To be clear, we’ll want to see how this performs against the hold-out set. If it holds up, though, it’s a big deal, and kind of in line with the vibes this year, which I’d typify as ‘harness matters’. Maybe we’d upgrade to ‘harness matters immensely’ if this can 100% ARC-AGI-3 on existing models (more in the 13% range without this harness).
I’m pretty excited to see what sort of generalization we come to over the next 12 months on the harness side: if it turns out this can be RLed in as ‘consider if building a world model might help here’ and we get this as another native capacity, that will be interesting. If we get 100 of those problem-solving strategies all included, feels like we will see another hurdle cleared in terms of usefulness.
Any custom harness for a problem shows that harness engineering is going away. Eventually models will introspect problems, then build custom harnesses tailored to that. Then use and modify the ephemeral harness as required.
Sol Ultra style is the path forward. The models are smart enough to self serve their tooling and processes. Given a problem they can figure it out and ask for directions when needed.
It's likely smart enough. It just needs to be told to do it and provided the ability to introspect it. How close could foundational models get to building this harness if explicitly prompted to?
We've only just started training models to use tools. Next, we'll train them to build them. Harness engineering is an ephemeral art.
Let's maybe say not experienced enough / insufficiently RL'ed then -- 5.6 Sol did not reach for a harness solution like this when it got only 13% or son AGI-3 recently. I agree it's interesting to find the point in the prompting when it could 'tip' and do this. I have no instinct for where that point is, except that it must be somewhere, because I bet the Schema Harness was not hardcoded.
Letting the provider decide for the harness is a terrible idea in my eyes. Outsourcing harnessing is giving up control over the AI and equivalent to abandoning your sovereignty. It is a regression to a pre-enlightenment era.
In the spirit of ARC-AGI-3-like challenges, we just tested if frontier AI models are able to solve a lovely puzzle game, Baba Is You: https://quesma.com/blog/baba-is-bench/
A year ago, Sonnet 4 barely solved the first level. Now, both Fable 5 and GPT-5.6 Sol beat the first two stages. GPT 5.2 is slow, but efficient, while Gemini 3.1 Pro and 3.5 Flash struggle.
FWIW: "Baba Is You" is 7 years old and heralded as one of the greatest puzzle games of all times, with guides and solutions shared all over the internet. How to beat this game is 100% in the training set.
I can't tell based on this article if the authors intend to submit this to Arc-AGI for the private / held-out set of games for a verified score. The final section sorta seems like they won't bother because Arc-AGI-3 is "now saturated"
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 animals behavior. The animal predicts and observes, updating it's predictions when it is wrong.
A system that could do this in a messy, dynamic, real-world environment would seem like genuine step in the direction of animal intelligence, especially if it could ditch the symbolic representations.
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.
> ARC-AGI-3 is an interactive reasoning benchmark which challenges AI agents to explore novel environments, acquire goals on the fly, build adaptable world models, and learn continuously.
This harness does nothing to actually accomplish those goals.
It's a clever trick, sure, but you aren't allowed to use a calculator on your basic algebra tests in school for a reason.
I don't think we got continuous learning here, but we very specifically got interim goal setting and custom world models; the thinking traces demonstrate this round trip of building a world model, mental or coded, then stopping when reality doesn't correlate, then hypothesizing and creating a new model.
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.
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?
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.
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
> simulator the model builds is comparable to the mental model of the game humans create
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.
Big jump for sure, but definitely comes with a giant grain of salt lacking open-sourcing the harness itself and measuring performance on the held-out set.
> Both scores come from a fixed fallback rule: Opus 4.8 and Sol xhigh run first; games scoring below 80 are rerun with Fable 5 and Sol max, respectively, and the higher per-game score is retained.
hmm, this is like pass@n until you get the high watermark? How would this mean anything?
(1) What does it score on the private test set?
(2) Does this approach generalize to, e.g., Atari or NES games, or is it just hard-coding priors about the games into the model (as Chollet specifically warned was a chronic problem in benchmarks in the original Arc-AGI paper)
Can someone tell me what the catch is? To outperform the state-of-the-art so drastically would be massive news, and surely the ARC Foundation would have tested this against the private data set, right?
This is not actually running the Arc-AGI-3 anymore. To summarize TFA:
1. The AI plays the game and records outputs.
2. The AI does TDD using those outputs to create its own copy of the game.
3. The AI then uses it's copy of the source code to understand the rules. This bypasses the intent for Arc-AGI-3 to test the underlying model's ability to intuit game rules naturally, like a human.
4. The AI then runs simulated moves on the copy of the game before playing them in the "real" game. This bypasses the intent for Arc-AGI-3 to test the underlying model's ability to plan and predict moves, and track world state in its "head" over time.
To make an apt comparison... You go to get your chess ELO. You don't know chess at all and you're really bad at it, so you pull out your laptop and write a chess engine. Then when you go to get ranked, you just copy the moves from the software. Now you're a grand master.
It doesn't necessarily mean anything to reach 99% on the public set. All of the public set is known in advance, so it's possible to hardcode rules that make this easy for the models. ARC-AGI-3 is supposed to measure generalization to unseen games, so the only score that matters is the score on the held out private test set that nobody outside the ARC prize foundation has access to. Also, I believe the private set is significantly harder than the public set.
I pretty much predicted this. If such smart models capable of doing math research fail so hard on such simple games the interface is the problem, not the model. Right harness provides a good interface between the problem and the intelligence.
> Schema, the harness we introduce today, reaches 99% on the ARC‑AGI‑3 Public set using Claude Opus 4.8 and Fable 5, and 95.35% using GPT‑5.6 Sol.
Impressive results. Will this translate to coding agents (and training general purpose and for coding LLMs) too?
---
> When Michelson and Morley could not detect the medium light was supposed to wave in, Lorentz took the first route: keep the aether, patch the rules with contraction hypotheses that absorbed the null result. Einstein took the second: in special relativity, he discarded the aether as part of the state and made simultaneity frame-relative, yielding a simple electrodynamics of moving bodies.
BECs, SVT, Superfluid Quantum Gravity
Massful photons are modeled with Proca fields. Like Einstein, Proca was also a student of Minkowski. The Mass-Equivalence principle ~~does not~~ still holds if photons have mass.
> could not detect the medium light was supposed to wave in,
Superfluid Quantum Gravity (Fedi,) says that there is a medium that light waves through; there is not nothing in space, space is a quantum dilatant superfluid with near-zero viscosity.
To be clear, we’ll want to see how this performs against the hold-out set. If it holds up, though, it’s a big deal, and kind of in line with the vibes this year, which I’d typify as ‘harness matters’. Maybe we’d upgrade to ‘harness matters immensely’ if this can 100% ARC-AGI-3 on existing models (more in the 13% range without this harness).
I’m pretty excited to see what sort of generalization we come to over the next 12 months on the harness side: if it turns out this can be RLed in as ‘consider if building a world model might help here’ and we get this as another native capacity, that will be interesting. If we get 100 of those problem-solving strategies all included, feels like we will see another hurdle cleared in terms of usefulness.
> To be clear, we’ll want to see how this performs against the hold-out set.
they could take open weight model, and check what will be impact from that harness on hold-out
Any custom harness for a problem shows that harness engineering is going away. Eventually models will introspect problems, then build custom harnesses tailored to that. Then use and modify the ephemeral harness as required.
Sol Ultra style is the path forward. The models are smart enough to self serve their tooling and processes. Given a problem they can figure it out and ask for directions when needed.
Indeed. And you can make this case about any tooling at all that is model adjacent.
Including by extension all programs, operating systems, or eventually hardware, I suppose.
I seriously doubt this, especially in a world in which there's not just one model.
This makes sense if the models some how become unified.
Except in this case, it isn't yet smart enough. But I agree, building this capability in is coming, and will be really awesome.
It's likely smart enough. It just needs to be told to do it and provided the ability to introspect it. How close could foundational models get to building this harness if explicitly prompted to?
We've only just started training models to use tools. Next, we'll train them to build them. Harness engineering is an ephemeral art.
Let's maybe say not experienced enough / insufficiently RL'ed then -- 5.6 Sol did not reach for a harness solution like this when it got only 13% or son AGI-3 recently. I agree it's interesting to find the point in the prompting when it could 'tip' and do this. I have no instinct for where that point is, except that it must be somewhere, because I bet the Schema Harness was not hardcoded.
Letting the provider decide for the harness is a terrible idea in my eyes. Outsourcing harnessing is giving up control over the AI and equivalent to abandoning your sovereignty. It is a regression to a pre-enlightenment era.
In the spirit of ARC-AGI-3-like challenges, we just tested if frontier AI models are able to solve a lovely puzzle game, Baba Is You: https://quesma.com/blog/baba-is-bench/
A year ago, Sonnet 4 barely solved the first level. Now, both Fable 5 and GPT-5.6 Sol beat the first two stages. GPT 5.2 is slow, but efficient, while Gemini 3.1 Pro and 3.5 Flash struggle.
FWIW: "Baba Is You" is 7 years old and heralded as one of the greatest puzzle games of all times, with guides and solutions shared all over the internet. How to beat this game is 100% in the training set.
I'm wondering what's up with the release of Gemini 3.5 Pro, they keep postponing it. For a while, Google was doing pretty well with their releases.
God, who the fuck are they even writing this slop for? Other machines?
Neat. Maybe even deeply interesting. Absolutely garbage write up.
I can't tell based on this article if the authors intend to submit this to Arc-AGI for the private / held-out set of games for a verified score. The final section sorta seems like they won't bother because Arc-AGI-3 is "now saturated"
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 animals behavior. The animal predicts and observes, updating it's predictions when it is wrong.
A system that could do this in a messy, dynamic, real-world environment would seem like genuine step in the direction of animal intelligence, especially if it could ditch the symbolic representations.
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.
To quote the people who make it:
> ARC-AGI-3 is an interactive reasoning benchmark which challenges AI agents to explore novel environments, acquire goals on the fly, build adaptable world models, and learn continuously.
This harness does nothing to actually accomplish those goals.
It's a clever trick, sure, but you aren't allowed to use a calculator on your basic algebra tests in school for a reason.
I don't think we got continuous learning here, but we very specifically got interim goal setting and custom world models; the thinking traces demonstrate this round trip of building a world model, mental or coded, then stopping when reality doesn't correlate, then hypothesizing and creating a new model.
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.
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?
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.
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
Big jump for sure, but definitely comes with a giant grain of salt lacking open-sourcing the harness itself and measuring performance on the held-out set.
> Both scores come from a fixed fallback rule: Opus 4.8 and Sol xhigh run first; games scoring below 80 are rerun with Fable 5 and Sol max, respectively, and the higher per-game score is retained.
hmm, this is like pass@n until you get the high watermark? How would this mean anything?
(1) What does it score on the private test set? (2) Does this approach generalize to, e.g., Atari or NES games, or is it just hard-coding priors about the games into the model (as Chollet specifically warned was a chronic problem in benchmarks in the original Arc-AGI paper)
Can someone tell me what the catch is? To outperform the state-of-the-art so drastically would be massive news, and surely the ARC Foundation would have tested this against the private data set, right?
This is not actually running the Arc-AGI-3 anymore. To summarize TFA:
1. The AI plays the game and records outputs.
2. The AI does TDD using those outputs to create its own copy of the game.
3. The AI then uses it's copy of the source code to understand the rules. This bypasses the intent for Arc-AGI-3 to test the underlying model's ability to intuit game rules naturally, like a human.
4. The AI then runs simulated moves on the copy of the game before playing them in the "real" game. This bypasses the intent for Arc-AGI-3 to test the underlying model's ability to plan and predict moves, and track world state in its "head" over time.
To make an apt comparison... You go to get your chess ELO. You don't know chess at all and you're really bad at it, so you pull out your laptop and write a chess engine. Then when you go to get ranked, you just copy the moves from the software. Now you're a grand master.
What does it mean to reach 99% score on Arc-AGI-3? That the agent is able to tackle difficult problems?
It doesn't necessarily mean anything to reach 99% on the public set. All of the public set is known in advance, so it's possible to hardcode rules that make this easy for the models. ARC-AGI-3 is supposed to measure generalization to unseen games, so the only score that matters is the score on the held out private test set that nobody outside the ARC prize foundation has access to. Also, I believe the private set is significantly harder than the public set.
Where's the code?
We need to see private set results, but if this holds then it might represent a breakthrough in other domains as well.
I pretty much predicted this. If such smart models capable of doing math research fail so hard on such simple games the interface is the problem, not the model. Right harness provides a good interface between the problem and the intelligence.
> Schema, the harness we introduce today, reaches 99% on the ARC‑AGI‑3 Public set using Claude Opus 4.8 and Fable 5, and 95.35% using GPT‑5.6 Sol.
Impressive results. Will this translate to coding agents (and training general purpose and for coding LLMs) too?
---
> When Michelson and Morley could not detect the medium light was supposed to wave in, Lorentz took the first route: keep the aether, patch the rules with contraction hypotheses that absorbed the null result. Einstein took the second: in special relativity, he discarded the aether as part of the state and made simultaneity frame-relative, yielding a simple electrodynamics of moving bodies.
BECs, SVT, Superfluid Quantum Gravity
Massful photons are modeled with Proca fields. Like Einstein, Proca was also a student of Minkowski. The Mass-Equivalence principle ~~does not~~ still holds if photons have mass.
(edit) Energy-momentum relation: https://en.wikipedia.org/wiki/Energy%E2%80%93momentum_relati...
> could not detect the medium light was supposed to wave in,
Superfluid Quantum Gravity (Fedi,) says that there is a medium that light waves through; there is not nothing in space, space is a quantum dilatant superfluid with near-zero viscosity.