TS-RAG boosts AI persuader win rates by 8% against stronger models

A new paper from researchers Pradyumna Narayana, Sana Ayromlou, and Purvi Sehgal tackles the problem of compounding failures in AI agents designed for multi-step persuasion tasks. They identify a key flaw in standard Retrieval-Augmented Generation (RAG): semantic leakage. Because standard RAG prioritizes vocabulary overlap over logical necessity, agents in open-ended debates suffer from problem drift and sycophantic conformity—essentially agreeing with the opponent rather than sticking to a coherent argument.

To fix this, the team introduces Taxonomic Strategy RAG (TS-RAG), a systems-level intervention that routes retrieved strategies through a discrete categorical bottleneck. This decouples argumentative structure from topical content, allowing agents to transfer abstract logic across domains without semantic contamination. In zero-shot cross-domain evaluations, TS-RAG acts as a "capability bridge," enabling lightweight persuaders to consistently defeat parametrically superior opponents. Win rates jumped from 70.5% to 78.5%, and argumentative efficiency improved. The paper also introduces trace-level diagnostics via a turn-by-turn Debate State Representation (DSR) to prevent evaluation collapse from default sycophancy.