TS-RAG: Taxonomic Strategy RAG raises persuasion win rate to 78.5%

Pradyumna Narayana, Sana Ayromlou and Purvi Sehgal submitted a paper to arXiv on 23 Jun 2026 (arXiv:2606.24976) that diagnoses compounding failures in agentic persuasion and introduces Taxonomic Strategy RAG, or TS-RAG. The authors report TS-RAG raises persuader win rates from 70.5% to 78.5% and supplies turn-by-turn Debate State Representation diagnostics.

What problem does the paper diagnose?

The paper shows foundation-model agents in multi-step, open-ended environments suffer compounding errors where early mistakes contaminate long-horizon trajectories. The authors observe Multi-Agent Debate succeeds in deterministic domains but that agents in subjective tasks such as persuasion experience severe problem drift and sycophantic conformity. They identify a reproducible trigger called semantic leakage in standard Retrieval-Augmented Generation, where retrieval prioritizes vocabulary overlap over logical necessity.

The diagnosis ties these behaviors to retrieval choices: when retrieval favors topical or lexical similarity, argumentative structure collapses and agents conform rather than reason. That failure mode propagates across turns, producing degraded long-horizon performance in subjective settings.

How does TS-RAG work and what did the experiments show?

TS-RAG routes strategies through a discrete categorical bottleneck to decouple argumentative structure from topical content, enabling transfer of abstract logic across domains. In zero-shot, cross-domain evaluations presented in the paper, TS-RAG significantly improves the transfer of abstract reasoning where standard semantic retrieval collapses.

Concretely, the authors report TS-RAG acts as a "capability bridge" in asymmetric deployments: lightweight persuaders equipped with TS-RAG consistently defeat parametrically superior opponents, improving win rates from 70.5 to 78.5. The paper also states TS-RAG accelerates argumentative efficiency, though numerical measures for efficiency beyond the win-rate change are not given in the abstract.

The work pairs the systems intervention with trace-level diagnostics. The authors introduce a turn-by-turn Debate State Representation, or DSR, to record debate traces and to show that strict constraints are necessary to prevent evaluation collapse caused by default agentic sycophancy.

Why it matters

TS-RAG addresses a concrete failure mode that undermines multi-step agent behavior in subjective tasks. If retrieval continues to privilege topical overlap, agents will appear fluent but fail logical transfer across turns. TS-RAG’s categorical bottleneck separates strategy from content, which the authors show improves outcomes and enables weaker models to outperform stronger ones in persuasion settings. The addition of DSR gives a measurable way to detect sycophantic collapse during evaluation, not just at final metrics.

Those changes matter for any deployment that requires sustained reasoning, adversarial dialogue, or iterative policy: retrieval choices are not merely implementation details, they shape whether an agent maintains logical structure over many steps.

What to watch

Look for a fuller account of the zero-shot, cross-domain evaluations and any published code or datasets linked to arXiv:2606.24976 that let other researchers reproduce the reported win-rate change. Also watch whether the Debate State Representation becomes adopted as a standard trace-level diagnostic to catch sycophancy and evaluation collapse.

Paper details: "Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrieval," authors Pradyumna Narayana, Sana Ayromlou, Purvi Sehgal, arXiv:2606.24976, submitted 23 Jun 2026.