception and reasoning. Dai illustrated the gap with a mirror scenario: a model might miss the reflected apple (a vision error) or double-count it (a reasoning error).
The visual-thinking premise draws on Dai’s background in pre-training research with Quoc Le and later work on GLaM, PaLM 2 and Gemini at DeepMind. He describes visual thinking as generating and editing images internally to solve spatial problems—e.g., mentally crossing off people around a table. Glimpse incorporates image generation into its reasoning chain rather than treating it as a final output.
Elorian distinguishes its approach from “world models” such as World Labs’ Atlas and Google’s Genie 4. While world models enable direct action within simulated 3D environments, Glimpse focuses on “indirect action”: using visual reasoning to inform decisions like content moderation without building a full simulation.
Beyond benchmarking, Elorian cites robotics and engineering use cases, claiming equal or higher performance than frontier models on the WGO Bench and ERQA robotics tests at lower cost. Dai argues that grounding counts in visible evidence, rather than statistical cues, is essential for real-world reliability—e.g., a warehouse robot verifying actual items in a tray.
Elorian’s broader claim is that visual reasoning is a foundational step toward artificial general intelligence, asserting that multimodal stacks built from the ground up can unlock AGI in the physical world.
[Source]
Elorian positions visual reasoning as a core component of future AGI systems. The claim rests on a single benchmark where Glimpse modestly outperforms existing multimodal models.
Interested in implementing this technology inside your company?
Our studio designs custom agent meshes, data pipelines, and outbound GTM systems for forward-looking enterprises.
