Research map / Continual learning / Domain and prompt-based
ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains
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 ReservoirTTA, a plug-in that maintains a reservoir of domain-specialized models, routed by online style-feature clustering, to make prolonged test-time adaptation robust to recurring and evolving domain shifts.
Problem
Test-time adaptation over prolonged streams where domains recur or evolve gradually causes error accumulation, catastrophic forgetting, and inter-domain interference in single-model TTA.
Approach
Maintain a pool of domain-specific models (per discovered style cluster) that are adapted independently, identify domains via online DP-Means-style clustering on VGG-19 log-variance style features, and predict with a soft-assignment-weighted ensemble of parameters.
Key results
- On CIFAR-10-C recurring CSC, ETA's error at visit 20 drops from 30.9% to 16.4% with ReservoirTTA.
- On ImageNet-C recurring CSC with ViT-B/16, ReservoirTTA reaches 31.9% error at visit 20 versus CoLA's 33.8%.
- EATA's ImageNet-C error at visit 20 under recurring CSC falls from 55.9% to 51.0% with ReservoirTTA (CNN backbone results also reported).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ReservoirTTA to ETA | ETA · CIFAR-10-C, recurring CSC, visit 20 | error (%): 30.9 -> 16.4 | Helped |
| Add ReservoirTTA to EATA | EATA · ImageNet-C, recurring CSC, visit 20 | error (%): 55.9 -> 51.0 | Helped |
| Add ReservoirTTA to ROID* | ROID* · ImageNet-C, recurring CSC, visit 20 | error (%): 55.5 -> 52.1 | Helped |
| Add ReservoirTTA to ETA | ETA · ImageNet-C, recurring CSC, visit 20, ViT-B/16 | error (%): 48.4 -> 31.9 | Helped |
| Add ReservoirTTA to DPCore-style prompt tuning using VPT | VPT · ImageNet-C, CCC, adaptation step 80k | error (%): 43.1 -> 39.2 | Helped |
| Add ReservoirTTA to TENT | TENT · Cityscapes→ACDC, Snow, visit 10 | mIoU (%): 52.0 -> 54.7 | Helped |
| Add ReservoirTTA to CoTTA* | CoTTA* · Cityscapes→ACDC, Fog, visit 10 | mIoU (%): 71.2 -> 72.7 | Helped |
| Add a Model Reservoir (ablation) | ReservoirTTA · CIFAR-100-C, recurring CSC, visit 20 | error (%): +1.9% | Helped |
| Add a Style Reservoir (ablation) | ReservoirTTA · ImageNet-C, CCC | accuracy: +1.2% | Helped |
| Add model-reservoir ensembling for prediction (ablation) | ReservoirTTA · ImageNet-C, CCC | accuracy: +1.5% | Helped |
| Add a Style Reservoir (ablation) | ReservoirTTA · CIFAR-100-C, recurring CSC and CDC | error | No clear effect |
| Use shallow VGG-19 layers for style extraction instead of middle-2 layers (ablation) | EATA + ReservoirTTA · CIFAR-100-C, recurring CSC, visit 20 | error (%): 28.75 -> 28.38 | Helped |
Surprising findings
- Style reservoir has limited effect on recurring CSC/CDC because frequent abrupt domain resets make styles easily separable, though it helps under gradual CCC shifts.
- TENT alone accumulates severe error (19.3% to 87.8% on CIFAR-10-C at visit 20), and plugging in ReservoirTTA only partially mitigates accumulation.
Limitations
- Up to 30% extra computation over lightweight methods such as EATA and ETA.
- Parameters are updated on every incoming batch regardless of convergence, with no adaptive update trigger.
- Some methods (RoTTA, PeTTA) are not integrated because their per-model memory banks make integration computationally prohibitive.
Experimental setup
Datasets: CIFAR-10-C CIFAR-100-C ImageNet-C DomainNet-126 PACS Cityscapes ACDC
Benchmarks: CIFAR-10-C CIFAR-100-C ImageNet-C DomainNet-126 PACS Cityscapes-to-ACDC
Metrics: average classification error (%) mIoU (%)
Baselines: TENT CoTTA RoTTA ETA SAR RDumb PeTTA EATA ROID CoLA DPCore BECoTTA
Builds on: EATA ETA CoTTA RDumb ROID DP-Means
Closest papers by meaning
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.90)
- Towards Robust Multimodal Open-set Test-time Adaptation via Adaptive Entropy-aware Optimization (ICLR 2025 · similarity 0.89)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.89)
- AETTA: Label-Free Accuracy Estimation for Test-Time Adaptation (CVPR 2024 · similarity 0.88)
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.