Research map / Label-efficient and robust learning / Recognition and long-tail

RADAR: Relative Angular Divergence Across Representations

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RADAR, a training-free transferability metric that compares within-domain vs cross-domain layer-wise angle and relative-distance distributions (fit with GMMs, compared by symmetric KL) to rank source-domain blends for target-domain classification.

Problem

Predicting which auxiliary source domains will help or hurt (negative transfer) a target task without retraining, using only frozen foundation-model features.

Approach

Measures domain transferability through the geometry of layer-to-layer representation trajectories: angles between separation and detour vectors and normalized excess path length, aggregated over 13 layers into a 24-D descriptor, then compared between within-domain and cross-domain pairs via GMM + symmetric KL divergence.

Key results

• On DomainNet, RADAR achieves MCI of -9.71 (CLIP) and -9.87 (DINOv3), the best among evaluated metrics for both.
• On EuroEval with EmbeddingGemma, RADAR reaches MCI -30.58, the only metric with a negative MCI there, beating the next best by over 30 points.
• On PACS, RADAR underperforms (e.g., -0.27 CLIP vs -25.45 for TransRate), which the authors attribute to an intermediate domain-separation regime.

Surprising findings

• Standardizing features helps text models substantially but slightly hurts vision models.
• Optimal geometry differs by modality: text favors Euclidean space, while DINOv3 features perform best in pseudo-Cartesian space.

Limitations

• Performance degrades on PACS and ImageNet-C severity 3, where domain separation is intermediate.
• Relies on GMM density estimation, which is a computational bottleneck with no standard GPU implementation.
• Unified Euclidean formulation may not suit future hypersphere-based foundation models.

Experimental setup

Datasets: DomainNet OfficeHome PACS AmazonReviews (MTEB) EuroEval ImageNet-C

Benchmarks: DomainNet OfficeHome PACS EuroEval AmazonReviews ImageNet-C

Metrics: Spearman rank correlation Mean Correlation Improvement (MCI)

Baselines: LEEP NCE H-Score Regularized H-Score LogME TransRate S-OTDD centroid distance

Builds on: LEEP LogME TransRate s-OTDD Sinkhorn divergence

Closest papers by meaning

• Cross-Architecture Transferability is Width-Confounded:Diagnosis and Subspace Correction (NeurIPS 2026 · similarity 0.84)
• Flowing from Words to Pixels: A Noise-Free Framework for Cross-Modality Evolution (CVPR 2025 · similarity 0.83)
• Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.82)
• Towards Reliable Identification of Diffusion-based Image Manipulations (NeurIPS 2025 · similarity 0.82)
• Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance (CVPR 2024 · similarity 0.82)
• Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.82)