Tag
This paper benchmarks LLMs against human verbal uncertainty markers from psychology literature and introduces VOCAL, an optimization-based algorithm that learns an optimal marker-to-uncertainty profile directly from LLM outputs, revealing systematic gaps between how LLMs and humans express confidence in language.
The post benchmarks a zero-GPU, CPU-only string-normalization entropy metric (Spanda) against semantic entropy across 1.5B–120B local and frontier models, finding that 7B–27B models are well-served by cheap exact-match entropy while 120B models collapse into identical confident hallucinations that even NLI models miss.
A new paper introduces CAGE, a framework combining generative models, adversarial discrimination, and conformal prediction to weight time series ensemble forecasts, showing improved robustness against outliers and noise on milk collection and monkeypox datasets.
The paper proposes EDAR, an entropy-driven adaptive routing framework that decides at inference time whether to answer a query with retrieval-augmented generation or escalate it to full long-context processing. EDAR retains 97.4% of the pure long-context baseline's accuracy while cutting token cost by 70.7%, escalating only 18.2% of queries on LongBench v2 and Infinity-Bench.
本文提出一种配对预测协议，用于衡量科学解释对实验预测的额外价值，并在 Tox21 与 OpenML 基准上测试；结论是自然解释的预测信用在测试分辨率下仍未得到确认，而 DeepSeek V4 系列模型的解释卡可降低部分次级漂移。
The article explains how to address hallucination in large language models by using probability distributions and calibration to quantify uncertainty, rather than unreliable confidence scores.
This paper derives sharp theoretical limits for honest uncertainty in repeated evaluation under a hard budget, with applications to language model and agent benchmarking, demonstrating practical improvements in interval width and MSE.
This paper introduces TAM-Chain, a multi-scale thyroid cytology classification method that employs absorbing Markov chains and Shannon entropy for uncertainty quantification, aiming to suppress false negatives and adapt to domain shifts.
The paper proposes the Slicing-Graphing-Alignment (SGA) method to quantify uncertainty in multi-step forecasting for time series foundation models, demonstrating superior performance and revealing an empirical scaling law.
This paper proposes a full-covariance smoothing technique for Bayesian neural networks to enable efficient online adaptation by propagating correlations through nonlinear activations, demonstrated in tasks like classification and control.
This paper proposes a tail-aware geometry learning framework for conformal ellipsoids that decouples tail sensitivity from coverage guarantees, improving uncertainty quantification in multivariate settings.
This paper combines transfer learning with conformalized quantile regression to improve solar PV forecasting accuracy and uncertainty quantification under data scarcity caused by load-shedding in Bangladesh, demonstrating significant performance gains over baseline methods.
The paper investigates uncertainty quantification methods for neural surrogates in computer-aided engineering, comparing Gaussian processes, Monte Carlo dropout, and deep ensembles on industrial datasets to evaluate their performance in various engineering contexts.
This paper develops a stochastic lexical calculus framework for semantic updates in language models, defining conditions for when language-derived probabilities support meaningful sequential state representations. Empirical experiments validate the framework's stability and coverage under calibrated conditions.
This paper introduces Compressed Active Subspaces (CAS), a method to enable scalable Bayesian inference in large neural networks by compressing parameter spaces and maintaining predictive performance.
The paper introduces pFedHGP, a personalized federated learning approach using hierarchical Gaussian processes for probabilistic modeling of heterogeneous distributed systems while preserving privacy through federated variational inference.
This paper proposes a variational Bayesian last layer method for graph neural networks to perform online node classification under distribution shift, achieving superior accuracy and uncertainty calibration compared to baselines.
This paper proposes the Label-Shift-Adjusted Bayesian Score for conformal prediction under label shift, demonstrating shorter prediction intervals with maintained coverage in molecular property prediction.
The paper introduces Physics-Informed Conformal Prediction (PI-CP), a framework that embeds PDE residuals into conformal prediction to provide distribution-free, spatially adaptive uncertainty quantification for neural operators like the Fourier Neural Operator (FNO).
The paper proposes a surrogate model approach to efficiently estimate inconsistency response surfaces in Cyber-Physical Systems under uncertainty, enabling scalability for sensitivity analysis and consistency repair.