We introduce AnoLLM, a novel framework that leverages large language models (LLMs) for unsupervised tabular anomaly detection. By converting tabular data into a standardized text format, we further adapt a pre-trained LLM with this serialized data, and assign anomaly scores based on the negative log likelihood generated by the LLM. Unlike traditional methods that can require extensive feature engineering, and often lose textual information during data processing, AnoLLM preserves data integrity and streamlines the preprocessing required for tabular anomaly detection. This approach can effectively handle mixed-type data, especially those containing textual features. Our empirical results indicate that AnoLLM delivers the best performance on six benchmark datasets with mixed feature types. Additionally, across 30 datasets from the ODDS library, which are predominantly numerical, AnoLLM performs on par with top performing baselines.

Addressing the critical need for robust safety in Large Language Models (LLMs), particularly against adversarial attacks and in-distribution errors, we introduce Reinforcement Learning with Backtracking Feedback (RLBF). This framework advances upon prior methods, such as BSAFE, by primarily leveraging a Reinforcement Learning (RL) stage where models learn to dynamically correct their own generation errors. Through RL with critic feedback on the model's live outputs, LLMs are trained to identify and recover from their actual, emergent safety violations by emitting an efficient "backtrack by x tokens" signal, then continuing generation autoregressively. This RL process is crucial for instilling resilience against sophisticated adversarial strategies, including middle filling, Greedy Coordinate Gradient (GCG) attacks, and decoding parameter manipulations. To further support the acquisition of this backtracking capability, we also propose an enhanced Supervised Fine-Tuning (SFT) data generation strategy (BSAFE+). This method improves upon previous data creation techniques by injecting violations into coherent, originally safe text, providing more effective initial training for the backtracking mechanism. Comprehensive empirical evaluations demonstrate that RLBF significantly reduces attack success rates across diverse benchmarks and model scales, achieving superior safety outcomes while critically preserving foundational model utility.

It’s commonplace in modern deep learning to achieve SOTA performance by fine-tuning a large, pretrained base model. Recent successes in natural language processing, attributed in part to knowledge transfer from large, pretrained, transformer-based language models, have sparked a similar revolution in computer vision via the introduction of Vision Transformers. As modern deep neural networks increase in performance, they also tend to increase in size. Key issues associated with fine-tuning such enormous models include storage overhead, as well as memory and / or latency requirements. Parameter efficient fine-tuning is a fairly recent paradigm which has been evolving alongside massive neural networks in part to address these issues. We showcase the effectiveness of parameter efficient fine-tuning of vision transformers, and introduce a simple yet effective method for learning a non-uniform parameter allocation given a fixed budget. We demonstrate our approach across a range of benchmark tasks in image classification and semantic segmentation.

Most deep neural networks use simple, fixed activation functions, such as sigmoids or rectified linear units, regardless of domain or network structure. We introduce differential equation units (DEUs), an improvement to modern neural networks, which enables each neuron to learn a particular nonlinear activation function from a family of solutions to an ordinary differential equation. Specifically, each neuron may change its functional form during training based on the behavior of the other parts of the network. We show that using neurons with DEU activation functions results in a more compact network capable of achieving comparable, if not superior, performance when compared to much larger networks.

Rapid eye movement (REM) sleep behavior disorder (RBD) is a sleep disorder that features loss of atonia, or REM sleep without atonia (RSWA). RBD and RSWA are early manifestations of degenerative neurological diseases such as Parkinson’s and Lewy Body Dementia. Accurate diagnosis of RBD is crucial for proper treatment planning and is invaluable for early detection of these neurodegenerative diseases. The current gold standard diagnosis of RSWA is through manual visual scoring by a clinician, which is labor-intensive, costly and error-prone. We develop a novel, efficient, and objective method using deep learning to detect RSWA events from polysomnography signals using a large cohort of 692 patients. Unlike previous automated methods that generate only a binary patient diagnosis, our method detects the location and class of all RSWA events. This finer-grained analysis forms the basis for subsequent diagnosis, and allows the quantification of event duration and frequency which in turn can help quantify disease load.