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

IT$^3$: Idempotent Test-Time Training

ICML 2025 · 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 IT3, an on-the-fly test-time training method that adapts a ZigZag-style network to each test input by minimizing its distance from idempotence, using a frozen or EMA copy for the second pass.

Problem

Deep models degrade under distribution shift at test time, and existing test-time training and adaptation methods often need extra data or domain-specific auxiliary tasks.

Approach

Uses the discrepancy ||f(x,f(x,0))-f(x,0)|| from a ZigZag-style network as a task-agnostic self-supervised test-time loss, minimized per input with a frozen copy of the network for the second application to avoid gradient pathways that expand the manifold.

Key results

• Fig. 2 shows OOD inputs move from near-random idempotence distances toward in-distribution levels after IT3 test-time optimization.
• Rebuttal reports higher corruption severity in ImageNet-C correlates with larger distance from idempotence.

Measured effects reported in the paper

Limitations

• No formal proof that test-time adaptation preserves idempotence on OOD samples; justification is largely heuristic.
• Predictions can be reinforced when the initial prediction is wrong, though authors report this is rare with the frozen network.
• Requires architecture modification to accept the label input, unlike BN-based methods such as TENT.

Experimental setup

Datasets: CIFAR-10 ImageNet-C

Benchmarks: CIFAR-10 ImageNet-C aerial road segmentation aerodynamics prediction tabular data

Metrics: accuracy error

Baselines: TTT (Sun et al. 2020) ActMAD TTT++ TENT EATA MEMO DEYO

Builds on: ZigZag test-time training ActMAD IGN

Closest papers by meaning

• A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.86)
• Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.85)
• Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.85)
• DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.85)
• Test-Time Adaptation with Binary Feedback (ICML 2025 · similarity 0.85)
• ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.84)