Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang, Angelika Romanou, Antoni-Joan Solergibert, Barna Pásztor, Bettina Messmer, Dhia Garbaya, Eduard Frank Ďurech, Ido Hakimi, Juan Garcia Giraldo, Mete Ismayilzada, Negar Foroutan, Skander Moalla, Tiancheng Chen, Vinko Sabolčec, Yixuan Xu, Michael Aerni, Badr AlKhamissi, Inés Altemir Marinas, Mohammad Hossein Amani, Matin Ansaripour, Ilia Badanin, Harold Benoit, Emanuela Boros, Nicholas John Browning, Fabian Bösch, Maximilian Böther, Niklas Canova, Camille Challier, Clément Charmillot, Jonathan Coles, Jan Milan Deriu, Arnout Devos, Lukas Drescher, Daniil Dzenhaliou, Maud Ehrmann, Dongyang Fan, Simin Fan, Silin Gao, Miguel Gila, María Grandury, Diba Hashemi, Alexander Miserlis Hoyle, Jiaming Jiang, Mark Klein, Andrei Kucharavy, Anastasiia Kucherenko, Frederike Lübeck, Roman Machacek, Theofilos Ioannis Manitaras, Andreas Marfurt, Kyle Matoba, Simon Matrenok, Henrique Mendonça, Fawzi Roberto Mohamed, Syrielle Montariol, Luca Mouchel, Sven Najem-Meyer, Jingwei Ni, Gennaro Oliva, Matteo Pagliardini, Elia Palme, Andrei Panferov, Léo Paoletti, Marco Passerini, Ivan Pavlov, Auguste Poiroux, Kaustubh Ponkshe, Nathan Ranchin, Javier Rando, Mathieu Sauser, Jakhongir Saydaliev, Mukhammadali Sayfiddinov, Marian Schneider, Stefano Schuppli, Marco Scialanga, Andrei Semenov, Kumar Shridhar, Raghav Singhal, Anna Sotnikova, Alexander Sternfeld, Ayush Kumar Tarun, Paul Teiletche, Jannis Vamvas, Xiaozhe Yao, Hao Zhao, Alexander Ilic, Ana Klimovic, Andreas Krause, Caglar Gulcehre, David Rosenthal, Elliott Ash, Florian Tramèr, Joost VandeVondele, Livio Veraldi, Martin Rajman, Thomas C. Schulthess, Torsten Hoefler, Antoine Bosselut, Martin Jaggi, Imanol Schlag. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large language models (LLMs) can transform education, but their optimization for direct question-answering often undermines effective pedagogy which requires strategically withholding answers.To mitigate this, we propose an online reinforcement learning (RL)-based alignment framework that can quickly adapt LLMs into effective tutors using simulated student-tutor interactions by emphasizing pedagogical quality and guided problem-solving over simply giving away answers.We use our method to train a 7B parameter tutor model without human annotations which reaches similar performance to larger proprietary models like LearnLM.We introduce a controllable reward weighting to balance pedagogical support and student solving accuracy, allowing us to trace the Pareto frontier between these two objectives.Our models better preserve reasoning capabilities than single-turn SFT baselines and can optionally enhance interpretability through thinking tags that expose the model's instructional planning.

Evaluating the pedagogical capabilities of AIbased tutoring models is critical for making guided progress in the field.Yet, we lack a reliable, easy-to-use, and simple-to-run evaluation that reflects the pedagogical abilities of models.To fill this gap, we present MATH-TUTORBENCH, an open-source benchmark for holistic tutoring model evaluation.MATHTU-TORBENCH contains a collection of datasets and metrics that broadly cover tutor abilities as defined by learning sciences research in dialogbased teaching.To score the pedagogical quality of open-ended teacher responses, we train a reward model and show it can discriminate expert from novice teacher responses with high accuracy.We evaluate a wide set of closed-and open-weight models on MATHTUTORBENCH and find that subject expertise, indicated by solving ability, does not immediately translate to good teaching.Rather, pedagogy and subject expertise appear to form a trade-off that is navigated by the degree of tutoring specialization of the model.Furthermore, tutoring appears to become more challenging in longer dialogs, where simpler questioning strategies begin to fail.We release the benchmark, code, and leaderboard openly to enable rapid benchmarking of future models. 1 github.com/eth-lre/mathtutorbench

Recent efforts in fine-tuning language models often rely on automatic data selection, commonly using Nearest Neighbors retrieval from large datasets. However, we theoretically show that this approach tends to select redundant data, limiting its effectiveness or even hurting performance. To address this, we introduce SIFT, a data selection algorithm designed to reduce uncertainty about the model's response given a prompt, which unifies ideas from retrieval and active learning. Whereas Nearest Neighbor retrieval typically fails in the presence of information duplication, SIFT accounts for information duplication and optimizes the overall information gain of the selected examples. We focus our evaluations on fine-tuning at test-time for prompt-specific language modeling on the Pile dataset, and show that SIFT consistently outperforms Nearest Neighbor retrieval, with minimal computational overhead. Moreover, we show that our uncertainty estimates can predict the performance gain of test-time fine-tuning, and use this to develop an adaptive algorithm that invests test-time compute proportional to realized performance gains. We provide the `activeft` (Active Fine-Tuning) library which can be used as a drop-in replacement for Nearest Neighbor retrieval.

One of the exciting capabilities of recent language models for dialog is their ability to independently search for relevant information to ground a given dialog response.However, obtaining training data to teach models how to issue search queries is time and resource consuming.In this work, we propose q2d: an automatic data generation pipeline that generates information-seeking dialogs from questions.We prompt a large language model (PaLM) to create conversational versions of question answering datasets, and use it to improve query generation models that communicate with external search APIs to ground dialog responses.Unlike previous approaches which relied on human written dialogs with search queries, our method allows to automatically generate querybased grounded dialogs with better control and scale.Our experiments demonstrate that: (1) For query generation on the QReCC dataset, models trained on our synthetically-generated data achieve 90%-97% of the performance of models trained on the human-generated data;(2) We can successfully generate data for training dialog models in new domains without any existing dialog data as demonstrated on the multi-hop MuSiQue and Bamboogle QA datasets.(3) We perform a thorough analysis of the generated dialogs showing that humans find them of high quality and struggle to distinguish them from human-written dialogs. 1 * Work done during an internship at Google Research. 1 We publicly release all prompts and will release all generated datasets.

As the field of deep learning progresses, and neural networks become larger, training them has become a demanding and time consuming task. To tackle this problem, distributed deep learning must be used to scale the training of deep neural networks to many workers. Synchronous algorithms, commonly used for distributing the training, are susceptible to faulty or straggling workers. Asynchronous algorithms do not suffer from the problems of synchronization, but introduce a new problem known as staleness. Staleness is caused by applying out-of-date gradients, and it can greatly hinder the convergence process. Furthermore, asynchronous algorithms that incorporate momentum often require keeping a separate momentum buffer for each worker, which cost additional memory proportional to the number of workers. We introduce a new asynchronous method, SMEGA2, which requires a single momentum buffer regardless of the number of workers. Our method works in a way that lets us estimate the future position of the parameters, thereby minimizing the staleness effect. We evaluate our method on the CIFAR and ImageNet datasets, and show that SMEGA2 outperforms existing methods in terms of final test accuracy while scaling up to as much as 64 asynchronous workers. Open-Source Code: https://github.com/rafi-cohen/SMEGA2

Training neural networks on large distributed clusters has become a common practice due to the size and complexity of recent neural networks. These high-end clusters of advanced computational devices cooperate together to reduce the neural network training duration. In practice, training at linear scalability with respect to the number of devices is difficult, due to communication overheads. These communication overheads often cause long idle times for the computational devices. In this paper, we propose LAGA (Lagged AllReduce with Gradient Accumulation): a hybrid technique that combines the best of synchronous and asynchronous approaches, that scales linearly. LAGA reduces the device idle time by accumulating locally computed gradients and executing the communications in the background. We demonstrate the effectiveness of LAGA in both final accuracy and scalability on the ImageNet dataset, where LAGA achieves a speedup of up to 2. 96x and 5. 24x less idle time. Finally, we provide convergence guarantees for LAGA under the non-convex setting.

We consider stochastic convex optimization problems, where several machines act asynchronously in parallel while sharing a common memory. We propose a robust training method for the constrained setting and derive non asymptotic convergence guarantees that do not depend on prior knowledge of update delays, objective smoothness, and gradient variance. Conversely, existing methods for this setting crucially rely on this prior knowledge, which render them unsuitable for essentially all shared-resources computational environments, such as clouds and data centers. Concretely, existing approaches are unable to accommodate changes in the delays which result from dynamic allocation of the machines, while our method implicitly adapts to such changes.

We propose a novel technique for faster deep neural network training which systematically applies sample-based approximation to the constituent tensor operations, i.e., matrix multiplications and convolutions. We introduce new sampling techniques, study their theoretical properties, and prove that they provide the same convergence guarantees when applied to SGD training. We apply approximate tensor operations to single and multi-node training of MLP and CNN networks on MNIST, CIFAR-10 and ImageNet datasets. We demonstrate up to 66% reduction in the amount of computations and communication, and up to 1.37x faster training time while maintaining negligible or no impact on the final test accuracy.