Measuring consistency via ensemble margin and local prediction variability: Auditing decision systems in the presence of predictive multiplicity
A paper on arXiv (cs.AI) proposes a consistency criterion for auditing ensemble predictions under predictive multiplicity. The criterion combines ensemble margin with local prediction variability of constituent models. The authors show that consistency scores of finite ensembles converge to the expected model's score as ensemble size and sample count increase. The framework is evaluated on transformer models for natural language understanding and parameter-efficient fine-tuning of large language models for tabular data classification.
The paper addresses the Rashomon effect in machine learning, where equally accurate models produce different predictions. It introduces a consistency criterion for auditing incorrect ensemble predictions, combining ensemble margin and local prediction variability. Theoretical results show convergence of finite ensemble consistency scores to the expected model's score under stability and smoothness assumptions. Empirical evaluation covers transformer models for NLU and parameter-efficient fine-tuning of LLMs for tabular data.
The proposed consistency criterion integrates ensemble margin with local prediction variability, providing a principled way to flag instances for human review. Convergence guarantees suggest the method scales with ensemble size and sampling, making it applicable to large ensembles. Evaluation on transformers and LLM fine-tuning indicates relevance to modern architectures.
This research addresses a practical challenge in deploying ML systems: detecting unreliable predictions when multiple models are equally accurate. The auditing approach could improve trust and safety in high-stakes applications by identifying cases where predictions are unstable across plausible models.
The method offers a way to reduce risk from predictive multiplicity in decision systems, potentially lowering error rates in critical applications and supporting compliance with emerging AI governance requirements.
Next signals include adoption of consistency-based auditing in ML pipelines, further validation on diverse tasks, and potential integration into model monitoring tools. The convergence results may encourage larger ensembles for more reliable consistency estimates.