Johan Land

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

Huge congrats to @LandJohan on verification of this amazing achievement!

@charlielidbury @arcprize 76.11%

@arcprize @LandJohan @arcprize @LandJohan What % did you get on the public eval set? arcprize.org/media/data/lea… only shows your semi-private scores

New SOTA public submission to ARC-AGI: - V1: 94.5%, $11.4/task - V2: 72.9%, $38.9/task Based on GPT 5.2, this bespoke refinement submission by @LandJohan ensembles many approaches together

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

@LandJohan If you know before hand when they are right or wrong then you would score 100% on any benchmark I guess ?

Here is the ensemble agent i was conceptually imagining, now scoring >70% on v2 of ARC-AGI.

@codewithimanshu @chatgpt21 Yeah, proves scaffolding is impactful!

@chatgpt21 @LandJohan @chatgpt21, that's a significant improvement in ARC-AGI performance, especially considering the cost per task.

New SOTA public ARC-AGI!! V1: 94.5%, $11.4/task V2: 72.9%, $38.9/task GPT 5.2 refine! Arc says - “submission by @LandJohan ensembles many approaches together” Congratulations!

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

@LandJohan Any underpinning research paper to share?

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): kaggle.com/code/johanland… @GregKamradt , holiday break is over 🙂 semi-private when? #ARCAGI #AIResearch

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

@arcprize @LandJohan tell him, he need to consider the x-axis

@thedjpetersen @ItsBrain4Brain @arcprize The code is open source: github.com/beetree/ARC-AGI Do whatever you want with it :) I don't think I even put in any license in there so it's free for all!

@LandJohan @ItsBrain4Brain @arcprize Curious could you share?

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.

@LandJohan @ender_mode_ Respect. Also congrats on reaching new high scores! I'd love to understand more about the "council" mechanics. Is there a paper?

Hall of fame twitter profile, guy has like 20 followers

@diegocabezas01 It's all public source: github.com/beetree/ARC-AGI 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.

@LandJohan That’s actually amazing any repo for the model? Or info on how it was built?

AI model based on GPT-5.2 got a new RECORD in ARC-AGI 2: 73%

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

@teortaxesTex Their scaling is just crazy, no chinese model even get past 20% yet, can whale bros did it?

Anyone still wants to argue that 5.2 is not the strongest model around? (yes this is a multi-model setup, but it's principally built around 5.2) ARC-AGI 2 will be solved this year.

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

GPT-5.2 is a beast, extremely underrated I hope @OpenAI keeps it.

@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.

Wow! New records on ARC-AGI, based on GPT 5.2! V1 is now at 94.5%, almost saturated! V2 is at 72.9%, I anticipate that it will near 100% in a few months! Next challenge is completing ARC-AGI-3!

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

Within just 10 months, performance on the ARC-AGI-2 benchmark surpassed 75%. Let that sink in.

@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.

Just to set the record straight 📌 — no single model did the heavy lifting here. GPT helped with the math and structural scaffolding 🧮 Gemini handled the core logic, reasoning flow, and split image generation with precision 🧠🖼️ Claude absolutely dominated the creative side — especially when it came to accurate reproduction, narrative cohesion, and score fidelity 🎼✍️ When the outputs were evaluated, the results weren’t “exact benchmark copies,” but they were so close in behavior and performance that the distinction becomes academic. That’s the real takeaway. This is exactly why I’ve been preaching multi-agent / multi-model orchestration for over a year now 🔁🤝. Different intelligences excel at different layers — forcing one model to do everything is leaving performance on the table. Shoutout to @RileyRalmuto as well — he’s been pushing polyphonic system designs in the same spirit. Different voices, different strengths, one coherent output 🎶🧩. This isn’t model worship. It’s systems engineering. And this is where the real gains are coming from 🚀

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

📢 @arcprize: Johan Land's AI ensemble hits 94.5% on ARC-AGI leaderboard Solo researcher outperforms GPT-5.2 Pro (90.5%) on the benchmark designed to test generalization. Uses an ensemble approach combining multiple techniques. Code open-sourced. Why it matters: The gap between frontier labs and indie researchers keeps shrinking.

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

Our active discord member "beetree" (@LandJohan) submitted a SOTA solution that scored 72.9% on ARC-AGI-2. Congrats! Come join us on discord.gg/9b77dPAmcA

@BillyHoy1_ Not really :)

Just a good old ensembles of models

@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.

Diverse ensembles of models are The Way. Thus saith Tomathustra. (See the model ranking criterion from 1994 in my profile pic. Ideally, the highly-ranked models would have mutually orthogonal error vectors.)