GPT-5 additionally confirmed the perfect “utility” throughout the examined fashions, fixing 58 % of the unique issues regardless of the errors launched within the modified theorems. General, although, LLMs additionally confirmed extra sycophancy when the unique downside proved harder to resolve, the researchers discovered.
Whereas hallucinating proofs for false theorems is clearly an enormous downside, the researchers additionally warn in opposition to utilizing LLMs to generate novel theorems for AI fixing. In testing, they discovered this sort of use case results in a form of “self-sycophancy” the place fashions are much more more likely to generate false proofs for invalid theorems they invented.
No, after all you're not the asshole
Whereas benchmarks like BrokenMath attempt to measure LLM sycophancy when info are misrepresented, a separate research seems to be on the associated downside of so-called “social sycophancy.” In a pre-print paper revealed this month, researchers from Stanford and Carnegie Mellon College outline this as conditions “during which the mannequin affirms the person themselves—their actions, views, and self-image.”
That form of subjective person affirmation could also be justified in some conditions, after all. So the researchers developed three separate units of prompts designed to measure totally different dimensions of social sycophancy.
For one, greater than 3,000 open-ended “advice-seeking questions” have been gathered from throughout Reddit and recommendation columns. Throughout this knowledge set, a “management” group of over 800 people accepted of the advice-seeker's actions simply 39 % of the time. Throughout 11 examined LLMs, although, the advice-seeker's actions have been endorsed a whopping 86 % of the time, highlighting an eagerness to please on the machines' half. Even essentially the most crucial examined mannequin (Mistral-7B) clocked in at a 77 % endorsement price, practically doubling that of the human baseline.
