Johan Land

[Johan Land @LandJohan]

@82deutschmark Thanks man! And, thanks a ton for all you do for the community!

[Johan Land @LandJohan]

@charlielidbury @arcprize 76.11%

[Johan Land @LandJohan]

@ndzfs No, you still need to find the solution, not just judge it

[Johan Land @LandJohan]

@codewithimanshu @chrisgpt Yeah, proves scaffolding is impactful!

[Johan Land @LandJohan]

@desai_pratik No, sorry. Code is open source though.

[Johan Land @LandJohan]

Just scored 76.11% on ARC-AGI 2 — beating public GPT-5.2 and Gemini-3-Pro baselines by >20%, and (as far as I know) the best publicly reported result so far.

Approach: what I’d call Multi-Model Reflective Reasoning
- Using GPT-5.2, Gemini-3, Opus 4.5
- Long-horizon/multi-step reasoning (~6hrs/problem)
- Agentic codegen (>100,000 python calls)
- Visual reasoning
- Council of judges

Fun fact: all solver code was written by Gemini-3-CLI.
Does this count as AI generating a new AI that beats the prior SOTA? 🤔

Full run + code (open source): https://t.co/8HZJV5XjIK

@GregKamradt , holiday break is over 🙂 semi-private when?

#ARCAGI #AIResearch

[Johan Land @LandJohan]

@pradanadimass @arcprize Oh, I am :) can trade y for x at efficient ratio. This was really going for max y though.

[Johan Land @LandJohan]

@thedjpetersen @ItsBrain4Brain @arcprize The code is open source: https://t.co/pqWRg79oqy

Do whatever you want with it :) I don't think I even put in any license in there so it's free for all!

[Johan Land @LandJohan]

No, but it's all open source. Maybe I should write a paper.

Essentially, it's gathering the reasoning traces for all possible solutions. Then it's exposing those to three different judges with slighly different roles. The judges then express their opinions after which a solution is picked.

[Johan Land @LandJohan]

@diegocabezas01 It's all public source: https://t.co/pqWRg79oqy

Check the v7 branch, that's the latest. Actually, go back a few commits and you'll find an even higher performing version - I had to dumb it down a bit for the submission.

[Johan Land @LandJohan]

@Viam_Invenias_0 @teortaxesTex I think you can get higher with the chinese models with an approach similar to the one I took here.

[Johan Land @LandJohan]

@captain_marrvel @OpenAI Beautiful model indeed.
Slow though :/

[Johan Land @LandJohan]

@DeryaTR_ Next challenge is indeed ARC-AGI-3!

The beautiful thing about ARC-AGI is that they allow "hobbyists" like myself to fairly be benchmarked against the labs.

[Johan Land @LandJohan]

@kimmonismus It's moving fast, indeed! Exciting times ahead!

[Johan Land @LandJohan]

@joshlee361 Largely I agree. Few other things to it, but the key thing indeed is that different models/prompts/modalities/chaining generate diverse results. But then, you also need to "know when you know" and "know when you don't know" which is the other half of the problem.

[Johan Land @LandJohan]

@BadTechBandit @arcprize "Indie researcher", I like it :)

[Johan Land @LandJohan]

@permaximum88 Yes! I love the community! Thanks @permaximum88 for everything you're doing!

[Johan Land @LandJohan]

@BillyHoy1_ Not really :)

[Johan Land @LandJohan]

@SuperbBias Diversity is the keyword indeed. Of the biggest insights I had was to induce diversity in the models by forcing them to thinking in different spaces and modalities.