Impossible Research releases agent harness that hits 99% on ARC-AGI-3

From our team at Impossible Research - an agent harness that plays games, writes code, reasons like a physicist, and saturates the ARC-AGI-3 benchmark. https://twitter.com/havenfeng/status/2077770348876247502

Having played with the ARC-AGI-3 games quite a bit and been impressed by their challenge, this seems quite amazing. I'm not sure I understand the implication of announcing the result of 99% on the public set but no results on the private set yet... is that just that those… Show more

Agree If you want a playground to build general purpose agents - v3 is the easiest place to start After you're happy w/ the core (intelligence via abstract reasoning), apply domain specific knowledge/training https://twitter.com/HavenFeng/status/2077770348876247502

"there's a chance ARC-AGI-3 is already solvable with GPT-5.5-xhigh + tools" now this is GPT-5.6-Sol and Opus/Fable; but still not surprising at all https://twitter.com/HavenFeng/status/2077770348876247502

Wow another one The amount of research built on top of ARC-AGI-3 coming out is starting to snowball The timeline is +3.5 months since V3 launch Looks cool - need to dig into it https://twitter.com/HavenFeng/status/2077770348876247502

Wow another one The amount of research built on top of ARC-AGI-3 coming out has The timeline is +3.5 months since V3 launch Looks cool - need to dig into it https://twitter.com/HavenFeng/status/2077770348876247502