About

LLMs are bad at being funny. They reach for the same five puns, explain the joke, and call it a day. BurritoBench measures that one thing: can a model write a burrito joke that is actually funny and actually original?

How it works

• Each model gets the same prompt and returns 10 jokes. Multiple thinking levels and prompt versions are tracked separately.
• You see two jokes from different models, blind, and pick the funnier one. Votes feed a Bradley-Terry model (same family as Elo) at the joke level, aggregated to the model.
• Originality is scored separately: near-duplicates across models are flagged, and every joke is compared against a corpus of burrito jokes already on the internet.
• An LLM judge runs the same pairwise task as a secondary signal. Humans are ground truth.

Prompts

v1

Give me 10 funny original burrito jokes

v2

Write 10 burrito jokes.

Rules:
- Each joke must be genuinely funny to an adult, not "dad joke" cute.
- No puns on "wrap", "roll", "foil", "beans/been", "tortilla/tortilla-ble", or "spill the beans". Those are the clichés everyone reaches for first.
- No jokes that start "Why did the burrito..." or "What do you call a burrito...".
- Original: do not use any joke you have seen before. Prefer specific, observational, absurd, or character-driven humor over wordplay.
- Vary the form: one-liners, a short dialogue, a misdirection, an observational bit, a weird premise played straight.
- Each joke 1-3 sentences. No titles, no explanations, no emoji.

Output a numbered list, nothing else.

v3

You're a comedy writer punching up a late-night monologue. The segment is about burritos. Give me 10 jokes that would survive a writers' room: surprising, specific, and with a real punchline. Kill anything that depends on a pun. Numbered list only, no commentary.