- Who it helps
- Researchers evaluating image reasoning and people interpreting vision benchmarks.
- Availability
- Public October 26 preprint. The abstract describes 1,660 images and seven decoy mechanisms; artifacts were not executed.
Paired questions
Neutral, cue-guided and multi-image questions probe the same counting problem. Better answers with a cue suggest that some failures concern choosing the intended reasoning process.
DecoyCount separates seeing objects from choosing what to count ↗
Scope of the finding
The authors report accuracy below 50% for tested frontier models and alignment gaps of 11.8–22.1 percentage points. A separate Glimpse claim in an interview uses a single internal run, so it is not an equivalent independent comparison.
DecoyCount separates seeing objects from choosing what to count ↗Elorian interview and comparison caveats ↗
What you can do
- Read the linked primary evidence and its access limits before planning a trial. Open reference ↗
Where the evidence stops
Paper abstract and interview read November 6. Neither results nor contamination checks were reproduced.
Sources read. Sources read November 6, 2026 UTC. No third-party project, model inference or complete product workflow was executed.
Sources & further reading
Paper · paper
DecoyCount separates seeing objects from choosing what to count ↗
Primary page read November 6. Claims were checked against the page; the workflow was not executed.
DecoyCount authors · page checked 2026-11-06
Web · analysis
Elorian interview and comparison caveats ↗
Primary page read November 6. Claims were checked against the page; the workflow was not executed.
Richard MacManus · page checked 2026-11-06
Coverage note: Original YouTube demonstrations and independent cross-platform attention were not established. No end-to-end workflow was executed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
DecoyCount uses visual decoys to test whether a model chooses the right counting rule without a hint.