Researchers propose Taxonomic Strategy RAG to address compounding failures in agentic persuasion

Foundation-model agents operating in multi-step, open-ended environments frequently accumulate compounding errors, where early mistakes degrade long-horizon trajectories. In subjective tasks such as persuasion, these errors manifest as problem drift and sycophantic conformity, undermining task performance.

The authors identify semantic leakage in standard Retrieval-Augmented Generation (RAG) as a reproducible trigger for these failures. Standard RAG prioritizes vocabulary overlap over logical necessity, which can contaminate argumentative structure with topical noise.

To mitigate this, the researchers introduce Taxonomic Strategy RAG (TS-RAG), a systems intervention that routes strategies through a discrete categorical bottleneck. This design decouples argumentative structure from topical content, enabling better transfer of abstract logic across domains.

Zero-shot, cross-domain evaluations show that TS-RAG significantly improves the transfer of abstract logic where standard semantic retrieval collapses. The method also acts as a "capability bridge" in asymmetric deployments, allowing lightweight persuaders to defeat parametrically superior opponents.

In experiments, TS-RAG improved win rates from 70.5 to 78.5 and accelerated argumentative efficiency. The paper also introduces a turn-by-turn Debate State Representation (DSR) for trace-level diagnostics, highlighting the need for strict constraints to prevent evaluation collapse via default agentic sycophancy.