Research map / Label-efficient and robust learning / Test-time adaptation
MIRA: Mutual Information guided calibration for Reliable Test-Time Adaptation
Hyeongyu Kim, Mingyeong Jang, Youngjun Song, GeonHui Han, Dosik Hwang
Read the paperCheck your idea against 137,217 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 MIRA, a source-free online test-time adaptation method that uses mutual information from lightweight last-layer Gaussian perturbations to guide calibrated, more reliable adaptation under distribution shift.
Problem
Test-time adaptation with only unlabeled target data; standard entropy-based methods rely on deterministic confidence and can reinforce overconfident errors under distribution shift.
Approach
Uses lightweight last-layer Gaussian perturbations to probe predictive uncertainty via mutual information, instead of deterministic entropy, to guide source-free online adaptation and calibration.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| attach MIRA reliability gate | TENT · CIFAR-100-C ResNeXt | accuracy (%): 67.84 -> 68.69 | Improved |
| attach MIRA reliability gate | EATA · ImageNet-C ResNet-50 | accuracy (%): 40.02 -> 43.18 | Improved |
| attach MIRA reliability gate | ROID · CIFAR-100-C ResNeXt | accuracy (%): 69.16 -> 69.21 | No clear effect |
| attach MIRA reliability gate | SAR · ImageNet-C ResNet-50 | accuracy (%): 36.72 -> 39.28 | Improved |
| minimize MI directly as the loss (ablation) | MIRA+TENT · ImageNet-C (loss ablation) | accuracy: 0.3421 -> 0.2977 | Hurt |
| use HE only or EH only instead of TENT entropy (ablation) | MIRA+TENT · ImageNet-C (loss ablation) | accuracy: 0.3421 -> 0.3408 | No clear effect |
| combine HE and EH losses (ablation) | MIRA+TENT · ImageNet-C (loss ablation) | accuracy: 0.3408 -> 0.3637 | Improved |
| add MI weight w_MI to HE+EH (ablation) | MIRA+TENT · ImageNet-C (loss ablation) | accuracy: 0.3637 -> 0.3658 | No clear effect |
| add confidence gate on top of MI weight (ablation) | MIRA+TENT · ImageNet-C (loss ablation) | accuracy: 0.3658 -> 0.3699 | Improved |
| perturb last 2-3 layers instead of classifier only (ablation) | MIRA · ImageNet-C ResNet-34 | ECE: 0.0458 -> 0.0649 | Hurt |
| increase Monte Carlo samples K from 100 to 1000 (ablation) | MIRA · ImageNet-C ResNet-50 | accuracy (%): 29.55 -> 29.76 | No clear effect |
| add MIRA to continual CoTTA | CoTTA · CIFAR-100-C continual | accuracy (%): 64.16 -> 65.20 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TENT — Tent: Fully Test-Time Adaptation by Entropy Minimization (ICLR 2021) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Baselines: entropy-based test-time adaptation methods
Builds on: entropy minimization
Closest papers by meaning
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.88)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.87)
- Unified Entropy Optimization for Open-Set Test-Time Adaptation (CVPR 2024 · similarity 0.87)
- Protected Test-Time Adaptation via Online Entropy Matching: A Betting Approach (NeurIPS 2024 · similarity 0.87)
- Bayesian Weight Enhancement with Steady-State Adaptation for Test-time Adaptation in Dynamic Environments (ICML 2025 · similarity 0.87)
- TEA: Test-time Energy Adaptation (CVPR 2024 · similarity 0.87)
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.
