Tag
The paper introduces EDB-Bio, a 96k-episode benchmark for scientific evidence deliberation across three biological domains, showing that frontier and open-weight models are systematically underconfidence-prone (Brier scores near a trivial base-rate forecaster) and that post-training with proper scoring rules improves calibration, yielding an Ibis0 model 2.5× more efficient than the best frontier model in sequential experiment selection.
提出一种面向救护车行程时间估计的不确定性感知方法，通过生成上下文自适应的预测区间来支持更透明、更风险敏感的紧急医疗调度决策，并在瑞典Skåne地区两年的真实救护数据上进行了回顾性评估。
This paper identifies and formalizes 'inference-time model shift'—a change in conformal score distributions caused by varying inference configurations like quantization and temperature rather than data distribution changes—showing it breaks split-conformal exchangeability and overstates deployment-time coverage across 8 language models and 4 finite-label benchmarks.
A design discussion asking when an agent should combine readings from two read-only tools that lack a shared snapshot, and what response contract should govern qualified availability estimates versus reporting readings separately while preserving uncertainty.
该论文提出"机制不确定性"概念，通过因果稳健性边际（causal robustness margin）量化大语言模型内部机制的稳定性，在多个开源模型与问答数据集上验证其对幻觉与不确定性检测的补充作用。
本文提出结合蒙特卡洛 dropout 不确定性量化的条件生成对抗网络数据增强框架，用于 Hawk T1A 飞机结构健康监测的合成信号生成与损伤分类，并与实验测量数据进行多维度比较验证。
论文提出 L-SplineFNO，一种将 FNO 主干与稀疏可学习张量积 B 样条传输耦合的内存高效神经算子，用于在非结构化网格上近似 PDE 解算子，在六个三维工程基准上实现高精度并可扩展至 4600 万节点，还支持零样本超分辨率和基于 Bernstein 凸包与 Taylor 模型的严格预测界。
提出 Uncertainty-aware Rubric Weighting (URW) 框架，通过 conformal 校准量化 LLM 评估中各 rubric 标准的判断不确定性，并加权聚合为奖励信号，显著提升成对偏好判断准确率及下游 RL 对齐效果。
提出 Selection-Safe Conformal Early Exit 方法，通过终端路径分数与单一 split-conformal 分位数校准，在数据依赖的自适应深度选择下仍保持无分布覆盖保证，无需 α/L 修正，并在 CIFAR-100/ImageNet 等任务上验证了路径选择相比朴素逐层选择的覆盖优势。
该论文提出了一种名为 Conformal Adaptation 的框架，通过共形风险控制（conformal risk control）在低分辨率下认证预测信号，再对已认证区域进行精细化计算，从而在 3D 医学图像分割、星系流量预测和钙信号预测等任务上实现 2.5–72 倍的墙钟加速，同时保留不确定性估计。
该论文提出了多层粒子边缘Metropolis-Hastings（MLPMMH）方法，通过伪边缘推断与跨网络宽度截断层级的耦合SMC相结合，实现分层贝叶斯深度学习中参数与超参数的联合推断，并在理论上证明了O(ε²)的均方误差与O(ε⁻²)的计算复杂度，同时给出可扩展的近似版本S-MLPMMH用于MNIST实验。
This paper introduces mechanistic uncertainty in LLMs, defined via a causal robustness margin measured using concept steering vectors, and shows that incorrect responses are more sensitive to internal interventions while learned facts gain robustness.
This paper proposes robust, adaptive Langevin Monte Carlo methods (R-LMC, A-LMC, RA-LMC) that handle generalized smoothness and state-dependent heavy-tailed noise, establishing Fisher information convergence guarantees without dissipativity assumptions, and demonstrates effective Bayesian fine-tuning of LLMs and vision-language models.
该论文提出U-Space，一种低维子空间方法，通过将“怀疑”与“确定”语义锚点映射到残差空间，实现对语言模型逐token的可解释不确定性量化，无需标签、重复采样或额外训练，并在推理基准上优于现有基线。
这篇论文通过分层诊断研究探讨了交通预测精度提升为何未能改善信号控制决策，发现预测价值取决于时间可观测性、动作可识别性、动力学一致性和目标对齐，并提出了一套实用的诊断协议。
U-Space introduces a low-dimensional subspace for making and interpreting uncertainty in language models, using semantic anchors of doubt and certainty mapped into residual space. The training-free U-Lens produces token-level uncertainty maps that outperform baselines on reasoning benchmarks, even under length-controlled evaluation.
The paper introduces Threshold-Aware Conformal Routing (TACR), a method that learns input-dependent interval scales to decide when a neural surrogate can replace an expensive scientific simulator, reducing simulator deferrals by 14–75% while preserving distribution-free marginal coverage guarantees.
OmniMed-Jev reframes medical LVLM outputs as distributions over a runtime-supplied candidate set, cutting calibration error by up to an order of magnitude and reliability error by two versus a generative baseline while keeping point-prediction performance comparable (except counting).
This paper studies how much normal evidence (images or categories) is needed to certify a false-alarm-rate-controlled threshold for few-shot industrial anomaly detection. Using a frozen DINOv2 PCA residual ranker on MVTec and VisA, the authors show target-only calibration is shift-fragile, derive lower bounds on the number of iid categories needed for valid distribution-free 95% upper confidence bounds (14–59 draws), and propose the CRESS protocol that largely fails closed with only a few certification categories.
The paper proposes a framework that preserves individual expert interval labels by modeling them with Beta distributions and decomposing predictive uncertainty into within-, between-, and model components, using a decomposition matching technique to align them. On sea-ice concentration data, it reduces MAE by 31% over hard-label baselines and improves correlation with between-expert variation to 0.500.