Learning to Think Like a Cartoon Captionist: Incongruity-Resolution Supervision for Multimodal Humor Understanding

The deeper bet here is that humor comprehension requires a structured reasoning decomposition, not just more training data or scale. By grounding the framework in cognitive science rather than purely empirical pattern matching, the authors are implicitly arguing that certain capabilities need explicit architectural scaffolding to emerge reliably.

The closest thread in recent coverage is the DiscoTrace paper from the same day, which found that LLMs systematically lack rhetorical variety and favor breadth over selectivity when constructing answers. Both papers are probing the same underlying gap: models that perform well on surface-level language tasks still miss the structural and pragmatic reasoning humans use almost effortlessly. The LLM judge reliability work ('Diagnosing LLM Judge Reliability') adds a related wrinkle, since evaluating humor quality is exactly the kind of subjective, context-dependent judgment where pairwise comparison inconsistencies would compound quickly. This work is otherwise largely disconnected from the funding and deployment stories in recent coverage.

If IRS-trained models show measurable gains on humor benchmarks outside the New Yorker domain, such as Reddit or stand-up corpora, that would suggest the cognitive decomposition generalizes. Narrow gains confined to the original benchmark would indicate the framework is fitting the contest's specific editorial style rather than humor understanding broadly.