![[2602.17144] When More Experts Hurt: Underfitting in Multi-Expert Learning to Defer](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.17144] When More Experts Hurt: Underfitting in Multi-Expert Learning to Defer
Summary
This article discusses the challenges of multi-expert learning in machine learning, highlighting how underfitting can occur when multiple experts are involved. It introduces a new method, PiCCE, to improve expert selection and prediction accuracy.
Why It Matters
Understanding the limitations of multi-expert systems is crucial in machine learning, especially as these systems become more prevalent in complex decision-making scenarios. The proposed method, PiCCE, offers a solution to enhance performance by effectively managing expert selection, which can lead to better outcomes in real-world applications.
Key Takeaways
- Multi-expert learning can lead to inherent underfitting, degrading performance.
- The challenge arises from identifying which expert to trust among a diverse pool.
- PiCCE method adapts expert selection to mitigate underfitting issues.
- Theoretical proofs support the consistency and effectiveness of PiCCE.
- Empirical results demonstrate improved performance in real-world scenarios.
Computer Science > Machine Learning arXiv:2602.17144 (cs) [Submitted on 19 Feb 2026] Title:When More Experts Hurt: Underfitting in Multi-Expert Learning to Defer Authors:Shuqi Liu, Yuzhou Cao, Lei Feng, Bo An, Luke Ong View a PDF of the paper titled When More Experts Hurt: Underfitting in Multi-Expert Learning to Defer, by Shuqi Liu and 4 other authors View PDF HTML (experimental) Abstract:Learning to Defer (L2D) enables a classifier to abstain from predictions and defer to an expert, and has recently been extended to multi-expert settings. In this work, we show that multi-expert L2D is fundamentally more challenging than the single-expert case. With multiple experts, the classifier's underfitting becomes inherent, which seriously degrades prediction performance, whereas in the single-expert setting it arises only under specific conditions. We theoretically reveal that this stems from an intrinsic expert identifiability issue: learning which expert to trust from a diverse pool, a problem absent in the single-expert case and renders existing underfitting remedies failed. To tackle this issue, we propose PiCCE (Pick the Confident and Correct Expert), a surrogate-based method that adaptively identifies a reliable expert based on empirical evidence. PiCCE effectively reduces multi-expert L2D to a single-expert-like learning problem, thereby resolving multi expert underfitting. We further prove its statistical consistency and ability to recover class probabilities and expert ac...
