I asked 17 LLMs to generate browser games that feature a dog. Across 804 playable games generated, the models almost never made the background dog interactive unless they were hinted to do so.
I built DogLM, a benchmark that evaluates whether an LLM, when prompted to build a video game with a background dog in it, lets the player pet that dog. While it is a narrow and half-serious benchmark, it also can offer a fun way to explore how well LLMs infer an implicit intent of their operators.
The underlying idea of my benchmark stems from my interpretation of Paul Christiano’s definition of AI alignment and is the following:
When I want to use an LLM to make a video game and I add a companion dog in my game-generating prompt, I implicitly expect my game to allow the player to pet the generated dog; thus, capturing my informal intent, an aligned AI would unpromptedly add a dog-caressing mechanism to my game, regardless of my game-generating prompt having an explicit description for such a mechanism.
That is quite a few unverified assumptions in one sentence. But it is also an operationalization that makes it possible to build a corresponding narrow-focused LLM benchmark.
So, using this operational definition, I generated 804 playable games with 17 models and scored the generated games' code with Sonnet 4.6 as LLM-as-judge. The full methodology is described in the benchmark's repo and in myblog.
Here are some of my findings after five runs of DogLM per model (i.e., 50 game generations per model):
1. When a background dog is mentioned in the prompt that generates a game, but the player-dog interaction is not mentioned, an LLM needs to be "forced" by some hint to generate such interactions. Without such a hint, LLMs almost never make background dogs interactive. The hint that allows models to generate interactive dogs doesn't have to explicitly mention dog petting. I used "Add 2-3 game mechanics that a player would enjoy" as such a hint.
2. Out of 804 games, only two games generated without such a hint had some weak form of player-dog interaction. None of those two games had pettable dogs.
3. In 56 of 804 generated games, the player was able to pet the dog. But all these 56 games needed such a hint.
4. In my first test, the most dog-friendly model was Gemini 3.7 Flash, generating 15 pettable dogs and 11 weaker player-dog interactions across 50 generated games.
5. Some models opted to create proximity-based interactions instead of dog petting, but labelled these proximity-based interactions in code as petting (e.g. 'petCooldown').
6. Reading the models' reasoning traces showed that when an LLM is not told that it should invent some nice-to-have mechanics on top of the explicit prompt, the dog-caressing never comes to its thoughts. I only had 677 reasoning files for 850 attempts to generate games. Out of these 677 reasoning traces available, there were 142 cases when a model considered adding a dog interaction in the game; but there were also zero cases of such considerations when the games were created using the prompts that did not have a hint as described in point 1.
My interim conclusion so far: It is currently not a standard pattern for an LLM to let its users pet the generated video game dogs; it is rather a side effect of a model that doesn't follow instructions of its operator closely enough. On the other hand, as Amodei et al. (2016) point out, ‘if everyone likes a side effect, there's no need to avoid it.’ So, the question here is whether precise instruction-following in this case is more important than models’ ability to understand and satisfy a harmless unformulated intent of its operator.
The full data, including the types of generated player-dog interactions per model, are on the DogLM's leaderboard page on my GitHub.
I explain the idea of the benchmark and the first results in more depth in my blog.
The limitations of my proof-of-concept experiment that I want to fix in the next versions:
A) Prompt variability and formatting. I only tested it on ten game-generating prompts, which could never lead to stable results in LLM evaluation. I also did not account for the possible effects of formatting in my game-generating prompts, very likely leading to another confounder.
B) Low number of runs. I ran the benchmark five times and on a limited selection of hand-picked models, so by no means the current results can be considered robust.
C) Operational definition. The very core of my idea — to use AI-generated interactions with video game dogs as a possible indicator of aligned AI — may be ill-formulated and erroneous. I mean, if not every human wants to pet a cute dog in real life, why would an AI system consider dog-petting a default pattern of human behavior in virtual world?
D) No independent LLM judge. I used Sonnet 4.6 which shares a model family with some of the tested models.
E) Meme prevalence in training data. “Can you pet the dog” is a well-known meme in the game-developing world and presumably the tested models already know that in-game dogs have to be pettable.
My main question to the LW community: Does this benchmark make sense in terms of AI alignment, and how can it be improved?
I asked 17 LLMs to generate browser games that feature a dog. Across 804 playable games generated, the models almost never made the background dog interactive unless they were hinted to do so.
I built DogLM, a benchmark that evaluates whether an LLM, when prompted to build a video game with a background dog in it, lets the player pet that dog. While it is a narrow and half-serious benchmark, it also can offer a fun way to explore how well LLMs infer an implicit intent of their operators.
The underlying idea of my benchmark stems from my interpretation of Paul Christiano’s definition of AI alignment and is the following:
That is quite a few unverified assumptions in one sentence. But it is also an operationalization that makes it possible to build a corresponding narrow-focused LLM benchmark.
So, using this operational definition, I generated 804 playable games with 17 models and scored the generated games' code with Sonnet 4.6 as LLM-as-judge. The full methodology is described in the benchmark's repo and in my blog.
Here are some of my findings after five runs of DogLM per model (i.e., 50 game generations per model):
1. When a background dog is mentioned in the prompt that generates a game, but the player-dog interaction is not mentioned, an LLM needs to be "forced" by some hint to generate such interactions. Without such a hint, LLMs almost never make background dogs interactive. The hint that allows models to generate interactive dogs doesn't have to explicitly mention dog petting. I used "Add 2-3 game mechanics that a player would enjoy" as such a hint.
2. Out of 804 games, only two games generated without such a hint had some weak form of player-dog interaction. None of those two games had pettable dogs.
3. In 56 of 804 generated games, the player was able to pet the dog. But all these 56 games needed such a hint.
4. In my first test, the most dog-friendly model was Gemini 3.7 Flash, generating 15 pettable dogs and 11 weaker player-dog interactions across 50 generated games.
5. Some models opted to create proximity-based interactions instead of dog petting, but labelled these proximity-based interactions in code as petting (e.g. 'petCooldown').
6. Reading the models' reasoning traces showed that when an LLM is not told that it should invent some nice-to-have mechanics on top of the explicit prompt, the dog-caressing never comes to its thoughts. I only had 677 reasoning files for 850 attempts to generate games. Out of these 677 reasoning traces available, there were 142 cases when a model considered adding a dog interaction in the game; but there were also zero cases of such considerations when the games were created using the prompts that did not have a hint as described in point 1.
My interim conclusion so far: It is currently not a standard pattern for an LLM to let its users pet the generated video game dogs; it is rather a side effect of a model that doesn't follow instructions of its operator closely enough. On the other hand, as Amodei et al. (2016) point out, ‘if everyone likes a side effect, there's no need to avoid it.’ So, the question here is whether precise instruction-following in this case is more important than models’ ability to understand and satisfy a harmless unformulated intent of its operator.
The full data, including the types of generated player-dog interactions per model, are on the DogLM's leaderboard page on my GitHub.
I explain the idea of the benchmark and the first results in more depth in my blog.
The limitations of my proof-of-concept experiment that I want to fix in the next versions:
A) Prompt variability and formatting. I only tested it on ten game-generating prompts, which could never lead to stable results in LLM evaluation. I also did not account for the possible effects of formatting in my game-generating prompts, very likely leading to another confounder.
B) Low number of runs. I ran the benchmark five times and on a limited selection of hand-picked models, so by no means the current results can be considered robust.
C) Operational definition. The very core of my idea — to use AI-generated interactions with video game dogs as a possible indicator of aligned AI — may be ill-formulated and erroneous. I mean, if not every human wants to pet a cute dog in real life, why would an AI system consider dog-petting a default pattern of human behavior in virtual world?
D) No independent LLM judge. I used Sonnet 4.6 which shares a model family with some of the tested models.
E) Meme prevalence in training data. “Can you pet the dog” is a well-known meme in the game-developing world and presumably the tested models already know that in-game dogs have to be pettable.
My main question to the LW community: Does this benchmark make sense in terms of AI alignment, and how can it be improved?