Tag
A preprint study uses scripted peer pressure (Asch-style) to show that LLM agents who conform in multi-agent debate still internally represent their original correct answer, revealed via Jacobian-lens analysis of residual streams — implying stated consensus can overstate true agreement.
This paper tests whether cognitive diversity (personas, temperature, model identity) drives multi-agent debate gains in small open-weight models, rejecting the hypothesis across 5,500+ budget-matched runs and arguing reported MAD improvements are essentially an ensemble-sampling effect rather than a diversity benefit.
This paper proposes TRACE, a fine-tuning framework for Retrieval-Augmented Generation (RAG) that uses multi-agent debate traces and answer completeness regularization to handle knowledge conflicts, improving robustness against misleading context and reducing incomplete answers.
This paper examines whether complex topology control in multi-agent debate for LLMs is necessary, finding that a simple random routing policy enhances cost-efficiency while maintaining competitive accuracy.
The paper proposes R2-MAD, a framework that enhances multi-agent debate in large language models by using experience memory and confidence estimation to address shared misconceptions, achieving consistent improvements over baselines.
This tweet discusses replicating a multi-agent debate system from US stocks to Chinese A-shares, with modifications for local market features like T+1 trading and price limits, emphasizing it's for research purposes only.
Debate-on-Graph (DoG) is a framework that enhances LLM reasoning by leveraging uncertain knowledge graphs (UKGs) with confidence scores, using a heuristic search and multi-agent debate mechanism to produce reliable answers. It achieves state-of-the-art performance on four QA benchmarks.
MAD-OPD utilizes a multi-teacher debate mechanism to break through the single-teacher distillation ceiling, enabling small models to surpass large teacher models in tool invocation and code generation tasks.
This paper introduces LegalHalluLens, a framework for auditing hallucinations in legal AI, providing typed hallucination profiles and a Risk Direction Index to improve trustworthy deployment.
MDForge is an LLM agent that automates the design of molecular dynamics pipelines for host-guest binding free-energy calculations, achieving human-expert competitive results and discovering a novel high-affinity binder.
This paper studies the relationship between token-level log-probability distributions, LLM-as-judge rubric scores, and final task accuracy in multi-agent debate systems. It finds a consistent four-phase confidence trajectory and role asymmetry between Constructor and Auditor agents.
RADAR introduces a role-anchored multi-agent debate framework where Politician and Scientist agents adversarially reason over evidence to detect misleading half-truths, outperforming baselines on omission-aware fact verification.