[arXiv 2026] COMIC: Beyond the "Dad Joke" — How Agentic Competition is Forging the Future of AI Comedy

TL;DR

Can AI truly be funny? While most LLMs are stuck in the realm of puns, a new framework called COMIC (Content Optimization via Multi-agent Iterative Competition) has bridged the gap. By abandoning single-pass generation in favor of a "survival of the funniest" evolutionary loop, COMIC produces full-length sketch comedy videos that rival professional human productions.

Background: Why Humor is the Final Frontier

In the world of AI, tasks like coding or math have clear "right" answers. Comedy, however, is a moving target. It is subjective, context-dependent, and loses its value upon repetition. Previous attempts at AI video generation (like MovieAgent or VGoT) often resulted in dry, narrative-less clips because they relied on a fixed list of instructions.

The authors of COMIC realized that great comedy isn't written; it's rewritten. They modeled their system after high-pressure writers' rooms (like Saturday Night Live), where ideas are constantly challenged, discarded, and refined.

Methodology: The "Island" of Misfit Scripts

The core innovation of COMIC is moving away from a single "God-mode" evaluator. Instead, it uses Island-based Evolution:

• Diverse Critics: The researchers analyzed nearly 5,000 YouTube videos to align LLM "critics" with real viewer engagement.
• Multi-Island Topology: Scripts are divided into different "islands," each governed by critics with different comedic philosophies (e.g., dry wit vs. slapstick).
• Semantic Crossover: Losing scripts are not just deleted; they are rewritten using feedback from the "winners," effectively performing a semantic mutation that carries forward the best comedic traits.

From Script to Screen: The Rendering Loop

Generating a funny script is only half the battle. Visualizing it requires extreme consistency. To solve this, COMIC introduces a History Tournament:

1. A Scene Director breaks the script into shots.
2. Rendering Critics evaluate each shot for character consistency and "visual artifacts."
3. The system tracks a history of shot versions and runs a single-elimination tournament to pick the best realization.

This "test-time scaling" allows the system to trade more GPU compute for higher quality, effectively "searching" the space of possible videos for the one that hits the comedic timing perfectly.

Experiments: Approaching Human Quality

In a blind study, COMIC didn't just beat LLM baselines—it crushed them. While earlier models were rated as "definitely not" worth re-watching, COMIC approached the "neutral to likely" range, a massive leap for fully automated content.

Table showing COMIC's superior Win Rate and Diversity across various channels like SNL and Key & Peele.

The system's ability to maintain Inter-Diversity (making many different kinds of funny videos) is credit to the island-based approach. As the generations progress, the "win rate" increases drastically, proving that the competitive pressure actually works to improve "fitness."

The "Red Queen" effect in action: Win rates climb as scripts evolve against increasingly harsh critics.

Critical Insights & The Future

COMIC proves that Iterative Competition > Single-Pass Logic. By embracing the "Red Queen Hypothesis"—where agents must constantly improve just to keep up with their rivals—we see a path toward AI that can handle open-ended, subjective creative work.

Limitations: The computational cost is still significant (about 24 hours on an H200 for a polished sketch), and relying on YouTube view counts as a proxy for humor can sometimes favor "clickbait" over high-brow comedy. However, as inference compute becomes cheaper, the "COMIC" approach could soon be producing our daily entertainment on demand.

Conclusion

COMIC isn't just about making people laugh; it's a blueprint for Agentic Orchestration. It demonstrates that when we treat AI agents not as simple tools, but as members of a competitive ecosystem, the emergent quality can surpass anything a single prompt could ever achieve.