Debate May Help AI Models Converge on Truth | Quanta Magazine

Two large AI models debate the answer to a given question, with a simpler model left to recognize the more accurate answer.

In theory, the two agents poke holes in each other’s arguments until the judge has enough information to discern the truth.

Debate emerged as a possible approach in 2018 , before LLMs became as large and ubiquitous as they are today .

A small subset of computer scientists and linguists soon began to look for the benefits of debate.

They found examples where it didn’t help humans, but there were hints that they could help language models.

In a 2023 paper, researchers reported that when multiple copies of an LLM were allowed to debate and converge on an answer, rather than convince a judge, they were more accurate.

Debate requires the debaters to be better than the judge, but “better” will depend on the task.

In tests, that’s knowledge; in tasks that require reasoning or reasoning, that dimension may be different.

Finding scalable oversight solutions is a critical open challenge in AI safety right now.