More AI agents do not decrease hallucination in clinical tasks

A research paper tested if teams of AI agents decrease incorrect statements, called hallucination, in clinical tasks. The authors compared four systems, one single model and three multi-agent set-ups, with four models and three tasks. Of 36 tests, 12 gave a clear result: four showed less hallucination and eight showed more. The authors say the model sets the result, not the number of agents.

How it works

The authors gave a score for hallucination in the written reasoning, and not only in the last answer. They say a doctor acts on the reasoning. They used a scheme that rates facts and faithfulness.

The numbers

Of 36 tests, 12 gave a clear result. Four showed a decrease and eight showed an increase. No model showed a decrease and an increase.

In a self-correction test, each model gave less hallucination when it saw its own flagged errors. Llama-3.2-3B-Instruct did not.

What is not known

The authors found no set-up that gives less hallucination in all conditions. They say collaboration moves around the work that the models can do, and does not add work that they cannot do. No other group has checked the result.

Insight: A correct answer can hide incorrect reasoning

In October 2026, Artificial Analysis said that more than 60% of legal results that are correct without a hallucination check contain a material hallucination. This paper finds that teams of agents increase task scores but not always trust. Tests that score only the last answer help sellers, and control seems to go to who measures the reasoning. Tell the seller to give that score.