ACAR: Adaptive Complexity Routing for Multi-Model Ensembles with Auditable Decision Traces

2026-02-26 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Advanced, long

Summary

ACAR (Adaptive Complexity & Attribution Routing) is a measurement framework for multi-model orchestration that uses self-consistency variance ($\\sigma$) to route tasks across single-model, two-model, and three-model execution modes. Evaluated across 1,510 tasks from four benchmarks (MathArena, Reasoning Gym, LiveCodeBench, SuperGPQA) using Claude Sonnet 4, GPT-4o, and Gemini 2.0 Flash, ACAR-U (without retrieval) achieved 55.6% accuracy, surpassing the two-model baseline (54.4%) while avoiding full ensembling on 54.2% of tasks. The mechanism is model-agnostic and requires no learned components. Notably, retrieval augmentation *decreased* accuracy by 3.4 percentage points due to low semantic alignment (median similarity 0.167), and "agreement-but-wrong" scenarios (where models agree on incorrect answers) bounded achievable accuracy 8 percentage points below full ensembling. The framework also found that attribution estimates based on proxy signals showed weak correlation with ground-truth values.

Key takeaway

For AI Architects designing multi-model LLM systems, ACAR's findings suggest prioritizing heuristic-based routing over learned classifiers for auditability and stability. You should implement self-consistency variance for adaptive compute allocation, as it improves accuracy over fixed two-model ensembles while reducing full ensemble usage. Critically, avoid naive retrieval augmentation without high semantic similarity thresholds, and recognize that "agreement-but-wrong" scenarios will inherently limit maximum achievable accuracy, necessitating strategies beyond simple ensembling for those cases.

Key insights

Self-consistency variance can effectively route multi-model ensembles, but retrieval and proxy attribution often fail.

Principles

• Self-consistency routing has fundamental limits.
• Retrieval requires semantic alignment.
• Attribution requires counterfactuals.

Method

ACAR routes tasks based on self-consistency variance ($\\sigma$) derived from N=3 probe samples, mapping variance to single-agent, arena-lite, or full-arena execution modes. This heuristic avoids learned components for auditability.

In practice

• Use similarity thresholds >0.7 for retrieval augmentation.
• Expect an accuracy ceiling from "agreement-but-wrong" failures.
• Implement explicit counterfactuals for practical attribution.

Topics

• Adaptive Complexity Routing
• Multi-Model Ensembles
• Self-Consistency Variance
• Auditable AI Systems
• Retrieval Augmentation

Code references

• mechramc/ACAR-TeamLLM

Best for: AI Scientist, Research Scientist, AI Architect, AI Researcher, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Rethinking the Necessity of Adaptive Retrieval-Augmented Generation through the Lens of Adaptive Listwise Ranking — Adaptive retrieval's role shifts from noise filtering for weaker LLMs to efficiency optimization for stronger ones.
• Value-and-Structure Alignment for Routing-Consistent Quantization of Mixture-of-Experts Models — MoE quantization requires preserving both router output values and structural relationships to maintain expert selection consistency and mitigate performance degradation.
• Adversarial Consensus Protocol (ACP): Validation by Contradiction in Multi-LLM Systems — Real architectural divergence between LLMs, not staged conflict, is key to robust validation.
• Self-Consistency from Only Two Samples: CoT-PoT Ensembling for Efficient LLM Reasoning — CoT-PoT ensembling improves LLM reasoning accuracy and efficiency by leveraging diverse, complementary reasoning modalities.
• A Single Rewrite Suffices: Empirical Lessons from Production Skill Description Optimization — A single LLM rewrite of skill descriptions significantly improves routing accuracy and engineering efficiency for AI agents.
• A Heterogeneous Temporal Memory Governance Framework for Long-Term LLM Persona Consistency — External temporal memory governance improves LLM long-term persona consistency and reduces drift.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.