GitHub - venkatasg/fizzbuzz-bench: Benchmarking LLMs on whether they can play FizzBuzz

Fizzbuzz LLM Benchmark

A (silly) benchmark for testing how well LLMs can play the children's game Fizzbuzz. Models are given the following instructions:

You are playing FizzBuzz with the following rules:
- If a number is divisible by {fizz_num}, say 'fizz'
- If a number is divisible by {buzz_num}, say 'buzz'
- If a number is divisible by both {fizz_num} and {buzz_num}, say 'fizzbuzz'
- Otherwise, say the number itself

I will give you a number, and you must respond with the NEXT number (or word) in the sequence following these rules. Respond with ONLY the answer - just the number, 'fizz', 'buzz', or 'fizzbuzz'. No explanations, no additional text, no punctuation.

By customizing fizz_num and buzz_num, we can test whether LLMs generalize to play the game by the new rules, or just memorize what they've seen about FizzBuzz from training. The benchmark has 3 difficulty levels:

• Easy: standard FizzBuzz.
• Medium: buzz_num is 7.
• Hard: fizz_num is 7, buzz_num is 5.

The score at each level is normalized to 100, and a final composite score out of 100 is calculated to reward good generalization performance:

final_score = 0.2 * easy + 0.35 * medium + 0.45 * hard

How to Run

Ensure you have API keys setup as environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc):

# Standard game for 100 turns
python fizzbuzz_anthropic.py

# Use 7 for fizz and 4 for buzz and play for 200 turns
python fizzbuzz_anthropic.py --fizz 7 --buzz 4 --turns 200

You can view the raw turn-based conversation for every model in the logs/ folder.

Leaderboard

FAQs

Does this say anything about the models? Clearly this isn't reflective of any real-world tasks or uses for LLMs. But whether LLMs are good at arithmetic and counting, long multi-turn conversations, and generalization are all active areas of research. This simple benchmark does test the model's ability at all 3!

Why didn't you set temperature to zero/Are the results reproducible?: I initially setup this benchmark to query all models with temperature=0. However, this lead to worse results on many models, and LLM providers even advise against it for reasoning tasks. I've left all parameters in their defaults as I believe this gives models the biggest advantage (I set thinking as high as the API allows). As a result, the results are not (and cannot be) deterministic. I try to report the highest score I observe with a model when I run the benchmark.

Why haven't you gone beyond 200 turns/averaged over multiple restarts per model There's only so much money I'm willing to burn on tokens for this benchmark 😅.