test-time-adaptation

SkillShift-Bench introduces the first benchmark for evaluating web agent skill learning under controlled open-environment shifts (contextual, perceptual, and execution), revealing systematic robustness degradation and proposing a test-time adaptation method to improve shifted skill execution without relearning.

提出 Dual-Branch Calibration Module (DBCM)，通过共享适配器与低秩通道适配器的双分支设计，在时序预测的测试时自适应中处理非平稳分布偏移，并在多个数据集、预测跨度和骨干模型上超越现有方法。

论文提出 Test-Time Memory Evolution (TTME)，一种即插即用的测试时适应方法，通过持续更新外部记忆来指导科学发现而无需更新模型权重。实验表明 TTME 在五个数学任务上全面优于测试时训练 (TTT)，同时将计算成本降低 82.8–84.6% FLOPs 和 47.7–58.4% 运行时间。

论文提出 Test-Time Memory Evolution（TTME），一种通过持续更新外部记忆而非模型权重来进行测试时自适应的方法，在五个数学任务上优于测试时训练（TTT），同时降低 82.8-84.6% 的 FLOPs 和 47.7-58.4% 的运行时间。

CF-NODE 提出了一种反事实空间节点干预的测试时间自适应框架，用于跨域糖尿病视网膜病变分级，在五个外部数据集上显著优于 TENT、EATA、SAR 和 T3A 等方法。

该论文提出 Error Feedback Calibration（EFC），一种完全测试时的确定性后处理校正框架，通过 Causal Error Diagonal 与 Cumulative Error Transport 将已揭示的预测误差转化为对未来预测的因果校正，从而在多种时序预测骨干和数据集上提升性能并抑制过度校正。

AdaExploreBench introduces a benchmark testing whether LLM agents can proactively discover hidden environmental rules and adapt to unannounced rule drift; experiments on eight leading LLMs show current models struggle to acquire relevant evidence and detect stale knowledge.

The paper proposes Energy Ring, a method that stabilizes the energy gap between known and unknown classes for reliable test-time discovery, using a lightweight adapter to enlarge norm-based gaps and a perturbation consistency module for compact unknown-class clusters.

REFIT introduces a label-free input calibration method for frozen activity-recognition models that corrects wearable sensor placement shifts (e.g., switching wrists or reattaching straps) by fitting axis-transform families to match training statistics, outperforming existing test-time adaptation approaches without retraining.

CMU paper introduces "harness learning," where an RL-trained proposer model revises an agent's executable harness from execution feedback without changing solver weights, achieving generalizable test-time adaptation across unseen tasks with a 4B proposer outperforming its 35B teacher.

InterEvolve introduces test-time evolution of reward programs for humanoid loco-manipulation, using an object-aware forward-backward behavioral foundation model and an LLM agent that revises staged reward programs in-context to unlock untrained skills, with evolved behaviors deployed autonomously on a physical Unitree G1 robot.

This position paper argues that offline RL should shift from learning static deployment policies to learning adaptive policy priors that can improve through subsequent online interaction, proposing a framework called Adaptive Offline RL (AORL) and highlighting Bayesian offline RL as a principled approach to preserving epistemic uncertainty.

WISE-ATTA introduces budgeted active test-time adaptation, deciding when to query labels over a test stream using lightweight online signals and a drift-based sample selection criterion, achieving competitive results with substantially fewer labels on ImageNet-C and ImageNet-R/K/A.

The paper introduces MASA, a method that uses frozen multimodal large language models to break the self-referential loop in wild test-time adaptation by providing structured semantic descriptions for more reliable adaptation.

This paper proposes TASCO, a framework for test-time adaptation in LLMs that incorporates local stability into confidence-based optimization to improve reasoning accuracy and token efficiency without updating model parameters.

CASTER introduces a gradient-free test-time adaptation method for frozen models, using affine statistics transport and a certificate to decide when to apply adaptation for improved performance without model updates.

Safin-1 presents a memory-routing architecture that embeds safety as an internal, evolving state in foundation models, enabling adaptive safety improvements without external constraints.

PETA proposes a parameter-efficient framework for test-time adaptation in virtual screening, improving performance by updating only LayerNorm parameters during inference.

This paper proposes ReNC, a method for open-world test-time adaptation that leverages neural collapse as a structural prior to filter out-of-distribution samples and refine prototypes for reliable model adaptation.

PRISM is a training-free test-time adaptation method that reverses low-rank affine noise distortions in audio-text models using frozen text prototypes, showing significant improvements under severe acoustic noise.