ACAR: A New Approach to Multi-Model Routing

A recent study published on arXiv introduces ACAR (Adaptive Complexity and Attribution Routing), a framework designed to analyze the orchestration of multiple models under auditable conditions. ACAR uses self-consistency variance (sigma), computed from three probe samples, to route tasks through execution modes involving a single model, two models, or three models.

The system is implemented on TEAMLLM, a deterministic execution substrate with immutable artifacts and complete decision traces. The evaluation of ACAR was conducted on 1,510 tasks, covering four benchmarks: MathArena, Reasoning Gym, LiveCodeBench, and SuperGPQA, using Claude Sonnet 4, GPT-4o, and Gemini 2.0 Flash, generating over 7,550 auditable runs.

Results and Limitations

The results show that sigma-based routing achieves 55.6% accuracy, exceeding the two-model baseline of 54.4% and avoiding full ensembling on 54.2% of tasks. The routing mechanism is model-agnostic and requires no learned components. However, the study also documents negative results. Retrieval augmentation reduced accuracy by 3.4%, due to low semantic similarity. Furthermore, when models agree on incorrect answers (sigma equals zero), no downstream ensemble can recover, limiting the maximum achievable accuracy. Finally, attribution estimates based on proxy signals show a weak correlation with ground-truth values, suggesting that practical attribution requires explicit counterfactual computation.

This work identifies the assumptions that fail in practice and provides falsifiable baselines for future research on routing, retrieval, and multi-model attribution.