COMED
A multi-model reasoning framework that enables selective cross-model collaboration via a post-anchor controller, positioned between single routing and dense collaboration. It uses three signals—self-consistency, router margin, and lightweight peer probes—to decide whether to escalate queries to additional models.
Timeline (last 90 days)
COMED (Controlled Model Escalation for Multi-LLM Deliberation) is a multi-model reasoning approach proposed in an arXiv paper, aimed at filling the middle ground between routing and dense collaboration.
COMED is designed as a post-anchor controller that enables selective cross-model collaboration.
COMED determines its processing strategy through three signals: anchor self-consistency, router margin, and lightweight peer probes.
COMED proposes a rescue-harm decomposition, where selective collaboration yields a net benefit when the number of rescued errors exceeds the harm introduced by collaboration.
Across 16 open-weight configurations, COMED improved upon both fixed-anchor and routed-anchor performance in all cases.
COMED achieves an improvement of up to +10.7 percentage points on the MedQA medical question-answering benchmark.
Compared to dense collaboration, COMED invokes fewer models and decodes fewer tokens, while simultaneously improving accuracy and reducing computational cost.
On the HLE benchmark, COMED improved GPT-5.5's performance from 23.1% to 28.1%.
COMED was validated across multiple benchmarks including medical, scientific, and general reasoning.
All Facts (9)
COMED (Controlled Model Escalation for Multi-LLM Deliberation) is a multi-model reasoning approach proposed in an arXiv paper, aimed at filling the middle ground between routing and dense collaboration.
50%UnverifiedCOMED is designed as a post-anchor controller that enables selective cross-model collaboration.
50%UnverifiedCOMED determines its processing strategy through three signals: anchor self-consistency, router margin, and lightweight peer probes.
50%UnverifiedCOMED proposes a rescue-harm decomposition, where selective collaboration yields a net benefit when the number of rescued errors exceeds the harm introduced by collaboration.
50%UnverifiedAcross 16 open-weight configurations, COMED improved upon both fixed-anchor and routed-anchor performance in all cases.
50%UnverifiedCOMED achieves an improvement of up to +10.7 percentage points on the MedQA medical question-answering benchmark.
50%UnverifiedCompared to dense collaboration, COMED invokes fewer models and decodes fewer tokens, while simultaneously improving accuracy and reducing computational cost.
50%UnverifiedOn the HLE benchmark, COMED improved GPT-5.5's performance from 23.1% to 28.1%.
50%UnverifiedCOMED was validated across multiple benchmarks including medical, scientific, and general reasoning.
50%