Online Reasoning Calibration: Test-Time Training Enables Generalizable Conformal LLM Reasoning
Department of Electrical Engineering and Computer Science (MIT EECS)(麻省理工学院电气工程与计算机科学系);Computer Science and Artificial Intelligence Laboratory (MIT CSAIL)(麻省理工学院计算机科学与人工智能实验室);Laboratory for Information and Decision Systems (MIT LIDS)(麻省理工学院信息与决策系统实验室);Computational Science and Engineering (MIT CSE)(麻省理工学院计算科学与工程);Massachusetts Institute of Technology(麻省理工学院)
Cai Zhou, Zekai Wang, Menghua Wu, Qianyu Julie Zhu, Flora C. Shi, Chenyu Wang, Ashia Wilson, Tommi Jaakkola, Stephen Bates
备注:
Published as a conference paper at COLM 2026; 22 pages
While test-time scaling has enabled large language models to solve highly difficult tasks, state-of-the-art results come at exorbitant compute costs. These inefficiencies can be attributed to the miscalibration of post-trained language models, and the lack of calibration in popular sampling techniques. Here, we present Online Reasoning Calibration (ORCA), a framework for calibrating the sampling process that draws upon conformal prediction and test-time training. Specifically, we introduce a meta-learning procedure that updates the calibration module for each input. This allows us to provide valid confidence estimates under distributional shift, e.g. in thought patterns that occur across different stages of reasoning, or in prompt distributions between model development and deployment. ORCA not only provides theoretical guarantees on conformal risks, but also empirically shows higher efficiency and generalization across different reasoning tasks. At risk level $δ=0.1$, ORCA improves Qwen2.5-32B efficiency on in-distribution tasks with savings up to 47.5% with supervised labels and 40.7% with self-consistency labels. Under zero-shot out-of-domain settings, it improves MATH-500 savings from 24.8% of the static calibration baseline to 67.0% while maintaining a low empirical error rate, and the same trend holds across model families and downstream benchmarks. Our code is publicly available at https://github.com/wzekai99/ORCA.
Karlsruhe Institute of Technology(卡尔斯鲁厄理工学院);ETH Zurich(苏黎世联邦理工学院);Tsinghua University(清华大学);HOCH Health Ostschweiz, Kantonsspital St. Gallen(瑞士东部健康组织,圣加仑州立医院)
Multi-Variable Conformal Prediction: Optimizing Prediction Sets using Multiple Calibration Variables
TU Munich & MCML, Germany(慕尼黑技术大学及MCML,德国);Politecnico di Milano, Italy(米兰理工学院,意大利);University of Brescia, Italy(布雷西亚大学,意大利);ETH Zürich, Switzerland(苏黎世联邦理工学院,瑞士)
Efficient Federated Conformal Prediction with Group-Conditional Guarantee
IoT Thrust(物联网 thrust);The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));Institute for Intelligent Networked Systems (INSI)(智能网络化系统研究所);Northeastern University London(伦敦东北大学);Department of ECE The Hong Kong University of Science and Technology HK SAR(电子工程系 香港科技大学香港特别行政区)
Reliable Real-Time Value at Risk Estimation via Quantile Regression Forest with Conformal Calibration
Institute of Statistics and Big Data, Renmin University of China(统计与大数据研究院,中国人民大学);Department of Decision Analytics and Operations, City University of Hong Kong(决策分析与运营部门,香港城市大学);School of Information, Renmin University of China(信息学院,中国人民大学)
Multi-Agent Conformal Prediction with Personalized Statistical Validity
Department of Electronic Systems(电子系统系);Aalborg University(奥尔堡大学);Department of Mathematical Sciences(数学科学系);Institut de Mathématiques de Toulouse(图卢兹数学研究所);Université de Toulouse(图卢兹大学)
Stable Localized Conformal Prediction via Transduction
School of Statistics;Data Science, Nankai University, Tianjin, China;School of Mathematics \& Statistics, The University of Melbourne, Melbourne, Australia
Conformalized Quantile Regression and Minimax Limits of Fixed-Score Calibration under Known Covariate Shift
HSE University(高等经济大学);Lomonosov Moscow State University(罗蒙诺索夫莫斯科国立大学);Laboratoire de Recherche d’EPITA(EPITA研究实验室);University of Duisburg-Essen(杜伊斯堡-埃森大学);Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)(穆罕默德·本·扎耶德人工智能大学(MBZUAI))
LoBoost: Model-Native Local Conformal Prediction for Gradient-Boosted Trees
Federal University of São Carlos(巴西圣卡洛斯联邦大学);University of São Paulo(圣保罗大学);Institute of Mathematics and Computer Science(数学与计算机科学学院);Université Grenoble Alpes(格勒诺布尔阿尔卑斯大学);Inria(法国国家信息与自动化技术研究院);CNRS(法国国家科学研究中心);Grenoble INP(格勒诺布尔国立应用科学研究院);LJK(格勒诺布尔联合实验室)
CoCopula: Conformal Calibration and Copula Aggregation for Reliable Renewable Energy Forecasting
NSF Artificial Intelligence Institute for Advances in Optimization(NSF人工智能优化研究所);H. Milton Stewart School of Industrial and Systems Engineering(H. Milton Stewart工业与系统工程学院);Georgia Institute of Technology(佐治亚理工学院)
Communication-Efficient Byzantine-Robust Federated Conformal Prediction via Partial Sharing
Department of Electronic Systems, Norwegian University of Science and Technology(电子系统系,挪威科学技术大学);Commonwealth Scientific and Industrial Research Organisation(澳大利亚联邦科学与工业研究组织);Department of Information and Communications Engineering, Aalto University(信息与通信工程系,艾尔托大学)
Division of Systems and Control, Department of Information Technology, Uppsala University, Sweden(信息科技系系统与控制 division,乌普萨拉大学,瑞典);Science for Life Laboratory, Uppsala, Sweden(生命科学实验室,乌普萨拉,瑞典)
A Critical Perspective on Finite Sample Conformal Prediction Theory in Medical Applications
Max Planck Institute for Intelligent Systems, Germany(德国马克斯·普朗克智能系统研究所);Tübingen AI Center, Germany(图宾根人工智能中心);ELLIS Institute Tübingen, Germany(图宾根ELLIS研究所);Department of Digital Medicine, University of Bern, Switzerland(伯恩大学数字医学系);Department of Diabetes, Endocrinology, Nutritional Medicine and Metabolism UDEM, Inselspital, Bern University Hospital, University of Bern, Switzerland(伯恩大学糖尿病、内分泌学、营养医学和代谢学系);Diabetes Center Berne, Bern, Switzerland(伯恩糖尿病中心);University of Tübingen, Germany(图宾根大学);University of Lucerne, Switzerland(卢塞恩大学)
Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions
Department of Biostatistics, Vanderbilt University Medical Center(生物统计学系,范德比尔特大学医学中心);Department of Informatics, King’s College London(信息学系,伦敦国王学院);Department of Biostatistics, MD Anderson Cancer Center(生物统计学系,MD安德森癌症中心)
Partial Model Sharing Improves Byzantine Resilience in Federated Conformal Prediction
Department of Electronic Systems, Norwegian University of Science and Technology(挪威科学技术大学电子系统系);CSIRO’s Data61(澳大利亚数据61机构);Department of Information and Communications Engineering, Aalto University(艾尔沃斯大学信息与通信工程系)
Center for Advanced Mathematics in Energy Research Applications(能源研究应用高级数学中心);Synchrotron Infrared Structural Biology Program(同步辐射红外结构生物学计划);Molecular Biophysics and Integrated Bioimaging Division(分子生物物理与整合生物成像部)