Efficient $Q$-Learning and Actor-Critic Methods for Robust Average Reward Reinforcement Learning

A recent thesis explores self-attention training for tabular classification through Optimal Transport (OT), developing an OT-based alternative that tracks the evolution of self-attention layers during training using discrete OT metrics like Wasserstein distance and Monge gap. The study reveals that while the final self-attention mapping approximates the OT optimal coupling, the training process remains inefficient.

A new implementation of next-generation reservoir computing (NGRC) has been introduced, designed for modeling dynamical systems using time-series data. This method employs a pseudorandom nonlinear projection of time-delay embedded inputs, enabling flexible feature-space dimensions and demonstrating effectiveness in tasks like attractor reconstruction and bifurcation diagram estimation, even with partial and noisy measurements.

A recent study published on arXiv evaluates the effectiveness of strong data augmentations in self-supervised contrastive learning for medical image segmentation, revealing that existing augmentations do not consistently enhance performance. The research suggests alternative augmentation techniques that yield better results in semantic segmentation tasks involving medical images.

The introduction of D4RT marks a significant advancement in the field of computer vision, focusing on the efficient reconstruction of dynamic scenes from video. This innovative feedforward model employs a unified transformer architecture to infer depth, spatio-temporal correspondence, and camera parameters from a single video, streamlining the process and enhancing performance.

A new study has introduced methods utilizing beam search to enhance consistency-based uncertainty quantification in large language models (LLMs), addressing issues with multinomial sampling that often leads to duplicates and high variance in uncertainty estimates. The research demonstrates improved performance across six question-answering datasets, establishing a theoretical lower bound for beam search effectiveness.

A new approach to supervised learning has been introduced, leveraging in-context learning with attention to enhance predictive accuracy for tabular data. This method adapts techniques like lasso regression and gradient boosting to create personalized models that focus on relevant training examples, improving interpretability and flexibility in predictions.

A new study presents an optimized behavior model for multi-agent driving simulation, focusing on enhancing realism and computational efficiency. The model utilizes an instance-centric scene representation and a query-centric context encoder, enabling effective interaction modeling among traffic participants. Adversarial Inverse Reinforcement Learning is employed to balance robustness and realism during training.

The Financial Instruction Following Evaluation (FIFE) introduces a benchmark to assess the instruction-following capabilities of language models (LMs) in financial analysis, highlighting their struggles with complex instructions. The benchmark consists of 88 human-authored prompts and employs a verification system for fine-grained reward signals, evaluating 53 models in a zero-shot setting.