Tag
Introduces Kalman Delta Networks, which improve language modeling by reformulating linear attention as a linear-Gaussian state-space model with Kalman-filter updates to track memory uncertainty, yielding efficient approximations that outperform existing linear-attention models.
This paper proposes methods for detecting hallucinations in black-box LLMs by combining semantic entropy and token-level uncertainty signals, evaluating techniques like TopK, CoCoA, Gated, and Stacked across multiple benchmarks to find that no single method is universally strongest but Stacked often performs best.
This paper introduces Disease Continuum Positioning (DCP), a longitudinal Bayesian learning framework that continuously estimates Alzheimer's disease severity from neuroimaging data, providing an uncertainty-aware score for disease progression prediction.
This paper proposes a statistical model to efficiently estimate uncertainty dynamics in text generation, smoothing noisy resampling data to significantly reduce computational costs while maintaining accuracy in analyzing LLM reasoning chains.
This paper studies confidence estimation and selective prediction for financial named entity recognition under domain shift, evaluating BERT and LoRA-tuned Qwen models to enhance reliability across different input distributions like SEC filings and social media.
The paper proposes H2EDL, a hyper evidential deep learning model for hierarchical classification that captures uncertainty at multiple levels of a label hierarchy, reducing calibration error and improving the preservation of coarse categories in fine-grained recognition tasks.
This interpretability study examines overconfidence in large language models, focusing on Qwen3-4B, by analyzing how uncertainty is expressed through verbal markers, abstention, and numeric scores, and proposes methods to identify and mitigate overconfident errors.
The paper proposes CASE, a dynamic selection combiner using a decodability criterion to predict when hidden-state selection outperforms majority voting in large language models, enhancing reliability on difficult questions.
The paper empirically studies self-correction in code generation using uncertainty estimation methods, finding that uncertainty-based approaches fail to improve Pass@1 accuracy, while verification-based methods yield significant gains.
This paper benchmarks auxiliary uncertainty signals from a BERT+GCN classifier to improve LLM-assisted systematic review screening, showing that targeted MAYBE-only routing maximizes efficiency with near-baseline cost.
This paper studies stable miscalibration in large language models, where high-confidence errors remain locally stable under perturbations, using diagnostics like audit scores and probes to assess calibration and internal sensitivity.
This paper introduces Fisher8, an output-layer gradient correction that uses Fisher geometry instead of Euclidean geometry to stabilize neural heteroscedastic regression, improving uncertainty calibration and likelihood-error tradeoffs.
This paper introduces FatigueCV, a physics-informed deep learning framework that predicts steel fatigue life from optical micrographs in under 65ms, using a CNN with uncertainty estimation. Validation on synthetic micrographs shows strong performance (R²=0.93), though real-world validation is noted as future work.
Proposes Collaborative Weighting Actor-Critic (CWAC), a framework that uses distributional critics and a collaborative weighting mechanism to mitigate overestimation bias in off-policy reinforcement learning.
This paper evaluates how small open-weight vision-language models (Qwen2-VL-2B and SmolVLM) handle realistic image degradations, finding that their verbalized confidence is unreliable while internal token probability provides much better error detection, though both fail under severe low-light conditions.
This paper shows that Monte Carlo dropout provides epistemic uncertainty signals for chest radiograph classifiers, which improves error detection and reduces confident misdiagnoses in clinical decision-support agents when communicated as a binary error-risk flag.
This paper introduces an uncertainty-aware trust estimation method for aggregating predictions from multiple LLMs, adapting structured expert judgment with Cooke-style log weighting to penalize overconfident incorrect predictions. Evaluations on MMLU and MMLU-Pro show that this approach achieves superior accuracy-reliability balance under heterogeneous and contaminated expert panels.
This paper investigates whether stochastic sampling (self-consistency) in LLMs can capture cross-question structure similar to diverse ensembles. Using a Marchenko–Pastur test, the authors find that within a single model, stochastic variation yields at most one significant dimension, while an ensemble of 24 models yields four, revealing a dimensionality gap that limits self-consistency as an ensemble substitute.
The article introduces CARE-PPO, a reinforcement learning framework that integrates confidence estimation with PPO fine-tuning for language-based quantitative prediction, enabling models to produce both accurate numerical estimates and reliable confidence signals, demonstrated in healthcare and finance tasks.
This paper introduces diversion decoding, a novel method for detecting hallucinations in large language models by actively challenging model responses during the decoding phase to extract features for training an uncertainty heuristic, achieving superior performance with lower computational complexity.