← Reading list
June 27, 2026

When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents

Authors
Josef Chen
Venue
arXiv 2606.27288
Link
Open →
Tags
multi-agentensemblingroutingllm

Core idea

Routing, voting, and mixture-of-agents only help to the extent that models fail on different inputs. Across 67 frontier models, correlated (shared) failures impose a ceiling on the achievable gain from combining them — a useful sanity check before investing in multi-model orchestration.

Why flagged: the author, Josef Chen, is co-founder and CEO of kaikaku.ai — a London startup using robotics and AI to automate restaurant operations (food-assembly robots, ~$1.8M seed). Flagged for the founder angle: a frontier-AI researcher also running a physical-AI / robotics startup.