Self-Verifying and Mutually Verifying Agents reduce hallucinations and false results

The ability to create agentic workflows that can call tools and LLMs (Large Language Models) is opening up a universe of possibilities (and dangers too, but that’s for another post). Among them is verification and hallucination reduction. A couple of articles in genetics just came out to show that indeed workflows that include verification can drastically reduce incorrect statements.

One such article is “GeneAgent: self-verification language agent for gene-set analysis using domain databases” by a team led by @Zhiyong Lu at @nih. Their GeneAgent is “an LLM-based AI agent for gene-set analysis that reduces hallucinations by autonomously interacting with biological databases to verify its own output.” On existing gene sets with known functional properties, GeneAgent was significantly more accurate than GPT-4 and on novel gene sets, it was able to produce “more relevant and comprehensive functional descriptions than GPT-4,” as confirmed by experts. It would be nice to see a comparison to GPT-o3 but the approach remains relevant.

This work uses a single “agent” but others have been using multi-agent systems that call on different LLMs to evaluate and verify the outputs of one LLM.

The potential of an agentic workflow in verifying results is underestimated. Hopefully that is changing.