A new arXiv paper examines whether instruction tuning gives language models a general-purpose ability to follow instructions or whether the behavior emerges from coordinated, task-specific components. The authors argue the evidence points to skillful coordination rather than one universal mechanism. The work adds to ongoing debate about what instruction tuning actually changes inside a model.
A new arXiv preprint proposes probing methods to trace where large language models source their knowledge, aiming to separate two kinds of hallucination. The authors distinguish faithfulness violations, where a model mishandles context it was given, from factuality violations, where its answers stem from incorrect stored knowledge. The work targets better attribution of model outputs to internal knowledge versus supplied context.
A new preprint introduces a decomposition method that extends feature-attribution explanations from single-number predictions to functional or multivariate outputs, such as demand forecasts that vary over time. The approach works in a Hilbert space so that the influence of each input feature can be separated across the whole output trajectory rather than summarized by one score. The authors position it as a general framework for settings where model predictions are curves or vectors instead of scalars.