transfer-learning

AgentMercury shows that training AI agents in simulated business environments generated from plain descriptions can transfer effectively to evaluation benchmarks, even if the training worlds are unrelated. The system improved performance through fine-tuning on construction traces.

This paper investigates AI learning and conceptual transfer in the Game of Hidden Rules, focusing on reinforcement learning with Transformer-based A2C framework, rule difficulty analysis, transfer learning, and generalization.

This paper proposes MeanFlow-Transfer (MF-T) and Continuous Adversarial MeanFlow (CAMF) to unify the adaptation and acceleration of pretrained diffusion and flow models, enabling high-quality few-step generation on new domains with limited data.

This paper analyzes curriculum learning in large language models by examining optimization dynamics across difficulty levels, proposing a new method called Transfer-aware Dynamic Curriculum Sampling (TDCS) that dynamically adjusts training data based on transfer relationships.

This paper proposes SW-ProxyCE, a zero-query adversarial attack framework that transfers from public EEG encoders to private downstream models, demonstrating security risks in EEG foundation models.

The paper introduces CEDAR-GRPO, a process-aware reinforcement learning framework that enhances abductive reasoning in large language models by combining final-answer correctness with abductive rewards, demonstrating transferable improvements across various unseen tasks.

This paper proposes the Transferable Latent Operator (TLO) for fluid dynamics, enabling zero-shot generalization from Eulerian field prediction to Lagrangian particle rollout without Lagrangian supervision.

The paper analyzes the safety of per-agent policy composition in multi-agent reinforcement learning, proving that independent composition can fail, and proposes MA-USFA, a hierarchical method that enables safe and flexible successor-feature transfer in cooperative MARL.

This paper proposes Hierarchical Federated Transfer Learning (HFTL) for Digital Twin-based Vehicular Ad hoc Networks, addressing data heterogeneity and sparsity via vehicle clustering and a data quality score mechanism to defend against malicious vehicles.

This paper proposes a market-information-aware gated LoRA framework to adapt the Chronos-2 time-series foundation model for day-ahead electricity price forecasting, improving cross-market transferability on Chinese provincial markets.

This paper proposes an LLM-assisted semantic stop embedding approach for reinforcement-learning-based bus holding control, reducing headway variability, bunching events, and passenger waiting time in simulations.

This paper proposes a transferable autologistic model for predicting rare equipment failures across heterogeneous sensor configurations, evaluating it on a synthetic refrigerator dataset.

This paper studies whether prompt-side agent playbooks (procedures distilled from prior runs) transfer across settings, finding that frozen transfer works only conditionally depending on compatibility with target deployment, with mixed results on ALFWorld, TAU2-Bench, and XBench-DeepSearch.

This paper introduces EddyFlow, a deep learning framework for kilometer-scale sea surface temperature downscaling that balances predictive accuracy, scale-dependent structure, and regional generalization. It achieves strong zero-shot performance and near-ideal spectral fidelity across multiple ocean regions.

This paper describes Team uOttawa's system for the EvaLatin 2026 shared task on Named Entity Recognition for Classical Latin, using prompt engineering of commercial LLMs (Gemini 2.5 Pro and Claude Sonnet 4-5) to achieve first place in both coarse- and fine-grained NER subtasks.

A research paper evaluating CHAP, a CNN-BiLSTM model, for classifying sedentary behavior from wearable accelerometer data, focusing on transfer from hip to wrist placement and the benefits of finetuning.

This paper proposes Frontier Learning, a framework that combines representations and predictions from multiple black-box and white-box pretrained models to construct a unified target-domain representation, guaranteeing performance no worse than any individual reuse baseline under distribution shift. Evaluations on visual domain adaptation and clinical mortality prediction show consistent gains over strong baselines.

This paper introduces a Capability-Driven Multimodal Scaling Law that predicts VLM performance from LLM textual benchmark capabilities, enabling principled backbone selection across model families without expensive training sweeps.

This arXiv paper introduces ProGFM, a Propagation-aware Graph Foundation Model that treats propagation relationships between edges and feature dimensions as transferable knowledge units, enabling adaptive aggregation and improved cross-domain generalization.

This paper introduces Knowledge-Geometry Decoupling (KGD), a method for pretrain-then-transfer in streaming recommendation systems. It separates pretrained behavioral knowledge from task-specific geometry, enabling continual model refresh without interference, and reports 4-12% improvements over baselines plus successful deployment at Shopee.