Minority Sentinel beats majority voting with 81% precision on LLM debates
When AI agents disagree, the minority opinion is right 25% of the time.
A new paper from Macquarie University identifies a critical flaw in multi-agent LLM debates: majority voting, the standard consensus mechanism, systematically suppresses correct minority opinions due to correlated errors from shared pretraining corpora. The researchers term this 'Minority Truth' and show that roughly one in four divergent cases has the minority holding the correct answer, creating a 10-percentage-point theoretical recovery margin.
To address this, the authors propose Minority Sentinel—a lightweight meta-classifier that extracts a multi-dimensional 'debate fingerprint' from conversation logs and trains a non-LLM LightGBM model to decide when to overturn the majority. Across six benchmarks and 20 random seeds, Minority Sentinel achieves a steady 81.2% flip precision with positive net gain. In contrast, an LLM-as-judge baseline yielded negative net gain despite higher recall, proving that safety—not recovery volume—determines intervention value. The work was accepted at the AgentSearch Workshop @ SIGIR 2026.
- Majority voting in multi-agent LLM debates suppresses correct minority opinions about 25% of the time due to correlated errors from shared training data
- Minority Sentinel uses LightGBM on debate fingerprints to decide when to overturn, achieving 81.2% flip precision
- LLM-as-judge baseline produced negative net gain, confirming that safe flips matter more than high recall
Why It Matters
Minority Sentinel could improve reliability of multi-agent AI systems by recovering suppressed correct answers without degrading overall accuracy.