Active Learning’s Leap Forward: From Edge AI to Unreliable Labels and Beyond

Latest 50 papers on active learning: Oct. 27, 2025

Active learning (AL) is undergoing a significant renaissance, driven by the escalating costs of data annotation and the increasing complexity of AI models. It offers a powerful paradigm where models intelligently select the most informative data points to be labeled, thereby maximizing learning efficiency with minimal human effort. This surge in interest is particularly evident in recent research, which showcases AL’s transformative impact across diverse domains—from bolstering the robustness of deep learning systems to optimizing computational processes and enabling novel applications in emerging fields.### The Big Idea(s) & Core Innovationslatest breakthroughs in active learning revolve around enhancing its reliability, efficiency, and applicability in challenging real-world scenarios. A key theme is making AL more robust to unreliable labels and complex data structures. For instance, Reliable Active Learning from Unreliable Labels via Neural Collapse Geometry by Atharv Goel, Sharat Agarwal, Saket Anand, and Chetan Arora (IIIT Delhi & IIT Delhi) introduces NCAL-R, an AL framework that leverages neural collapse geometry to handle noisy labels. This innovation mitigates noise amplification, a common pitfall, by incorporating geometric insights into sample selection through “Class-Mean Alignment Perturbation” and “Feature Fluctuation” scores.significant stride is the unification of AL with other crucial ML tasks. The paper A Unified Approach Towards Active Learning and Out-of-Distribution Detection by Sebastian Schmidt and colleagues (Technical University of Munich, BMW Group, Sprin-D) presents SISOM, a method that simultaneously performs active learning and out-of-distribution (OOD) detection. This simplifies the application lifecycle by eliminating the need for separate OOD design, achieving top performance in both tasks by intelligently leveraging ambiguous unlabeled samples.in data-scarce and high-stakes environments is also a major focus. Young In Kim and Andrea Agiollo (Purdue University, Delft University of Technology), in SAMOSA: Sharpness Aware Minimization for Open Set Active learning, introduce an open-set AL algorithm that uses sharpness-aware minimization to select informative atypical samples near decision boundaries, boosting accuracy by up to 3% without extra computational cost. Similarly, Budget-constrained Active Learning to Effectively De-censor Survival Data by Ali Parsaee and co-authors (University of Alberta) tackles the unique challenge of censored survival data, demonstrating superior performance by strategically acquiring labels under budget constraints.### Under the Hood: Models, Datasets, & Benchmarksadvancements are underpinned by sophisticated models, novel datasets, and rigorous benchmarks that push the boundaries of what’s possible with active learning. Here’s a glimpse:NCAL-R Framework (from Reliable Active Learning from Unreliable Labels via Neural Collapse Geometry): Leverages neural collapse geometry with two novel scores (Class-Mean Alignment Perturbation and Feature Fluctuation) for robust sample selection, showing consistent improvements in accuracy and out-of-distribution generalization across multiple benchmarks.SISOM (from A Unified Approach Towards Active Learning and Out-of-Distribution Detection): A unified framework for joint OOD detection and AL, utilizing latent space analysis and uncertainty-diversity fusion. Evaluated extensively on common image benchmarks and OpenOOD benchmark, achieving top performance in both tasks. Code available at https://www.cs.cit.tum.de/daml/sisom.SAMOSA (from SAMOSA: Sharpness Aware Minimization for Open Set Active learning): An open-set active learning algorithm that integrates sharpness-aware minimization and data typicality to identify and query informative atypical samples. Demonstrated on multiple datasets with up to 3% accuracy improvement. Code is available through a 4open.science repository.CUSAL (from Calibrated Uncertainty Sampling for Active Learning): An acquisition function prioritizing samples with the highest calibration error using a kernel-based estimator. Empirically validated on MNIST, F-MNIST, CIFAR-10, and SVHN, showing lower calibration and generalization errors. The paper itself contains the description.DMLE (from Dependency-aware Maximum Likelihood Estimation for Active Learning): A novel MLE approach that accounts for sample dependencies in active learning, correcting the i.i.d. assumption of conventional MLE. Provides superior performance in early learning cycles on benchmark datasets. Code is publicly available at https://github.com/neu-spiral/DMLEforAL.TUTORBENCH (from TutorBench: A Benchmark To Assess Tutoring Capabilities Of Large Language Models): The first comprehensive benchmark for evaluating LLM tutoring capabilities, featuring 1490 samples across STEM subjects with a rubric-based evaluation system. Dataset available on Hugging Face.Active Attacks Framework (from Active Attacks: Red-teaming LLMs via Adaptive Environments): An RL-based red-teaming algorithm combining active learning and GFlowNet multi-mode sampling to generate diverse adversarial prompts for LLMs. Code available at https://github.com/mila-udem/active-attacks.LogAction (from LogAction: Consistent Cross-system Anomaly Detection through Logs via Active Domain): A framework for cross-system anomaly detection using active domain adaptation, achieving 93.01% F1 with only 2% labeled data. Resources and code are available at https://logaction.github.io.PCoreSet (from PCoreSet: Effective Active Learning through Knowledge Distillation from Vision-Language Models): A selection strategy for active learning that leverages vision-language models (VLMs) and knowledge distillation, maximizing coverage in probability space for improved student model performance. Code at https://github.com/erjui/PCoreSet.NFPF (from Unsupervised Active Learning via Natural Feature Progressive Framework): An unsupervised active learning framework that identifies informative samples using Reconstruction Difference (RD) and inter-model discrepancy. Achieves 20x fewer training steps on CIFAR-100. Code: https://github.com/Legendrobert/NFPF.DAM (from DAM: Dual Active Learning with Multimodal Foundation Model for Source-Free Domain Adaptation): A framework that integrates Vision-and-Language models with active learning for source-free domain adaptation. Achieves state-of-the-art performance using Dual-Focused active Supervision (DFS) and Alternative Distillation (ADL). Code at https://github.com/xichen-hit/DAM.### Impact & The Road Aheadinnovations collectively underscore active learning’s pivotal role in overcoming the data labeling bottleneck, a perennial challenge in AI/ML. From reducing annotation costs in medical imaging with systems like TissueLab (University of Pennsylvania) in A co-evolving agentic AI system for medical imaging analysis, to enabling personalized human activity recognition with FedAR (University of Milan) in Personalized Semi-Supervised Federated Learning for Human Activity Recognition, AL is making AI more accessible, efficient, and reliable. The integration of AL with foundation models, as seen in Google Research’s On-the-Fly OVD Adaptation with FLAME for few-shot object detection in remote sensing, and in Vector Institute’s Automated Capability Evaluation of Foundation Models for scalable model evaluation, signifies its potential for a new era of adaptable AI.future of active learning lies in further improving its theoretical guarantees, enhancing its robustness to diverse data challenges (such as unreliable labels and complex dependencies), and integrating it more seamlessly into real-world applications. The continued research into multi-fidelity approaches like Murray Cutforth and colleagues’ (Stanford University, University College London) Multi-fidelity Batch Active Learning for Gaussian Process Classifiers and adaptive frameworks such as ETH Zurich’s PANAMA (PArametric Neural Amp Modeling with Active Learning) in Parametric Neural Amp Modeling with Active Learning promises to unlock even greater efficiencies and capabilities. As AI systems become more complex, active learning will remain an indispensable tool for building smarter, more data-efficient, and ultimately, more impactful solutions.