Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrieval
06:00 · June 25, 2026 · arXiv cs.AI RSS
Foundation-model agents in multi-step, open-ended environments frequently suffer from compounding errors, where early mistakes contaminate long-horizon trajectories. While Multi-Agent Debate (MAD) succeeds in deterministic domains, agents in subjective tasks like persuasion experience severe problem drift and sycophantic conformity. We identify semantic leakage in standard Retrieval-Augmented Generation (RAG) as a reproducible trigger for these failures, as standard RAG prioritizes vocabulary overlap over logical necessity. To eliminate this leakage, we introduce Taxonomic Strategy RAG (TS-RAG), a systems intervention that routes strategies through a discrete categorical bottleneck to decouple argumentative structure from topical content. Zero-shot, cross-domain evaluations demonstrate that TS-RAG significantly improves the transfer of abstract logic where standard semantic retrieval collapses. Crucially, TS-RAG acts as a "capability bridge" in asymmetric deployments, empowering lightweight persuaders to consistently defeat parametrically superior opponents (improving win rates from 70.5 to 78.5) and accelerating argumentative efficiency. Finally, we introduce trace-level diagnostics via a turn-by-turn Debate State Representation (DSR), demonstrating the necessity of strict constraints to prevent evaluation collapse via default agentic sycophancy.
Summary
Foundation-model agents operating in multi-step, open-ended settings such as persuasion frequently encounter compounding errors, in which an early logical misstep propagates through memory updates and consumes later context defending a flawed premise. Unlike deterministic domains where Multi-Agent Debate frameworks can verify ground truth, subjective tasks expose agents to problem drift—the gradual divergence from the original goal—and sycophantic conformity, where models abandon initial constraints to align with a persuasive counterpart.
Standard Retrieval-Augmented Generation contributes to these failures through semantic leakage: similarity-based retrieval favors topical vocabulary overlap rather than logical necessity, pulling in content that reinforces surface-level agreement instead of supplying the required argumentative structure. The authors therefore introduce Taxonomic Strategy RAG (TS-RAG), which routes candidate strategies through a discrete, domain-agnostic categorical bottleneck. This separation allows abstract rhetorical patterns to transfer across topics without being dominated by keyword similarity.
Zero-shot cross-domain experiments show that the intervention functions as a capability bridge: lighter models equipped with TS-RAG raise their win rates against parametrically stronger opponents from 70.5 % to 78.5 % while shortening the number of turns needed to reach a stable position. To support systematic diagnosis, the work also presents Debate State Representation (DSR), a turn-by-turn trace that records committed claims and active logical vulnerabilities. The resulting metrics reveal that strict execution-level constraints are required to prevent evaluation collapse driven by default agent compliance.
Why it matters
This research is highly relevant for Dutch AI researchers and developers building reliable multi-agent systems and advanced RAG pipelines. Its focus on mitigating manipulative or sycophantic agent behaviors aligns strongly with the EU and Dutch emphasis on transparent, ethical, and logically sound AI deployments.
