Comment by teravor
2 hours ago
it looks like what they are doing is using a frontier model to write a simulator for a game and then solve using it.
it's not as impressive as it looks. the goals of Arc-AGI-like constructs is to get an IQ-like figure using raw'ish 2D measurement 'games' in the hope that it would signify something meaningful.
what this harness does is get the model to write a simulator first, it's measuring something entirely different.
> State grounding turns raw observations into objects, variables, and relations that can be tracked. Mechanism discovery finds how that state changes under an action and writes the rule as an executable program
The way I'm reading this isn't that they are writing a game simulator, but rather that they have two things they are evolving - a perceptual model of the game mapping from pixels to objects, and a behavioral model of how each action acts upon these perceptual objects. The behavioral model is written as a program that can be backtested by the game states and actions they have already taken to see if they are correctly predicting the resulting next game state.
The ARC AGI 3 games are non-trivial, and I think it's very impressive to see them doing well using this approach.
I'd agree with their conclusion:
> We read a saturated ARC‑3 as the new beginning: mechanism discovery as a general capability — grounding the causal structure of a world through the agentic loop of action and perception, in environments far richer than a 64×64 grid. This is where we are heading to.
This is the way that an animal learns about it's environment - by observation (and innate biases) to recognize the objects in the environment, and predict their behavior, both autonomous (which AGI ARC 3 doesn't test - the objects in the environment are passive), and in reaction to the animal's behavior. The animal predicts and observes, updating its predictions when it is wrong.
A system that could do this in a messy, dynamic, real-world environment would seem a like genuine step in the direction of animal intelligence, especially if it could ditch the symbolic representations.
Writing a simulator requires understanding the game well enough to spec the simulator. Acquiring said understanding - within the action limits of the benchmark – seems like the heart of the challenge, so this doesn't strike me as "cheating" at all.
This is classic goalpost movement. Arc-AGI-3 was launched this year with roughly 0.5% success for frontier models. being able to 99% it in less than six months sets a new record for Arc-AGI saturation timeline. Speaking of singularity measures. It is definitely a big deal, not least in that Chollet needs to cancel his summer vacation and write Arc-AGI-4 now.
Arc AGI are simple games, the hardness comes from the input being basically adversarial to LLM training. if you use an LLM scaffold that removes the adversarial part you are measuring something else.
the harness basically outsources the alien nature of what the LLM is asked to do to algorithms it writes. this would actually be impressive if you got it to do that for a much more complicated game than Arc.
with this harness the ARC AGI test becomes a test of whether or not the model can work out the transition rules in a very simple game.
A great deal of mathematics is transforming nonlinear problems into linear ones and solving them with linear techniques. Others are solving non linear problems through stochastic methods. In almost all cases most non trivial math is done by transforming a harder problem into a simpler one.
I get what you mean in terms of testing the model itself to see its improvement in some domain. However if you can transform the domain to be better adapted to the model and achieve the desired results, this is indeed an accomplishment because a whole domain of problems is shown to be practically feasible with this technique without expensive model improvements. Of course the benchmark still exists without the harness, but the harness also exists which allows these problems to be solved.
As noted elsewhere the models themselves were used to build the harness, which means the models can in fact score this scores without intervention but building a harness for themselves adapted to the domain and using it. Is this cheating by the goal posts you’re setting?
There’s a real tension between “I want to solve problems and this technique shows how to solve the problem domain,” and the “I want to measure how something performs unassisted with other techniques.” Fortunately it’s not a mutually exclusive situation. You can do both simultaneously, gain the benefit of the technique to transform the problem into something tractable and keep measuring using the benchmark.
2 replies →
The point of Arc-AGI-3 is to measure model performance. We already know that models can one-shot and iterate on very rudimentary game implementations. And, naturally, once it effectively has a copy of the source code, it can use that to play the game better.
This harness is really moving the goalpost by defeating the entire point of the test. Instead of seeing the strength of a model's world view, its ability to internally derive and intuit rules, and its ability to keep track of game state over time, we're just letting the AI cheat. This is just the LLM equivalent of running a chess engine to the side.
And this harness would not work in a remotely complex game and relies on the fact that Arc-AGI-3 is a focused test that only made the games as complicated as they needed to be for current model performance.
I think this is just too simplistic a take; Arc-AGI-1 was wide open to all models, harnesses, etc, and had quite a lot of innovative structures implemented by hobbyists. At the time, this was seen as a good thing (it was), because we don't know the best system architecture for all sorts of problems right now -> innovation is good.
The games are designed to allow assessment of a system. Knowing better systems to solve the games is a step forward. If any of the frontier labs could have one-shotted -3 in March with a custom harness, they would have done so.
Sounds like a distinction between sport and work. How useful is pure model performance if it's known that there are conditions in which even greater performance can be achieved on real tasks? How useful is it to know how fast/far a person can run if they can ride a bicycle or drive a vehicle?
okay, what if this benchmark were just called Arc, and all you knew about it was what questions it asked? it just looks like a bunch of arcade games, which should tell you that it doesn't test AGI at all.
on the flip side, the idea that most tests are bad, even standardized tests, the tests that you scored well on that gave you all your opportunities in life: it cuts to the emotional, grounded core, the absolute foundation, of too many people. in the crowd of hacker news commenters; people who buy anthropic shares at retail; the people who work at tech companies; and their kids, families, etc., who are a bunch of nobodies, there are a lot more incentives to believe "stupid fucking arcade games test AGI" than not.
The simulator the model builds is comparable to the mental model of the game humans create. It is also much more efficient, GPT 5.6 Sol cost $25,000 to run on ARC-AGI-3
then they should try to use that for a more complicated game than Arc AGI. Arc games are simple by design, if you have the model simulate them they become trivial.
>if you have the model simulate them they become trivial
Eh, this is kind of sounds like being a prey animal that develops an almost unbeatable colored camouflage and then the predator develops infrared vision making it useless and the prey saying "no fair, you cheated".
People use algorithmic models all the time on problems that are far too difficult or large for their minds to conceptualize, is this not just an extension of that?