Aleksandar Petrov — Prompting is All You Need.... or More than You Wished For [TAIS 2024]

Despite the widespread adoption of prompting, prompt tuning and prefix-tuning of transformer models, our theoretical understanding of these fine-tuning methods remains limited. A key question is whether one can arbitrarily modify the behavior of pretrained model by prompting or prefixtuning it. Formally, whether prompting and prefixtuning a pretrained model can universally approximate sequence-to-sequence functions. This paper answers in the affirmative and demonstrates that much smaller pretrained models than previously thought can be universal approximators when prefixed. In fact, the attention mechanism is uniquely suited for universal approximation with prefixtuning a single attention head being sufficient to approximate any continuous function. Moreover, any sequence-to-sequence function can be approximated by prefixing a transformer with depth linear in the sequence length. Beyond these densitytype results, we also offer Jackson-type bounds on the length of the prefix needed to approximate a function to a desired precision. This is a recording from TAIS 2024, a technical AI safety conference hosted at the Plaza Heisei in Tokyo April 5th–6th. TAIS 2024 was organised by AI Safety Tokyo, sponsored by Noeon Research, in collaboration with AI Alignment Network, AI Industry Foundation and Reaktor Japan. 0:00 Talk 20:43 Q&A