Non Verbis, Sed Rebus: Large Language Models are Weak Solvers of Italian Rebuses

TL;DR

This paper evaluates large language models' ability to solve Italian rebuses, revealing their limitations and the tendency of fine-tuning to lead to memorization rather than genuine reasoning skills.

Contribution

It introduces a new Italian rebus dataset and systematically assesses LLMs' performance, highlighting their weaknesses and the impact of fine-tuning.

Findings

LLMs perform poorly on Italian rebuses

Fine-tuning improves performance but mainly causes memorization

Rebus solving is a challenging benchmark for linguistic reasoning

Abstract

Rebuses are puzzles requiring constrained multi-step reasoning to identify a hidden phrase from a set of images and letters. In this work, we introduce a large collection of verbalized rebuses for the Italian language and use it to assess the rebus-solving capabilities of state-of-the-art large language models. While general-purpose systems such as LLaMA-3 and GPT-4o perform poorly on this task, ad-hoc fine-tuning seems to improve models' performance. However, we find that performance gains from training are largely motivated by memorization. Our results suggest that rebus solving remains a challenging test bed to evaluate large language models' linguistic proficiency and sequential instruction-following skills.