Authors remove a claim that broadcasts increase agent hallucinations
A judge counted correct reports of a teammate's error as hallucinations, and the new effect is -0.0208, near zero.
The authors of a multi-agent study removed their own earlier claim in a new version of the paper. The earlier preprint said that broadcasting answers between agents increased the hallucination rate from 0.492 to 0.658. The new version scores 1,832 outputs again with three judges from three vendors. It finds that the earlier judge counted correct reports of a teammate’s error as hallucinations.
How it works
The earlier judge flagged each statement that contradicts a known fact. Agents quote and dispute each other. A correct report of a teammate’s error thus looked incorrect. The new plan gives each fact one of four labels:
- asserted incorrect
- attributed to a different source
- asserted correct
- missing
The numbers
On values that agents assert as their own, the broadcast effect was -0.0208 in 10 trials. This is not distinguishable from zero. For mentions, the judges found an effect of +0.1996.
What is not known
No person checked the labels. Labels from a person for 120 outputs are pending. The outputs come from one harness and 13 hand-written scenarios, and the shares do not hold for other systems.
Insight: The judge rule sets the hallucination score
Artificial Analysis counts a hallucination only when it is material, which means that it can mislead a reader. It says that more than 60% of results that are correct without this check contain one. This paper shows that a loose judge can count correct quotes as errors. A hallucination rate depends on the rule of the judge, and the party that writes the rule decides the score. Find the rule.
Sources
Posted
