New arXiv paper reveals why multi-agent LLM debates often fail

A new paper on arXiv dissects why multi-agent LLM pipelines—such as multi-agent debate, intrinsic self-correction, and retrieval-augmented verification—exhibit puzzling behaviors like accuracy plateaus, non-replication of gains on frontier models, and qualitative divergence across providers. The authors decompose downstream agent response into two coupled decisions: detection (whether to treat upstream content as authoritative) and conditional generation (what to produce if not). This yields four response regimes, of which 'detection-without-correction' is the load-bearing failure mode: agents correctly detect errors but then generate incorrect corrections.

Empirically, the team tested across nine configurations spanning four model families (unnamed in abstract, but frontier models), four benchmarks (GSM8K, MATH-500, GPQA-Diamond, AIME), and two methods (multi-agent debate, intrinsic self-correction). The conditional miscorrection rate consistently dominated at 53-94% across cohorts, while detection rate varied by an order of magnitude. The framework unifies four observed phenomena as signatures of a common mechanism and characterizes detection threshold as a stable model/protocol-level regularity. For practitioners building multi-agent systems, this suggests focusing on improving detection accuracy rather than just correction ability, as miscorrection is the primary bottleneck.