Skip to main content
Search

# Test-Time Adaptation

10 related articles

Subscribe to updates
  1. arXiv

    World-model test-time adaptation stops fitting prediction alone: OpTTA adds an optimistic planning term and beats AdaJEPA on final success in 14 of 15 shifted settings

    Pretrained latent world models are usually frozen at test time, and distribution shift can degrade planning with them; updating only on prediction loss fits the outcomes of the agent's own actions and can stay wrong about actions that would reach the goal. OpTTA adds the planner's goal-reaching cost on candidate imagined rollouts to the prediction loss, biasing adaptation toward model parameters that are optimistic about reaching the goal. Across three JEPA world models and 15 shifted settings, adapting only the predictor raises final success over the frozen model by 13.5 percentage points on average, versus 6.9 for prediction-only AdaJEPA; with the encoder also adapted, the gains are 13.5 versus 7.5 points.

  2. arXiv

    Frozen world models skip retraining: Library of Hands calibrates only the "hand" and lifts success from 28.6% to 62.3% when hands change and return

    When a camera moves or a controller is replaced, the same command moves the robot in a different direction or by a different magnitude, while a frozen world model still plans under its training-time action convention, so every plan is expressed in the wrong action frame. Library of Hands (LoH) trains nothing and leaves the world model frozen: it uses the frozen predictor to infer the applied action from two successive camera images, estimates the map from commands to applied actions, and stores each identified map in a library for reuse. In closed-loop tests where hands change and come back, LoH raises a frozen planner's success from 28.6% to 62.3%, and loses only 0.6 percentage points when nothing changes.

  3. arXiv

    Confidence estimators drift after agentic RL: StaR-TTA reorders them by trajectory stability, reaching 0.7208 average AUROC on four QA benchmarks

    Agentic reinforcement learning shifts the reasoning trajectories of large language models, and trajectory-based confidence estimators trained on a base model lose much of their ranking ability after that shift. StaR-TTA treats trajectory stability as a label-free reliability signal: it learns the link between stability and correctness on labeled base trajectories, then uses stability rankings from unlabeled RL trajectories to correct confidence at test time. Across four QA benchmarks it reaches an average AUROC of 0.7208, with gains on unseen datasets, different model families and scales, and both PPO and GRPO.

  4. arXiv

    Query-conditioned distributional hypernetworks predict LoRA weight updates, where the distribution mean alone beats deterministic hypernetworks and weight sampling enables a new form of test-time scaling

    This work studies hypernetworks conditioned only on an LLM's input query to estimate LoRA, and introduces distributional hypernetworks that output a distribution over LoRAs via an end-to-end loss with a differentiable Monte Carlo approximation; results show that even the mean of the learned distribution can outperform deterministic hypernetworks, that sampling weight updates yields a form of test-time scaling distinct from token sampling with performance improving as more weight samples are considered and remaining stronger than corresponding token-sampling adaptation baselines, and that generated updates can transfer across queries.

  5. arXiv

    MAGEFormer embeds voxel spacing into positional encoding and attention, reaching the lowest HD95 on anisotropic abdominal CT segmentation across BTCV and FLARE 22

    The authors propose MAGEFormer, which embeds physical spacing constraints into representation learning: MASE calibrates positional frequencies by voxel spacing, GCA suppresses physically implausible cross-scale feature correlations, and GVV performs deterministic sub-voxel view aggregation at inference; under a unified 5-fold protocol on BTCV and FLARE 22 it attains the strongest boundary accuracy among compared methods (10.58 mm and 3.40 mm HD95) with consistent Dice gains, and its advantage grows with anisotropy severity.

  6. arXiv

    WISE-ATTA shifts active test-time adaptation from which sample to label to when to label, reaching lower error on ImageNet-C/R/K/A with fewer labels

    The work introduces budgeted active test-time adaptation (ATTA), in which labels are available for only a fraction of batches in a long test stream, and proposes WISE-ATTA: lightweight online signals (the fraction of low-entropy predictions) plus a sliding-history quantile threshold and a budget-debt controller decide when to request supervision, while prediction drift relative to an EMA anchor selects a single sample to label; on ImageNet-C under CTTA/FTTA and on ImageNet-R/K/A it matches or improves on recent ATTA methods at one label per batch and reports up to 50% fewer labels.

  7. arXiv

    BiasRecon hits top PSNR in cross-anatomy, cross-center and cross-modality MRI reconstruction with fewer than 100 tunable parameters

    The work proposes BiasRecon, a bias-calibrated adaptation framework grounded in a minimal-intervention principle, which alternates frequency-guided prior calibration (learnable scalars α and β modulating low- and high-frequency components at each U-Net skip layer, adding only 2L parameters), score-based denoising, and adaptive regularization that uses Stein's Unbiased Risk Estimator (SURE) to tune the regularization parameter γ; trained on FastMRI knee (973 volumes) and evaluated without retraining on FastMRI-Brain (cross-anatomy), Stanford Knee (cross-center) and CMRxRecon (combined shift) under 8× Gaussian 1D undersampling, it improves PSNR over DDS by +1.19 dB, +2.06 dB and +1.60 dB respectively, and is best in 10 of 12 sampling/acceleration settings.

  8. arXiv

    HD-TTA chooses between competing 'compact or inflate' hypotheses, cutting HD95 by about 6.4 mm and raising precision by over 4% in cross-domain brain tumor segmentation

    The work proposes Hypothesis-Driven Test-Time Adaptation (HD-TTA): with a frozen nnU-Net v2 backbone and optimization only over test-sample logits, a Gatekeeper first decides whether a case needs refinement, two competing geometric hypotheses (compact denoising vs. diffuse recovery) are generated in parallel, and a representation-guided selector picks the safest output using intrinsic texture consistency; trained on BraTS 2023 GLI and evaluated with strictly fixed hyperparameters on unseen pediatric (PED) and meningioma (MEN) target domains, HD-TTA keeps Dice comparable while improving safety metrics, reducing HD95 from 70.96 mm to 64.55 mm (about 6.4 mm) and raising precision from 15.36% to 19.64% on MEN relative to the strongest baseline TCA.

  9. VesselSim trains a 3D segmentation model on synthetic vessels only, matching vascular foundation models zero-shot on real brain and kidney MR/CT

    VesselSim proposes a two-stage framework for universal 3D blood vessel segmentation: a stochastic, geometry-driven vascular simulation (recursive branching, curvature-controlled growth, collision-aware topology) with domain-randomized intensity synthesis generates 16,500 anatomically plausible 3D angiographic volumes, a 3D U-Net is trained solely on this synthetic data, and a test-time adaptation strategy via a self-supervised mask reconstruction decoder adapts at inference, achieving zero-shot performance competitive with state-of-the-art vascular segmentation foundation models on multiple real datasets spanning MR and CT and several anatomical regions including the brain and kidneys, without real annotated data during training.

  10. SurgDepth: Test-Time Adaptation of Depth Foundation Models for Surgical Scene Understanding

    The work proposes SurgDepth, a test-time adaptation framework that adapts natural-image-pretrained depth foundation models to surgical endoscopy without any labeled surgical data, combining two self-supervised signals (stereo photometric consistency and flip equivariance, the latter requiring only single images) with selective decoder adaptation and two tuning-free reliability mechanisms (progress-aware anchor regularization and a structural-trust safeguard for poorly-illuminated sequences); the authors report AbsRel reductions of up to 63% with stereo pairs and 42% without, improvement in four of five cross-domain evaluation settings, generalization across seven foundation models, and the lowest AbsRel (0.

You’re all caught up