Research map / Label-efficient and robust learning / Test-time adaptation
TRUST: Test-Time Refinement using Uncertainty-Guided SSM Traverses
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 TRUST, a test-time adaptation method for VMamba that adapts SS2D parameters across the top-K lowest-entropy traversal permutations with pseudo-labels and averages the adapted weights.
Problem
VMamba-style state space models lose generalization under distribution shift, and existing test-time adaptation methods are not designed to exploit SSM architectural properties such as traversal order.
Approach
Generate multiple SS2D cross-scan traversal permutations, select the lowest-entropy ones, adapt the Mamba-specific SS2D parameters on pseudo-labels for each, and average the resulting weights to get one model used with the default traversal at inference.
Key results
- TRUST reaches 77.5% mean accuracy on CIFAR10-C (level 5), versus 65.9% for source only and 66.5% for Tent.
- TRUST reaches 56.1% mean accuracy on ImageNet-C, versus 38.7% for source only and 41.7% for SHOT.
- Using high-entropy instead of low-entropy traversal permutations drops CIFAR10-C accuracy to 53.4%, below the 65.9% source-only baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TRUST traversal permutations and weight averaging instead of no adaptation | VMamba · CIFAR10-C | Top-1 accuracy (%): 65.9 -> 77.5 | Helped |
| Use TRUST traversal permutations and weight averaging instead of no adaptation | VMamba · CIFAR100-C | Top-1 accuracy (%): 41.2 -> 54.3 | Helped |
| Use TRUST traversal permutations and weight averaging instead of no adaptation | VMamba · ImageNet-C | Top-1 accuracy (%): 38.7 -> 56.1 | Helped |
| Use TRUST traversal permutations and weight averaging instead of no adaptation | VMamba · ImageNet-S | Top-1 accuracy (%): 31.4 -> 41.5 | Helped |
| Use TRUST traversal permutations and weight averaging instead of no adaptation | VMamba · ImageNet-V2 | Top-1 accuracy (%): 62.2 -> 64.0 | Helped |
| Use TRUST traversal permutations and weight averaging instead of no adaptation | VMamba · ImageNet-R | Top-1 accuracy (%): 31.3 -> 44.3 | Helped |
| Use TRUST traversal permutations and weight averaging instead of no adaptation | VMamba · PACS | Top-1 accuracy (%): 66.7 -> 69.9 | Helped |
| Add traversal-permutation-based weight averaging to SS2D-only adaptation (ablation) | TRUST naive on CIFAR10-C · CIFAR10-C | Top-1 accuracy (%): 74.2 -> 77.5 | Helped |
| Increase traversal permutations from 2 to 8 (ablation) | TRUST · CIFAR10-C | Top-1 accuracy (%): 75.6% to 77.6% | Helped |
| Increase traversal permutations from 2 to 8 (ablation) | TRUST · CIFAR100-C | Top-1 accuracy (%): 51.7% to 54.7% | Helped |
| Increase traversal permutations from 2 to 8 (ablation) | TRUST · ImageNet-C | Top-1 accuracy (%): 54.4% to a peak of 56.1% at six permutations before slightly dropping to 55.5% at eight | Mixed |
| Average traversal model weights rather than predictions (ablation) | TRUST · CIFAR10-C | Top-1 accuracy (%): 68.1% -> 75.6% | Helped |
Surprising findings
- Using the top-k highest-entropy traversal permutations drops accuracy by 12.5 points below source-only on CIFAR10-C.
- Standard augmentations (rotation, cropping, jitter) yield only modest gains over source-only, unlike the traversal-based method.
Limitations
- Applicable only to VMamba-style architectures with SS2D traversal orders, not CNNs or ViTs without modification.
- Adaptation cost scales linearly with the number of selected permutations in sequential mode.
- Pseudo-label updates may reinforce errors under severe or unseen shifts.
Experimental setup
Datasets: CIFAR10-C CIFAR100-C ImageNet-C PACS ImageNet-S ImageNet-V2 ImageNet-R Pascal VOC21 Pascal Context 59
Benchmarks: CIFAR10-C CIFAR100-C ImageNet-C ImageNet-S ImageNet-V2 ImageNet-R PACS
Metrics: top-1 accuracy
Baselines: Source only ETA LAME SAR SHOT Tent FOA TRUST naive
Builds on: Tent VMamba SWAD SHOT Model Soups
Closest papers by meaning
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.86)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.86)
- VMamba: Visual State Space Model (NeurIPS 2024 · similarity 0.85)
- Mamba-Adaptor: State Space Model Adaptor for Visual Recognition (CVPR 2025 · similarity 0.85)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- Implicit Modeling for Transferability Estimation of Vision Foundation Models (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.