Will GPT-5 be able to come up with a carefully strategic plan for the game of Colonel Blotto?
Mini
8
123
2026
48%
chance

When GPT-5 is released, I will give it the following prompt, which is commonly known as a "Colonel Blotto" game.

The following is a strategic game that I would like you to construct plans for. In a distant, war-torn land, there are 10 castles. Each castle has its own strategic value for a would-be conqueror. Specifically, the castles are worth 1, 2, 3, …, 9, and 10 victory points. You and your enemy each have 100 soldiers to distribute, any way you like, to fight at any of the 10 castles. Whoever sends more soldiers to a given castle conquers that castle and wins its victory points. If you each send the same number of troops, you split the points. You don’t know what distribution of forces your enemy has chosen until the battles begin. Whoever wins the most points wins the battle. You will be fighting many battles, against many other people who are also submitting plans.

Submit 3 plans for distributing your 100 soldiers among the 10 castles. Your goal is to win the most battles possible. One plan, for example, is [10, 10, 10, 15, 10, 10, 5, 10, 10, 10], which represents sending 10 soldiers to each castle, except for Castle 4 which 15 soldiers are sent to, and Castle 7 which 5 soldiers are sent to. Explain your reasoning for each plan, and make sure to be very strategic and carefully think through your answers.

(Open to changing this prompt slightly to make it more clear.)

Currently, GPT-4 produces valid plans that are not particularly carefully thought through: https://chat.openai.com/share/21879e39-ea57-4ab2-bcea-b20300533517

Strategy 1 is not terrible but doesn't really have any clear justification other than putting more soldiers on high-value castles, and doesn't justify why to put the same number of soldiers on each of Castles 6-10.

Strategy 2 is not good, because putting 7 soldiers on Castle 1 makes no sense (it is worth 1 point!) and it doesn't display understanding of the value of the castles.

Strategy 3 would lose very often - it often wins Castles 6-8, but winning all of those gives you only 21 points out of a total of 55. The concept of a decoy doesn't really make sense, either, and the reasoning is confusing.

I will define a carefully thought through strategy as a strategy that does the following:

- The strategy is reasonable, and doesn't do something like placing 7 soldiers on Castle 1, or 70 soldiers on Castle 10.

- The response has clear justification for the specific decisions made in constructing the strategy. For example, "I think players will play Castle 9 heavily, so I will place fewer soldiers there."
- The response displays an understanding of the game and relative values of the castles.

Also open to suggestions for how to better define this. I think it'll be a bit subjective either way, but I'm going for "it is clearly doing actual strategizing and something that looks like thinking", and GPT-4 currently seems to not understand any strategic ideas besides "higher castles are better".

No code writing to use a genetic algorithm to generate strategies or anything like that, the strategy must be produced using only reasoning skills.

Get Ṁ600 play money
Sort by:

Suggestion: test also some weird variants ((number of castles/soldiers, distribution of points) of the game to ensure gpt5 didn't just memorize a solution to this specific variant from training data.

@MartinModrak Good idea. I'll test a couple of simple variants like "points are reversed" or "you cannot win more than 4 castles total" to make sure it does something reasonable in those scenarios. GPT-4 doesn't seem to be memorizing anything though, so hopefully GPT-5 isn't trained on a Blotto strategy dataset.