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

EVA-0: Test-Time Model Evolution with Only Two Forward Passes per Sample

NeurIPS 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Realistic and stable 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 EVA-0, a backprop-free test-time adaptation method that uses sample-wise symmetric two-sided zeroth-order updates, shortcut-resistant scale-invariant entropy, and anchor-guided perturbation/relaxation, so inference and adaptation both fit in two forward passes per sample.

Problem

Test-time adaptation normally relies on backpropagation, which is memory-heavy and infeasible for quantized, edge, accelerator or black-box models; existing forward-only methods need many forward passes (FOA: 28) or give noisy updates (one-sided SPSA in ZOA), and degrade over continual streams.

Approach

Use the two symmetric perturbed forward passes both as a two-sided SPSA gradient estimate and, averaged, as the inference output; make the entropy objective scale-invariant to block logit-norm shortcuts; draw per-sample independent perturbations; and pull weights toward a fixed pretrained anchor to prevent drift in continual adaptation.

Key results

• On ViT-Base ImageNet-C (severity 5, single-domain), EVA-0 reaches 66.7% average accuracy with 2 forward passes, versus 65.7% for BP-based DeYO and 66.3% for 28-forward FOA.
• On ResNet50-GN, EVA-0 gets 45.4% average accuracy, outperforming FOA (35.3%) by about 10 points.
• Under continual adaptation on ViT-Base, FOA-dagger falls from 55.9% to 22.3% while EVA-0 keeps 67.0%; on 8-bit ViT EVA-0 reaches 65.6% vs FOA 63.5%.

Surprising findings

• Smaller test batches gave better zeroth-order adaptation than larger batches, contrary to prior offline ZO findings.
• Naive ZO updates cause weight norm growth and drift in continual streams, which BP-based updates did not show under the same objective.

Limitations

• Evaluated only on ImageNet-C with ViT-Base and ResNet50-GN; no language or other modalities.
• Relies on a pretrained anchor and hyperparameters (mu, gamma, lambda) tuned per architecture.
• Code release is promised ('Code will be released') rather than provided.

Experimental setup

Datasets: ImageNet-C ImageNet-1K

Benchmarks: ImageNet-C

Metrics: accuracy wall-clock time peak memory forward passes per sample

Baselines: TENT CoTTA SAR DeYO ZOA FOZO FOA

Builds on: FOA ZOA FOZO SPSA TENT MeZO

Closest papers by meaning

• Test-Time Model Adaptation with Only Forward Passes (ICML 2024 · similarity 0.91)
• A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
• Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.88)
• ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
• Backpropagation-free Network for 3D Test-time Adaptation (CVPR 2024 · similarity 0.87)
• Unraveling Batch Normalization for Realistic Test-Time Adaptation (AAAI 2024 · similarity 0.87)