AI agents built to argue and persuade collapse faster than expected when deployed across multiple steps — and researchers now have a clearer picture of why.
A paper published on arXiv identifies what the authors call semantic leakage in standard Retrieval-Augmented Generation, the technique that lets AI models pull in relevant context before responding. Standard RAG, they found, matches text by vocabulary overlap rather than logical structure — so an agent arguing one topic accidentally imports reasoning patterns from a loosely related one. That contamination compounds across turns, causing what the paper terms problem drift and sycophantic conformity: agents gradually abandon their position and start agreeing with whoever they are debating. The proposed fix is Taxonomic Strategy RAG, or TS-RAG, which routes retrieved strategies through a categorical classification layer to separate argumentative structure from topic-specific language. In evaluations, TS-RAG raised win rates from 70.5 to 78.5 for smaller models competing against parametrically larger ones.
The finding matters because multi-agent debate is increasingly pitched as a reliability mechanism — a way to stress-test AI outputs by having models argue against each other. If the debating agents are prone to sycophantic collapse, the oversight mechanism defeats itself. The paper also introduces turn-by-turn Debate State Representation diagnostics, which let researchers watch exactly where an agent's argument breaks down rather than only observing the final verdict.
The irony here is worth noting: the same sycophancy problem that researchers worry about in single-model assistants reappears when you pit models against each other, just with a longer delay before it surfaces.