A Comprehensive Survey of Logit Adjustment Methods for Long-Tailed Recognition

Long-tailed data distributions, where a few head classes are well-represented and most classes are highly underrepresented, are common in real-world recognition problems. Empirical risk minimization-trained deep neural networks show a strong prediction bias towards wellrepresented classes, hindering the performance on rare classes. Recent findings show that this prediction bias mainly comes from the classifier decision-making process rather than representation learning, leading to the development of logit adjustment methods that directly modify decision scores. This survey offers a thorough and critical overview of logit adjustment methods for long-tailed learning. We formulate a general mathematical framework to represent existing logit adjustment methods as class-wise transformations zj = zj +∆j(x; S) in logit space, where the adjustment signal S forms the basis of our taxonomy: (i) prior-driven static transformations based on class frequencies πj, (ii) distribution-aware modeling leveraging feature statistics Σj, (iii) optimization-dependent adjustment tracking gradient evolution ∇ θ , (iv) instance-conditioned refinement adapting to sample difficulty, and (v) noise-resilient correction in the presence of unreliable supervision. For each category, we analyze theoretical assumptions, computational tradeoffs, decision boundary effects, and failure modes, while clarifying connections to margin-based losses and classical reweighting.