Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrieval

2026-06-23

AI Summary

This work addresses the challenges of error accumulation, topic drift, and blind conformity in multi-agent systems performing subjective tasks such as persuasion, which stem from semantic leakage and retrieval deficiencies in standard Retrieval-Augmented Generation (RAG). To mitigate these issues, the authors propose TS-RAG, a novel approach that introduces a classification-strategy bottleneck to decouple argumentative structure from topical content, thereby eliminating semantic leakage and enhancing long-range reasoning stability. Additionally, they design a Debate State Representation (DSR) for fine-grained diagnostic capability and incorporate a lightweight adversarial intervention mechanism that serves as a “competence bridge” for weaker agents facing stronger opponents. Experimental results demonstrate that TS-RAG improves persuasion win rates from 70.5% to 78.5%, significantly boosts abstract logical transferability, and effectively suppresses conformity behavior.

Abstract

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.

Problem

• compounding failures
• agentic persuasion
• semantic leakage
• problem drift
• sycophantic conformity

Innovation

• Taxonomic Strategy RAG
• compounding failures
• semantic leakage
• Debate State Representation
• multi-agent persuasion