An updated arXiv paper presents a representation for spatial information called a stochastic map, along with methods for constructing it, extracting information from it, and revising it step by step as new data arrives. The work targets robotics settings where position estimates are uncertain and must be maintained incrementally rather than recomputed from scratch.
A new preprint examines how correctness is defined when vision-language models are used to make clinical predictions from medical images and electronic health records. The authors argue that current uncertainty estimation methods may be evaluated in ways that do not reflect whether a prediction is actually reliable. The work targets safer deployment by improving how unreliable outputs are detected.
A newly posted arXiv thesis investigates whether large language models can gauge the reliability of the code they produce, addressing the problem of fluent but functionally incorrect output. The work focuses on introspective uncertainty estimation as a way to flag low-confidence generations in software engineering workflows. The abstract is truncated, so the full methods and results are not yet detailed in the listing.
A new arXiv paper introduces ABSOL, a method that pairs aggregated Bayesian subsampling with large language models to improve reasoning over structured data. The approach targets cases where reliable answers depend on consistent evidence, dependency-aware reasoning, and estimated uncertainty. The authors frame the work as addressing the unreliability of LLMs used as natural-language interfaces to Bayesian networks.
A new arXiv paper addresses the risk of overconfident errors when large language models classify clinical text, where a wrong label can affect patient care. The authors note that current black-box approaches simply attach a confidence score to an unchanged LLM prediction, and they propose an uncertainty-aware method designed to produce better-calibrated results. The work targets medical NLP settings where knowing when a model is unreliable matters as much as the predicted label itself.
A new arXiv preprint addresses a limitation in DAREK, a computationally cheap bottom-up scheme for estimating uncertainty in Kolmogorov-Arnold Networks that use spline activations. The authors describe the problem as "fictitious knots" that weaken distance-awareness guarantees in high-dimensional settings, and propose a fix to restore them. The work targets interpretable function approximation, where reliable uncertainty estimates matter for trusting model outputs.
A study examines whether the per-pixel confidence values produced by feed-forward 3D reconstruction models can be treated as reliable uncertainty estimates. Because these scores are trained mainly as loss weights, the authors audit how well they are calibrated for downstream use. The paper reports on the limits of reusing them as an uncertainty signal.
A new arXiv preprint describes a hallucination detection method that combines a fine-tuned DeBERTa-v3 classifier with Monte Carlo dropout uncertainty estimates and other signals. The approach targets domain-specific settings where unfaithful model outputs are a concern. The work is cross-listed under cs.AI.
Researchers describe a machine learning approach that classifies sea ice by its stage of development, using labels drawn from several operational ice charts compiled by human analysts. The method also estimates uncertainty, which is relevant for navigation and ice monitoring where chart interpretations can vary.