zero-shot-learning

This paper proposes a large-scale multi-modal dataset (MMIO) for zero-shot industrial defect detection and introduces the Refined Text-Visual Prompt (RTVP) method, achieving state-of-the-art results on the benchmark.

GlucoFM-Bench evaluates time-series foundation models for blood glucose forecasting across 15 datasets, showing strong zero-shot/few-shot transfer by Chronos-2 and TimesFM but superior performance of a lightweight LSTM when full training data is available.

Researchers from University of Technology Sydney compare fine-tuned transformers (DistilBERT, RoBERTa) against zero-shot LLMs (Llama variants, Claude, Gemini) for classifying misinformation responses on Reddit, finding that fine-tuned RoBERTa achieves 0.62 macro-F1 versus 0.50 for the best zero-shot model. The study shows that task-specific fine-tuning outperforms larger generalist models, particularly for detecting belief propagation, and that safety-alignment artifacts in frontier models can degrade performance.

Introduces RESCAST-100K, a large-scale benchmark dataset for cross-domain residential load and indoor temperature forecasting, featuring simulated and real data to evaluate transfer learning, domain adaptation, and zero-shot generalization.

This paper introduces a zero-shot strategy for chart summarization using Program-of-Thoughts prompting, where lightweight visual language models (VLMs) generate Python programs to compute statistics, improving factual accuracy over existing methods.

AnyMo is a geometry-aware framework for setup-agnostic human motion modeling using physics-grounded IMU simulation and graph encoding, achieving significant improvements in zero-shot activity recognition, cross-modal retrieval, and motion captioning across multiple datasets.

This academic paper investigates using LLMs for zero-shot prediction of psychological well-being scores from spontaneous speech, evaluating 12 models and achieving high correlation with clinical metrics.

This paper introduces the Neural Rule Inducer (NRI), a foundation model for zero-shot logical rule induction that uses domain-agnostic statistical properties to generalize across tasks without retraining.

DiZiNER is a framework that uses disagreement between multiple LLMs to refine task instructions for zero-shot named entity recognition, achieving state-of-the-art results on 14 out of 18 benchmarks and significantly reducing the performance gap between zero-shot and supervised systems.

Researchers introduce Zero-shot World Models (ZWM), an approach that achieves visual competence comparable to state-of-the-art models while trained on minimal data (single child's visual experience) without task-specific training. This work demonstrates a path toward more data-efficient AI systems that match human developmental learning efficiency.