On OOD dataset StyleMirror, we find that LLM-Judge novelty scores are associated with expert preferences to a larger extent than a previously proposed n-gram novelty metric, Creativity Index, suggesting our operationalization yields a more aligned metric for textual creativity.

Writing quality reward model scores are associated with both creativity and pragmaticality judgements, but are not interpretable. LLM-judge can replicate some expert novelty judgements but struggle with identifying non-pragmatic expressions.

In a follow-up study with GPT-5 and Claude, we observe that the rate of human-judged creative expressions in AI-written text is significantly lower than in human-written text.

Further, we find that both open source models tested, OLMo-1 and 2 of 7B and 32B size, exhibit a negative relationship between n-gram novelty and pragmaticality. As open-source LLMs try to generate text not present in data, their expressions tend to make less sense in context.

N-gram novelty is not a reliable metric of creativity: over *90%* of top-quartile n-gram novelty expressions were not judged as creative. We find many examples of low n-gram novelty expressions rated creative and high n-gram novelty expressions rated as non-pragmatic.

We recruit expert writers with MFA/MA/PhD background. They rated expressions in human- and AI-generated (from fully (code + DATA) open-source OLMo models) passages for if they make sense, are pragmatic, and are novel; they could also highlight any creative expressions.

The standard definition of creativity states the product has to be both novel AND appropriate. Similarly, we operationalize textual creativity as human-judged expression novelty AND sensicality (making sense by itself) + pragmaticality (making sense in context).

N-gram novelty is widely used as a measure of creativity and generalization. But if LLMs produce highly n-gram novel expressions that don’t make sense or sound awkward, should they still be called creative? In a new paper, we investigate how n-gram novelty relates to creativity.

Even powerful models achieve only 50% explanation adequacy rate, suggesting difficulties in reasoning about figurative inputs. Hallucination & unsound reasoning are the most prominent error categories.

Our main results are:
1. VLMs struggle to generalize from literal to figurative meaning understanding (training on e-ViL only achieves random F1 on our task)
2. Figurative meaning in the image is harder to explain compared to when it is in the text
3. VLMs benefit from image data during fine-tuning

Via human-AI collaboration, we augment existing datasets for multimodal metaphors, sarcasm, and humor with entailed/contradicted captions and textual explanations. The figurative part can be in the image, caption, or both. We benchmarks a variety of models on the resulting data.

We frame the multimodal figurative meaning understanding problem as an explainable visual entailment task between an image (premise) and its caption (hypothesis). The VLM predicts whether the image entails or contradicts the caption, and shows the reasoning steps in a textual explanation.

Can vision-language models understand figurative meaning in multimodal inputs, like visual metaphors, sarcastic captions or memes? Come find out at our #NAACL2025 poster on Friday at 9am!

New task & dataset of images and captions with figurative phenomena like metaphor, idiom, sarcasm, and humor.