Sarvapali D. Ramchurn

Collective Emergent Alignment: Measuring and Amplifying Interaction-Induced Reasoning Alignment in Multi-Agent LLMs

According to the abstract, Wang, Li, Ramchurn and colleagues introduce REMAP, a framework for measuring whether interaction between LLM agents improves reference-grounded reasoning beyond individual self-refinement. They report mostly positive gains on two multi-hop benchmarks and propose a compact fact-level communication scheme.