DARL: Mitigating Gradient Conflicts in Long-Tailed Out-of-Distribution Learning

Long-tailed out-of-distribution learning aims to reduce performance bias in long-tailed in-distribution (ID) data while rejecting out-of-distribution (OOD) samples, which are often mistaken for under-represented tail classes. To achieve OOD detection, existing methods incorporate an outlier exposure (OE) term into the long-tailed recognition (LTR) loss. However, as we prove in this paper, the OE term induces a gradient conflict with the ID objectives, especially for tail classes, thereby contradicting the core motivation of LTR. To avoid the ID-OOD dilemma, we propose Dynamic Ambiguity-aware Recalibration for Logits (DARL), an ambiguity-guided long-tailed OOD learning approach, grounded on two theoretical insights. First, we show that the mixed ID data can mitigate the conflict in OE training and exhibits higher intrinsic ambiguity than the original ID data, thus able to serve as a surrogate for real OOD data. Second, we introduce an ambiguity-aware logit adjustment that can dynamically calibrate the class margins using energy-based ambiguity metrics, effectively reducing early-stage bias while avoiding late-stage overfitting. Extensive experiments show that DARL achieves the overall state-of-the-art performance of long-tailed OOD learning. Moreover, compared with the OE methods, DARL trains solely on the ID data, which can reduce the data requirements by 80%. The code is available in https://github.com/XuanZhang-A/DARL.