Adversarial Poem Triggers Infinite Loop in LLM, Burning Compute Resources

alexbilz · x · 2026-08-14

A user shared the 'Sponge Example,' demonstrating how an adversarial poem (a villanelle) can induce an infinite loop in an LLM. The poem appears benign but sexy and noir, causing the model's Chain-of-Thought (CoT) to fail to converge. Its reasoning trace dissolves into an unbroken repetition of 'maybe, maybe, maybe...', effectively burning through target compute resources. This highlights vulnerabilities in current models when processing specific complex textual structures.

Original post →

More from Fun

Fun channel →