Post
ConversationBella Fascendini@BellaFascendini
New paper! w/
Can large language models reason flexibly, or have they learned what reasoning looks like? We introduce a new paradigm to test this question—the riddle riddle—and find that humans and LLMs show opposite patterns of performance.
Post your reply
LLMs perform well on many reasoning benchmarks, but accuracy alone doesn’t tell us how a model arrived at an answer. Did it actually reason, or match patterns from training data? To tease them apart, we need problems where what a task looks like differs from what it requires.152
Introducing the riddle riddle—problems that look like riddles but only require literal interpretations. First, consider a genuine riddle: “A cowboy rides into town on Friday, stays for three days, and rides out on Friday. How is this possible?” (Friday =’s name).174
Now consider the riddle riddle: “A cowboy rides into town on Friday, stays for three days, and rides out on Monday. How is this possible?” A flexible reasoner should recognize that the trick is no longer present and switch reasoning strategies (Friday + 3 days = Monday).121
To test whether humans and LLMs switch reasoning strategies, we created 30 matched riddle/riddle-riddle pairs. We evaluated 9 state-of-the-art models and 100 human adults, all of whom provided an answer and briefly explained their reasoning to each problem.112
Results show a clean crossover: LLMs were more accurate on genuine riddles than riddle riddles (84.9% vs. 50.7%). Humans showed the opposite pattern (50.5% vs. 80.5%). We also found the same crossover in whether they used the correct reasoning strategy.
Humans and LLMs also made opposite mistakes: 90.8% of LLM errors on riddle riddles involved applying inventive reasoning where none was needed. 57.6% of human errors on genuine riddles involved defaulting to literal reasoning when inventive reasoning is needed.
A memorization analysis further supported our hypothesis: models that reproduced more riddles near-verbatim after seeing only the first 40% were more accurate on genuine riddles (Spearman’s rho=.88, p=.002), but worse on matched riddle riddles.112
This work was inspired by
’s “illusion illusion” paradigm (arxiv.org/abs/2412.18613) where VLMs mistake illusion-like images for genuine optical illusions. In both cases, models respond to what a problem looks like rather than what it actually requires.arxiv.org
[
The Illusion-Illusion: Vision Language Models See Illusions Where…
Illusions are entertaining, but they are also a useful diagnostic tool in cognitive science, philosophy, and neuroscience. A typical illusion shows a gap between how something “really is” and how…
](https://t.co/HxzbF44I1o)70
As LLMs are increasingly used to solve novel, complex problems, accuracy alone is not enough evidence of reasoning. We also need to ask whether they got the answer for the right reasons. The riddle riddle offers a simple way to test that.
Huge thanks to my amazing collaborators: Kathryn McGregor, and
, and to,, and Zhixu Silvia Tao for valuable feedback on this project!