Learning to Trust the Crowd: A Multi-Model Consensus Reasoning Engine for Large Language Models

TL;DR

This work tackles the unreliability of large language models by leveraging a supervised meta-learning framework that ingests outputs from multiple heterogeneous LLMs. It builds a rich feature suite—spanning semantic embeddings, similarity clustering, lexical cues, reasoning quality, and model priors—and applies gradient-boosted trees, listwise ranking, and graph neural networks over a similarity graph of candidate answers. Across four compact benchmarks (GSM8K, ARC-Challenge, HellaSwag, TruthfulQA) using three open-weight models on a single GPU, the graph-attention consensus model yields substantial gains over the best single model and majority voting, while improving calibration and reducing hallucinations. The results demonstrate that supervised multi-model consensus is a practical route to more reliable LLM behavior, especially in multi-step reasoning and truthfulness domains, with explicit avenues for future cost-aware and domain-adaptive extensions.

Abstract

Large language models (LLMs) achieve strong aver- age performance yet remain unreliable at the instance level, with frequent hallucinations, brittle failures, and poorly calibrated confidence. We study reliability through the lens of multi-model consensus: given responses from several heterogeneous LLMs, can we learn which answer is most likely correct for a given query? We introduce a Multi-Model Consensus Reasoning Engine that treats the set of LLM outputs as input to a supervised meta-learner. The system maps natural language responses into structured features using semantic embeddings, pairwise similarity and clustering statistics, lexical and structural cues, reasoning-quality scores, confidence estimates, and model-specific priors, and then applies gradient-boosted trees, listwise ranking, and graph neural networks over similarity graphs of answers. Using three open-weight LLMs evaluated on compact, resource- constrained subsets of GSM8K, ARC-Challenge, HellaSwag, and TruthfulQA, our best graph-attention-based consensus model improves macro-average accuracy by 4.6 percentage points over the strongest single LLM and by 8.1 points over majority vote, while also yielding lower Brier scores and fewer TruthfulQA hal- lucinations. Ablation and feature-importance analyses show that semantic agreement and clustering features are most influential, with reasoning-quality and model-prior features providing com- plementary gains, suggesting supervised multi-model consensus is a practical route toward more reliable LLM behavior, even in a modest single-machine setup.

Figures (3)

• Figure 1: Overview of the Multi-Model Consensus Reasoning Engine. A query is sent to multiple base LLMs; their responses are mapped to feature vectors (embeddings, similarity, clustering, priors), and a consensus meta-model selects the final answer $\hat{a}$.
• Figure 2: Accuracy of baselines and the GAT-based consensus model across mini-benchmarks. Error bars indicate one standard deviation over three runs of the meta-model (for baselines, they reflect small variability from resampling). The consensus model consistently outperforms the best single LLM and majority vote on all datasets.
• Figure 3: Ablation study for the GBDT consensus model on the mini-benchmarks. Removing semantic similarity and clustering features leads to the largest drop in macro-averaged accuracy, followed by removal of confidence/model priors and reasoning-quality scores.