CONCAT framework boosts multi-agent LLM efficiency by 2x with no training
Cut latency by 50% while outperforming trained methods—no fine-tuning needed.
A team of researchers from multiple institutions has introduced CONCAT (Consensus- and Confidence-Driven Ad Hoc Teaming), a novel framework designed to make large language model (LLM) based multi-agent systems dramatically more efficient without requiring any task-specific training. The core problem with existing multi-agent systems is the massive computational overhead from heavy communication between agents. Prior solutions involved training sparse communication graphs or fine-tuning planners, which limits generalizability and adds cost.
CONCAT solves this by first clustering agents based on their initial answers, then selecting cluster leaders according to each agent's confidence. A heuristic function grounded in Theory of Mind predicts the collaboration benefits between leader pairs using their answers and confidence scores. Finally, CONCAT prunes a percentage of inter-leader communications based on those predicted benefits, forming an ad hoc network. Across three LLMs and three benchmarks, CONCAT delivers up to 2.02x higher efficiency (accuracy per latency) versus LLM-Debate and reduces average latency by 50.1% on Qwen2.5-14B-Instruct, outperforming training-dependent methods like AgentDropout.
- Training-free: no fine-tuning or domain-specific adaptation required, preserving generalizability across tasks and models.
- Achieves up to 2.02x higher accuracy/latency ratio than LLM-Debate and reduces latency by 50.1% on Qwen2.5-14B-Instruct.
- Uses Theory of Mind-inspired heuristic to predict collaboration benefits and prune low-value agent communications.
Why It Matters
Enables faster, cheaper multi-agent LLM deployments without sacrificing accuracy or adaptability across domains.