CAMO: A Class-Aware Minority-Optimized Ensemble for Robust Language Model Evaluation on Imbalanced Data

Class-imbalance methods implicitly assume that the minority class is uniformly undersampled relative to the majority class. However, in many real-world settings, minority instances may be disproportionately under-observed in certain regions of the feature space. For example, small businesses that go bankrupt may disappear from records, while those that survive remain visible, making bankruptcy appear less common among small firms than it actually is. This gives rise to localized minority imbalance (LMI), a challenge that is often overlooked and extends beyond general class-count imbalance. We show that under LMI, existing imbalance mitigation techniques can fit observation-induced biases in the training data and generalize poorly to under-observed regions of the true minority distribution. To address this, we propose a tree-based stratified approach that recursively partitions the feature space with the goal of reducing within-stratum LMI distortion. For each resulting stratum, we pair its majority instances with the full observed minority set and train a base classifier to create an ensemble. Extensive experiments over benchmark tabular datasets simulated with LMI show that our stratified ensembling approach outperforms popular and state-of-the-art imbalance mitigation techniques. We also introduce a gold-standard evaluation protocol that uses unbiased test sets, and demonstrate that conventional hold-out evaluation from the same LMI-biased data can substantially mislead performance. Overall, our results highlight that the cause of imbalance is as important as the correction method.

Mixture-of-Experts (MoE) models have become the standard architecture for frontier LLMs. The multiple parallel experts in each transformer block theoretically enable expert specialization according to categories such as domains, languages, or modalities during pretraining. However, when data is imbalanced across categories, minority categories can become concentrated in a small fraction of the expert pool or spread inefficiently across them, effectively limiting their available capacity. In this work, we demonstrate that category-conditioned expert-selection biases can preferentially route tokens from different categories toward specialized expert pools. By applying biases based on input category, we provide a simple way to induce soft category-aligned specialization. This intervention encourages categories toward distinct expert subsets while retaining cross-category sharing when useful. More importantly, it enables the over-allocation of expert capacity to categories that would otherwise be disadvantaged by data scarcity. We observe performance improvements across modalities, low-resource languages, and code in models up to 10.6B parameters. At our smallest and largest model scales, these gains are equivalent to allocating 18.5% and 2.7% additional baseline training tokens, respectively. Overall, our results show that expert biases can serve as a lightweight control for category-level expert allocation, making room for low-resource categories while retaining standard MoE inference.

Class-level evaluation can conceal substantial performance disparities across subconcepts within the same class, causing models that perform well on average to fail on specific subpopulations. Prior work has shown that common evaluation measures for imbalanced classification are biased toward larger minority subconcepts and that utility-based reweighting using true subconcept labels can mitigate this bias; however, such labels are rarely available at test time. We introduce a practical utility-weighted evaluation that replaces unavailable subconcept labels with predicted posterior probabilities from a multiclass subconcept model. Evaluation weights are defined as the expected utility under this posterior, yielding a soft, uncertainty-aware metric we call predicted-weighted balanced accuracy (pBA). Experiments on tabular benchmarks as well as medical-imaging and text datasets show that unweighted scores can be misleading under within-class heterogeneity, while pBA provides more stable and interpretable assessments when subconcept distributions are uneven but not pathological. Our code is available at: https://anonymous.4open.science/r/correcting-bias-imbalance-9C6C/.