Research map / Label-efficient and robust learning / Test-time adaptation

Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts

NeurIPS 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Pseudo-label test-time adaptation · 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 Test-Time Correction (TTC), a training-free VLM test-time adaptation method that re-ranks top-k candidate labels by a reconstruction-based divergence shift computed against class memory subspaces.

Problem

VLM test-time adaptation mostly adjusts predictions (logits, prototypes, caches, priors), which adds computational overhead; the paper asks whether candidate verification via subspace stability can correct predictions more efficiently under distribution shift.

Approach

Treat each top-k candidate label as a hypothesis, hypothetically insert the test feature into that class's memory bank, and select the candidate with minimum aggregated divergence shift (ridge-regularized nearest-subspace reconstruction error plus pairwise shifts), without gradient updates.

Key results

• TTC improves over training-free baselines by 2.5 pp (ResNet50) and 2.4 pp (ViT-B/16) in zero-shot classification.
• TTC reaches about 2x the throughput of BCA with ViT-B/16 while using roughly 3x less CPU memory.
• In base-to-novel generalization, TTC improves the best average harmonic mean by 0.9 pp.

Surprising findings

• The L2 distance divergence metric falls below the initial CLIP prediction on average, showing that the choice of divergence matters.
• TTC overturns a correct initial prediction in only about 3.3% of cases, with failures concentrated where candidates are nearly tied.

Limitations

• Cannot recover the correct label when the true class is excluded from the original top-k candidate set.
• For unseen classes or short streams with sparse memories, TTC defaults to the initial prediction.
• Gains are modest on some settings and failures concentrate on nearly tied, semantically similar candidates.

Experimental setup

Datasets: ImageNet ImageNet-A ImageNet-R ImageNet-Sketch ImageNet-V2 Caltech101 DTD EuroSAT FGVC-Aircraft Flowers102 Food101 Oxford Pets

Benchmarks: ImageNet ImageNet-A ImageNet-R ImageNet-Sketch ImageNet-V2 Caltech101 DTD EuroSAT FGVC-Aircraft Flowers102 Food101 Oxford Pets

Metrics: top-1 accuracy harmonic mean throughput CPU memory GPU memory

Baselines: TPT DiffTPT MTA TDA OGA BCA CoOp CLIP

Builds on: CLIP TPT TDA CoOp BCA

Closest papers by meaning

• The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.94)
• Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · similarity 0.93)
• Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.92)
• Realistic Test-Time Adaptation of Vision-Language Models (CVPR 2025 · similarity 0.92)
• Noisy Test-Time Adaptation in Vision-Language Models (ICLR 2025 · similarity 0.91)
• DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.91)