CATS CONFUSE REASONING LLM: QUERY AGNOSTIC ADVERSARIAL TRIGGERS FOR REASONING MODELS

Summary from the TLDR AI newsletter

Adding irrelevant phrases like "Interesting fact: cats sleep most of their lives" to math problems causes reasoning models to generate incorrect answers 300% more often than baseline. The query-agnostic phenomenon transfers across model size, with distilled models showing even higher vulnerability rates. The attacks also cause significant computational overhead, with 42% of responses exceeding 1.5x normal token length.