CILA: Confusion-based Instance-wise Logit Adjustment for Long-Tailed Recognition

Jiyoung Lim; Jee-Hyong Lee

cs.CVOct 20, 2026

CILA: Confusion-based Instance-wise Logit Adjustment for Long-Tailed Recognition

Authors: Jiyoung Lim, Jee-Hyong Lee

Organizations: Sungkyunkwan University

Abstract

Real-world datasets are inherently class-imbalanced, leading models to favor majority classes and struggle with minority ones. Existing logit adjustment methods address this by modifying logits based on class frequencies, but their class-wise adjustments ignore instance-wise confusion and often suppress informative signals from confusing non-target classes. This results in insufficient learning of such classes, increased prediction uncertainty, and limited generalization. To address this, we propose two complementary methods. First, Confusion-based Instance-wise Logit Adjustment (CILA) adaptively adjusts logits based on instance-wise confusion between the target and non-target classes, restoring informative gradients that class-wise adjustment suppresses. Second, Batch-adaptation Noise-Injection (BNI) perturbs the target logit during training, extending the confusion-aware update to correctly classified but near-boundary samples that the penalty alone cannot reach. CILA is a loss-agnostic plug-in module that can be seamlessly integrated into existing long-tailed learning methods, and consistently improves performance across multiple benchmarks. Our code will be available at https://github.com/maya0395/CILA.

Figures & tables

Figure 1: Visualization of the Non-Target Logit Over-Suppression Problem on CIFAR100-IF100. Logit adjustment increases head class target probabilities while reducing those of tail classes. However, it also excessively suppresses non-target probabilities of middle and tail classes, weakening informative gradients from confusing classes and hindering discrimination between confusing classes.

Figure 2: Illustration of class-wise logit adjustment in non-confusing and confusing cases. (a) shows a non-confusing case that is easily distinguished. However, (b) represents a confusing case that is easily confused with another class ( Dog ). Class-wise logit adjustment increases a positive sample loss for the correct class while simultaneously reducing negative sample losses for confusing classes.

		IF 10				IF 50				IF 100
Method	Loss	Head	Med	Tail	All	Head	Med	Tail	All	Head	Med	Tail	All
Loss	CE	67.4 $\pm$ 0.5	45.7 $\pm$ 0.5	-	60.7 $\pm$ 0.5	69.7 $\pm$ 0.4	40.2 $\pm$ 0.6	18.0 $\pm$ 0.7	48.3 $\pm$ 0.3	72.0 $\pm$ 0.3	43.5 $\pm$ 0.6	10.0 $\pm$ 0.4	43.4 $\pm$ 0.3
	LDAM	67.1 $\pm$ 0.2	46.5 $\pm$ 0.8	-	60.8 $\pm$ 0.4	69.8 $\pm$ 0.3	39.3 $\pm$ 0.5	16.3 $\pm$ 0.5	47.7 $\pm$ 0.0	71.4 $\pm$ 0.4	42.4 $\pm$ 1.1	10.5 $\pm$ 0.5	43.0 $\pm$ 0.4
	DBM	67.7 $\pm$ 0.5	46.8 $\pm$ 0.7	-	61.2 $\pm$ 0.4	69.8 $\pm$ 0.2	40.0 $\pm$ 0.5	17.7 $\pm$ 0.5	48.2 $\pm$ 0.2	71.7 $\pm$ 0.8	42.3 $\pm$ 0.5	10.2 $\pm$ 0.4	43.0 $\pm$ 0.2
	CILA	66.8 $\pm$ 0.1	54.1 $\pm$ 0.3	-	62.7 $\pm$ 0.3	68.0 $\pm$ 0.2	44.6 $\pm$ 0.5	29.0 $\pm$ 0.4	51.2 $\pm$ 0.3	70.3 $\pm$ 0.5	46.1 $\pm$ 0.2	19.2 $\pm$ 0.8	46.1 $\pm$ 0.1
	CE-DRW	65.1 $\pm$ 0.7	54.2 $\pm$ 1.6	-	61.7 $\pm$ 0.2	65.9 $\pm$ 0.7	44.9 $\pm$ 1.0	28.4 $\pm$ 0.5	50.5 $\pm$ 0.1	67.6 $\pm$ 0.9	47.3 $\pm$ 0.8	19.0 $\pm$ 0.3	45.9 $\pm$ 0.2

Table 1: Performance comparison with loss function-based on the CIFAR-100-LT, re-implemented from official code (CUDA/DODA/SHIKE: respective repositories; others: DBM [ 29 ] )

Figure 3: Overview of the proposed method.

	ImageNet-LT				iNaturalist 2018
Method	Head	Med	Tail	All	Head	Med	Tail	All
CE †	64.0	33.8	5.8	41.6	73.9	63.5	55.5	61.0
cRT †	58.8	44.0	26.1	47.3	69.0	66.0	63.2	65.2
LWS †	57.1	45.2	29.3	47.7	65.0	66.3	65.5	65.9
CE-DRW †	61.7	47.3	28.8	50.1	68.2	67.3	66.4	67.0
LDAM-DRW †	60.4	46.9	30.7	49.8	-	-	-	66.1

Table 2: Accuracy (%) on ImageNet-LT and iNaturalist 2018. † and ‡ denote results from DODA [ 33 ] and DBM [ 29 ] , respectively.

Figure 4: Confusion penalty and accuracy across training epochs under three noise settings: no noise (zero), fixed-scale noise (fix), and batch-adaptive noise (mu) (CIFAR-100-LT, IF=100).

Method	Inter-class Separability $\uparrow$				Intra-class Compactness $\downarrow$
Method	Head	Med	Tail	All	Head	Med	Tail	All
LA [ 23 ]	6.0	6.0	5.7	5.9	28.6	28.8	29.6	29.0
DBM-LA [ 29 ]	7.0	6.3	6.0	6.2	31.0	31.9	33.1	32.0
CILA-LA (Ours)	6.8	6.9	6.1	6.6	28.1	28.3	29.3	28.6

Table 3: Analysis of inter-class separability and intra-class compactness on CIFAR100 (IF 100).

Figure 5: Mean prediction entropy on test data.

Figure 6: Difference in confusion matrices (CILA $-$ LA) for semantically similar class pairs on CIFAR-100-LT (IF=100). Positive values (green) indicate improved correct predictions by CILA, confirming reduced misclassification between fine-grained classes.

Components			IF 50				IF 100
CILA	BNI	$D_{y}$	Head	Med	Tail	All	Head	Med	Tail	All
			62.7	47.5	36.8	51.8	62.3	47.1	25.3	45.9
✓			64.6	46.8	36.1	52.1	64.9	47.0	25.8	46.9
✓		✓	64.3	47.0	37.3	52.3	64.6	47.6	28.0	47.6
	✓		69.8	42.8	22.1	50.1	70.3	45.4	15.2	45.1
	✓	✓	63.3	47.9	37.7	52.4	62.2	47.1	29.1	47.0

Table 4: Ablation study on the three components under different imbalance factors (IF = 50, 100). ✓indicates the corresponding component is enabled.

Upper Bound	IF 50				IF 100
Upper Bound	Head	Med	Tail	All	Head	Med	Tail	All
0.5	62.3	48.0	37.2	51.9	62.7	47.4	29.3	47.4
1.0	63.2	47.0	39.1	52.2	64.3	48.4	28.1	47.9
1.5	64.3	47.4	37.3	52.5	64.8	48.2	28.8	48.2
2.0	61.3	49.3	42.2	53.0	64.7	47.6	30.1	48.3

Table 5: Performance comparison across penalty upper bound.

Noise Type	IF 50				IF 100
Noise Type	Head	Med	Tail	All	Head	Med	Tail	All
Chi	63.3	48.4	35.5	52.2	63.4	48.5	28.5	47.7
Gaussian	65.0	47.0	39.0	52.9	64.3	48.7	28.9	48.2
Rayleigh	61.3	49.3	42.2	53.0	64.7	47.6	30.1	48.3

Table 6: Performance comparison across noise types.

$\rho$ in Eq. 10	IF 50				IF 100
$\rho$ in Eq. 10	Head	Med	Tail	All	Head	Med	Tail	All
0.1	64.0	48.3	38.0	52.9	64.7	48.0	27.9	47.8
0.2	61.3	49.3	42.2	53.0	64.7	47.6	30.1	48.3
0.3	63.2	47.9	37.7	52.4	62.5	50.0	27.5	47.6

Table 7: Effect of noise ratio

\rho

on accuracy.

Class-Imbalanced Learning

Class-imbalanced learning develops classifiers that recognize rare classes without sacrificing performance on common ones.

See all 257 papers

CUE: Concept-Aware Multi-Label Expansion to Mitigate Concept Confusion in Long-Tailed Learning
May 2, 2026Ruichi Zhang, Chikai Shang, Jiacheng Yang +4Long-Tailed Classification Logit Adjustment
Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China · College of Computer Science and Software Engineering, Shenzhen University, China · Institute of High Performance Computing, A*STAR, Singapore
CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels
May 22, 2026Mengke Li, Haiquan Ling, Lihao Chen +3Class-Imbalanced Learning Learning with Noisy Labels
College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China
Beyond Static Costs: Learning-Dynamics Aware Loss Functions for Long-Tailed Classification
Jul 28, 2026Varad Shinde, Nikhil Kumar Shrey, Magesh Rajasekaran +5Class-Imbalanced Learning Long-Tailed Classification

May 2, 2026J/K move · Enter open · S save

CUE: Concept-Aware Multi-Label Expansion to Mitigate Concept Confusion in Long-Tailed Learning

Ruichi Zhang, Chikai Shang, Jiacheng Yang, Mengke Li+3

Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China · College of Computer Science and Software Engineering, Shenzhen University, China · Institute of High Performance Computing, A*STAR, Singapore

Long-Tailed Classification Logit Adjustment Vision Foundation Model Adaptation

Open paper AskPDF arXiv

Long-tailed distributions are common in real-world recognition tasks, where a few head classes have many samples while most tail classes have very few. Recently, fine-tuning foundation models for long-tailed learning has gained attention due to their excellent performance. However, most existing methods focus solely on mitigating long-tailed distribution bias while overlooking concept confusion caused by the long-tailed distribution. In this paper, we study this problem and attribute it to the mutual exclusivity of single-label supervision under long-tailed distributions, which suppresses feature sharing among related classes and amplifies the dominance of head classes, leading to disrupted inter-class discriminability. To address this, we propose CUE, Concept-aware mUlti-label Expansion, which introduces multi-label concept signals to preserve disrupted inter-class relationships. Specifically, CUE constructs concept sets by (i) extracting instance-level visual cues from zero-shot CLIP and (ii) generating class-level semantic cues with LLM; the two cues are incorporated via separately weighted Binary Logit-Adjustment (BLA) auxiliary losses and jointly optimized with the baseline Logit-Adjustment (LA) loss. Experiments on several long-tailed benchmarks, CUE achieves balanced and strong performance, surpassing recent state-of-the-art methods. Code is available at: https://github.com/zhangruichi/CUE.

Abstract

Class-Imbalanced Learning

Explore similar work

CUE: Concept-Aware Multi-Label Expansion to Mitigate Concept Confusion in Long-Tailed Learning

CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels

Beyond Static Costs: Learning-Dynamics Aware Loss Functions for Long-Tailed Classification