A tour of inference-time factuality methods for LLMs - DoLa, SLED, DELTA, Lookback Lens and Think@n - that fight hallucination by changing decoding instead of retraining.
A survey of out-of-distribution detection for vision-language models like CLIP, covering training-free scoring, prompt-based detection and LLM-augmented pipelines.
Reviewing Reasoning's Razor, a paper showing extended chain-of-thought reasoning can hurt LLM recall at the low-FPR thresholds that matter for safety and hallucination detection.
Paper review showing that keeping LoRA modules unmerged at inference yields embeddings whose Mahalanobis distance is a strong near-OOD detector for fine-tuned LLMs.
An overview of weight space learning, where neural network weights themselves become data for predicting generalization, robustness and even synthesizing new models.
A practical MLOps tutorial on developing ML workloads in Docker and deploying them to HPC clusters with Singularity (Apptainer) for rootless, GPU-ready execution.
GROOD detects out-of-distribution inputs by measuring gradient sensitivity to an artificial OOD prototype, leveraging neural collapse to improve safety in deep learning systems.
A deep-dive summary of an LLM hallucination survey, with a taxonomy of factuality and faithfulness errors plus the root causes, detection and mitigation techniques behind them.
Paper review using topological data analysis and persistent homology on penultimate-layer embeddings to characterise OOD samples that resist a network's topology simplification.
