In a debate test, 37% of AI agents changed answers with no peer

A test of multi-agent debate looked at some large language models. They give an answer to a question and read the answers of the other agents. The researchers used 5 GPT-4o agents on MMLU-Pro questions on which the agents did not agree. In 37% of agent-question pairs, the agent changed its answer with no peer information. The authors find that agreement alone is an unclear sign of value.

How it works

The researchers gave the agents one of 3 inputs: self-reflection (the agent thinks again with no peer), peer stances only, or full peer reasoning. They compared the 3 to find the increase from peer information.

The numbers

Some pairs did not change in the control test. Of these, 29.3% used a peer answer after they saw peer stances. These changes were corrections, new errors and changes between incorrect answers. Full reasoning gave no more majority-vote increase than self-reflection, and it used 57% more tokens.

What is not known

When a correct peer answer was available, correction was more frequent than dangerous adoption. Responses to constructed peer material differed across 3 models. A step that ignores the answer cut the good and the bad adoption, with no detectable precision increase. The source does not tell if other groups checked the work.

Insight: Agents change stance easily, so a change proves little

Earlier work found that 10 comments moved the stance of one agent by 0.171. This test adds that 37% of agents change answers with no peer. A change of stance in agent teams is easy to cause. It reads as a weak sign of quality. If you assemble agent teams, compare them with one agent that thinks again with no peer.