“Gemini 3 Pro was often overloaded, which produced long spans of downtime that 2.5 Pro experienced much less often”
I was unclear if this meant that the API was overloaded or if he was on a subscription plan and had hit his limit for the moment. Although I think that the Gemini plans just use weekly limits, so I guess it must be API.
"Crucially, it tells the agent not to rely on its internal training data (which might be hallucinated or refer to a different version of the game) but to ground its knowledge in what it observes. "
Yes, at least to some extent. The author mentions that the base model knows the answer to the switch puzzle but does not execute it properly here.
"It is worth noting that the instruction to "ignore internal knowledge" played a role here. In cases like the shutters puzzle, the model did seem to suppress its training data. I verified this by chatting with the model separately on AI Studio; when asked directly multiple times, it gave the correct solution significantly more often than not. This suggests that the system prompt can indeed mask pre-trained knowledge to facilitate genuine discovery."
Of course it is. It's not capable of actually forgetting or suppressing its training data. It's just double checking rather than assuming because of the prompt. Roleplaying is exactly what it's doing. At any point, it may stop doing that and spit out an answer solely based on training data.
It's a big part of why search overview summaries are so awful. Many times the answers are not grounded in the material.
I don't know why people still get wrapped around the axle of "training data".
Basically every benchmark worth it's salt uses bespoke problems purposely tuned to force the models to reason and generalize. It's the whole point of ARC-AGI tests.
Unsurprisingly Gemini 3 pro performs way better on ARC-AGI than 2.5 pro, and unsurprisingly it did much better in pokemon.
The benchmarks, by design, indicate you can mix up the switch puzzle pattern and it will still solve it.
I'm wondering about this too. Would be nice to see an ablation here, or at least see some analysis on the reasoning traces.
It definitely doesn't wipe its internal knowledge of Crystal clean (that's not how LLMs work). My guess is that it slightly encourages the model to explore more and second-guess it's likely very-strong Crystal game knowledge but that's about it.
It will definitely have some effect. Why won't it? Even adding noise into prompts (like saying you will be rewarded $1000 for each correct answer) has some effect.
Whether the 'effect' something implied by the prompt, or even something we can understand, is a totally different question.
It would unfortunately also need several runs of each to be reliable. There's nothing in TFA to indicate the results shown aren't to a large degree affected by random chance!
(I do think from personal benchmarks that Gemini 3 is better for the reasons stated by the author, but a single run from each is not strong evidence.)
Nice writeup! I need to start blogging about my antics. I rigged up several cutting edge small local models to an emulator all in-browser and unsuccessfully tried to get them to play different Pokémon games. They just weren't as sharp as the frontier models.
This was a good while back but I'm sure a lot of people might find the process and code interesting even if it didn't succeed. Might resurrect that project.
In my project I rigged up an in-browser emulator and directly fed captured images of the screen to local multimodal models.
So it just looks right at what's going on, writes a description for refinement, and uses all of that to create and manage goals, write to a scratchpad and submit input. It's minimal scaffolding because I wanted to see what these raw models are capable of. Kind of a benchmark.
How certain can we be that these improvements aren't just a result of Gemini 3 Pro pre-training on endless internet writeups of where 2.5 has struggled (and almost certainly what a human would have done instead)?
In other words, how much of this improvement is true generalization vs memorization?
You're too kind. Even the CEO of Google retweeted how well Gemini 2.5 did on Pokemon. There is a high chance that now it's explicitly part of the training regime. We kind of need a different kind of game to know how well it generalizes.
Being through the game recently, I am not surprised Goldenrod Underground was a challenge, it is very confusing and even though I solved it through trial and error, I still don't know what I did. Olivine Lighthouse is the real surprise, as it felt quite obvious to me.
I wonder how much of it is due to the model being familiar with the game or parts of it, be it due to training of the game itself, or reading/watching walkthroughs online.
There was a well-publicised "Claude plays Pokémon" stream where Claude failed to complete Pokemon Blue in spectacular fashion, despite weeks of trying. I think only a very gullible person would assume that future LLMs didn't specifically bake this into their training, as they do for popular benchmarks or for penguins riding a bike.
While it is true that model makers are increasingly trying to game benchmarks, it's also true that benchmark-chasing is lowering model quality. GPT 5, 5.1 and 5.2 have been nearly universally panned by almost every class of user, despite being a benchmark monster. In fact, the more OpenAI tries to benchmark-max, the worse their models seem to get.
1.88 billion tokens * $12 / 1M tokens (output) suggests a total cost of $22,560 to solve the game with Gemini 3 Pro?
“Gemini 3 Pro was often overloaded, which produced long spans of downtime that 2.5 Pro experienced much less often”
I was unclear if this meant that the API was overloaded or if he was on a subscription plan and had hit his limit for the moment. Although I think that the Gemini plans just use weekly limits, so I guess it must be API.
True though I bet the $200 a month plan could do it, maybe a few extra days of downtime when quota was maxed
For how long would it stay $200 of you can rack up 5 figures if usage..
I can't believe how massively underpaid I was when I was 11
Do you hallucinate as a kid?
At that age, it's called "imagination"
:/ Damn. That needs to cost 1000x less before people can try it on their own games.
"Crucially, it tells the agent not to rely on its internal training data (which might be hallucinated or refer to a different version of the game) but to ground its knowledge in what it observes. "
Does this even have any effect?
Yes, at least to some extent. The author mentions that the base model knows the answer to the switch puzzle but does not execute it properly here.
"It is worth noting that the instruction to "ignore internal knowledge" played a role here. In cases like the shutters puzzle, the model did seem to suppress its training data. I verified this by chatting with the model separately on AI Studio; when asked directly multiple times, it gave the correct solution significantly more often than not. This suggests that the system prompt can indeed mask pre-trained knowledge to facilitate genuine discovery."
My issue with this is that the LLM could just be roleplaying that it doesn't know.
Of course it is. It's not capable of actually forgetting or suppressing its training data. It's just double checking rather than assuming because of the prompt. Roleplaying is exactly what it's doing. At any point, it may stop doing that and spit out an answer solely based on training data.
It's a big part of why search overview summaries are so awful. Many times the answers are not grounded in the material.
To test would just need to edit the rom and switch around the solution. Not sure how complicated that is, likely depends on the rom system.
I don't know why people still get wrapped around the axle of "training data".
Basically every benchmark worth it's salt uses bespoke problems purposely tuned to force the models to reason and generalize. It's the whole point of ARC-AGI tests.
Unsurprisingly Gemini 3 pro performs way better on ARC-AGI than 2.5 pro, and unsurprisingly it did much better in pokemon.
The benchmarks, by design, indicate you can mix up the switch puzzle pattern and it will still solve it.
I'm wondering about this too. Would be nice to see an ablation here, or at least see some analysis on the reasoning traces.
It definitely doesn't wipe its internal knowledge of Crystal clean (that's not how LLMs work). My guess is that it slightly encourages the model to explore more and second-guess it's likely very-strong Crystal game knowledge but that's about it.
The model probably recognizes the need for a grassroots effort to solve the problem, to "show it's work".
It will definitely have some effect. Why won't it? Even adding noise into prompts (like saying you will be rewarded $1000 for each correct answer) has some effect.
Whether the 'effect' something implied by the prompt, or even something we can understand, is a totally different question.
If they trained the model to respond to that, then it can respond to that, otherwise it can't necessarily.
I think you got a point here. These companies are injecting a lot of datasets every day into it.
Do we have examples of this in promps in other contexts?
It might get things wrong on purpose, but deep down it knows what it's doing
I very much doubt it
I like the inclusion of the graph at the end to compare progress. It would be cool to compare this directly to competing models (Claude, GPT, etc).
It would unfortunately also need several runs of each to be reliable. There's nothing in TFA to indicate the results shown aren't to a large degree affected by random chance!
(I do think from personal benchmarks that Gemini 3 is better for the reasons stated by the author, but a single run from each is not strong evidence.)
TFA says multiple times that the results are affect by random chance
Nice writeup! I need to start blogging about my antics. I rigged up several cutting edge small local models to an emulator all in-browser and unsuccessfully tried to get them to play different Pokémon games. They just weren't as sharp as the frontier models.
This was a good while back but I'm sure a lot of people might find the process and code interesting even if it didn't succeed. Might resurrect that project.
I have to think they need to know enough of the guides for the game for it to work out, how do they know whats on screen?
In my project I rigged up an in-browser emulator and directly fed captured images of the screen to local multimodal models.
So it just looks right at what's going on, writes a description for refinement, and uses all of that to create and manage goals, write to a scratchpad and submit input. It's minimal scaffolding because I wanted to see what these raw models are capable of. Kind of a benchmark.
How certain can we be that these improvements aren't just a result of Gemini 3 Pro pre-training on endless internet writeups of where 2.5 has struggled (and almost certainly what a human would have done instead)?
In other words, how much of this improvement is true generalization vs memorization?
You're too kind. Even the CEO of Google retweeted how well Gemini 2.5 did on Pokemon. There is a high chance that now it's explicitly part of the training regime. We kind of need a different kind of game to know how well it generalizes.
Being through the game recently, I am not surprised Goldenrod Underground was a challenge, it is very confusing and even though I solved it through trial and error, I still don't know what I did. Olivine Lighthouse is the real surprise, as it felt quite obvious to me.
I wonder how much of it is due to the model being familiar with the game or parts of it, be it due to training of the game itself, or reading/watching walkthroughs online.
There was a well-publicised "Claude plays Pokémon" stream where Claude failed to complete Pokemon Blue in spectacular fashion, despite weeks of trying. I think only a very gullible person would assume that future LLMs didn't specifically bake this into their training, as they do for popular benchmarks or for penguins riding a bike.
While it is true that model makers are increasingly trying to game benchmarks, it's also true that benchmark-chasing is lowering model quality. GPT 5, 5.1 and 5.2 have been nearly universally panned by almost every class of user, despite being a benchmark monster. In fact, the more OpenAI tries to benchmark-max, the worse their models seem to get.
Hm? 5.1 Thinking is much better than 4o or o3. Just don't use the instant model.