We present an empirical study examining how experienced tutors (experts) and non-tutors (novices) evaluate the correctness of tutor praise responses under different AI-assisted decision-support interfaces and explanation styles. We examine human-AI reliance patterns by decomposing interaction errors into over-reliance (accepting incorrect AI suggestions) and under-reliance (rejecting correct AI suggestions), together with time cost as a process-level indicator. Across conditions, human-AI collaboration improved accuracy compared to humans working alone, but consistently underperformed an AI-only baseline, indicating that human judgment introduced additional errors even when assisted by a highly accurate model. Novices benefited more from AI support since they tend to follow AI suggestions, whereas experts frequently overrode correct AI advice, resulting in lower overall performance, revealing a paradox of expertise in educational decision-making. We further compare two explanation modalities: textual reasoning and inline highlighting. Textual reasoning reduced under-reliance when the AI was correct but increased over-reliance when the AI was wrong, while inline highlighting exerted minimal influence on either behavior. Notably, neither explanation modality improved accuracy, and both increased time costs. As a contribution to learning analytics, we demonstrate how reliance patterns (over-reliance and under-reliance) and time cost function as process-level indicators that reveal how users integrate, or fail to integrate, AI recommendations. Our findings underscore the need for adaptive, trust-calibrated explanation strategies in tutor-facing decision support systems that balance accuracy, efficiency, and accountability in human-AI collaboration.

Jeffrey Li, Mohammadreza Armandpour, Seyed Iman Mirzadeh, Sachin Mehta, Vaishaal Shankar, Raviteja Vemulapalli, Samy Bengio, Oncel Tuzel, Mehrdad Farajtabar, Hadi Pouransari, Fartash Faghri. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Scaling laws are useful guides for derisking expensive training runs, as they predict performance of large models using cheaper, small-scale experiments. However, there remain gaps between current scaling studies and how language models are ultimately trained and evaluated. For instance, scaling is usually studied in the compute-optimal training regime (i.e., "Chinchilla optimal" regime). In contrast, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but models are usually compared on downstream task performance. To address both shortcomings, we create a testbed of 104 models with 0.011B to 6.9B parameters trained with various numbers of tokens on three data distributions. First, we fit scaling laws that extrapolate in both the amount of over-training and the number of model parameters. This enables us to predict the validation loss of a 1.4B parameter, 900B token run (i.e., 32$\times$ over-trained) and a 6.9B parameter, 138B token run (i.e., a compute-optimal run)––each from experiments that take 300$\times$ less compute. Second, we relate the perplexity of a language model to its downstream task performance by proposing a power law. We use this law to predict top-1 error averaged over downstream tasks for the two aforementioned models, using experiments that take 20$\times$ less compute.

We introduce DataComp for Language Models, a testbed for controlled dataset experiments with the goal of improving language models.As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations.Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing atmodel scales ranging from 412M to 7B parameters.As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set.The resulting dataset, DCLM-Baseline, enables training a 7B parameter language model from scratch to 63% 5-shot accuracy on MMLU with 2T training tokens.Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6 percentage point improvement on MMLU while being trained with half the compute.Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation. We release the \dclm benchmark, framework, models, and datasets at https://www.datacomp.ai/dclm/

Weak supervision (WS) is a popular approach for label-efficient learning, leveraging diverse sources of noisy but inexpensive weak labels to automatically annotate training data. Despite its wide usage, WS and its practical value are challenging to benchmark due to the many knobs in its setup, including: data sources, labeling functions (LFs), aggregation techniques (called label models), and end model pipelines. Existing evaluation suites tend to be limited, focusing on particular components or specialized use cases. Moreover, they often involve simplistic benchmark tasks or de-facto LF sets that are suboptimally written, producing insights that may not generalize to real-world settings. We address these limitations by introducing a new benchmark, BOXWRENCH, designed to more accurately reflect real-world usages of WS. This benchmark features tasks with (1) higher class cardinality and imbalance, (2) notable domain expertise requirements, and (3) opportunities to re-use LFs across parallel multilingual corpora. For all tasks, LFs are written using a careful procedure aimed at mimicking real-world settings. In contrast to existing WS benchmarks, we show that supervised learning requires substantial amounts (1000+) of labeled examples to match WS in many settings.

OBJECTIVES: Distributed computations facilitate multi-institutional data analysis while avoiding the costs and complexity of data pooling. Existing approaches lack crucial features, such as built-in medical standards and terminologies, no-code data visualizations, explicit disclosure control mechanisms, and support for basic statistical computations, in addition to gradient-based optimization capabilities. MATERIALS AND METHODS: We describe the development of the Collaborative Data Analysis (CODA) platform, and the design choices undertaken to address the key needs identified during our survey of stakeholders. We use a public dataset (MIMIC-IV) to demonstrate end-to-end multi-modal FL using CODA. We assessed the technical feasibility of deploying the CODA platform at 9 hospitals in Canada, describe implementation challenges, and evaluate its scalability on large patient populations. RESULTS: The CODA platform was designed, developed, and deployed between January 2020 and January 2023. Software code, documentation, and technical documents were released under an open-source license. Multi-modal federated averaging is illustrated using the MIMIC-IV and MIMIC-CXR datasets. To date, 8 out of the 9 participating sites have successfully deployed the platform, with a total enrolment of >1M patients. Mapping data from legacy systems to FHIR was the biggest barrier to implementation. DISCUSSION AND CONCLUSION: The CODA platform was developed and successfully deployed in a public healthcare setting in Canada, with heterogeneous information technology systems and capabilities. Ongoing efforts will use the platform to develop and prospectively validate models for risk assessment, proactive monitoring, and resource usage. Further work will also make tools available to facilitate migration from legacy formats to FHIR and DICOM.

Labeling training data is a critical and expensive step in producing high accuracy ML models, whether training from scratch or fine-tuning. To make labeling more efficient, two major approaches are programmatic weak supervision (WS) and semi-supervised learning (SSL). More recent works have either explicitly or implicitly used techniques at their intersection, but in various complex and ad hoc ways. In this work, we define a simple, modular design space to study the use of SSL techniques for WS more systematically. Surprisingly, we find that fairly simple methods from our design space match the performance of more complex state-of-the-art methods, averaging a 3 p.p. increase in accuracy/F1-score across 8 standard WS benchmarks. Further, we provide practical guidance on when different components are worth their added complexity and training costs. Contrary to current understanding, we find using SSL is not necessary to obtain the best performance on most WS benchmarks but is more effective when: (1) end models are smaller, and (2) WS provides labels for only a small portion of training examples.

Due to the popularity of social networks and human-carried/human-affiliated devices with sensing abilities, like smartphones and smart wearable devices, a novel application was necessitated recently to organize group activities by learning historical data gathered from smart devices and choosing invitees carefully based on their personal interests. We proposed a private and efficient social activity invitation framework. Our main contributions are ( 1 ) defining a novel friendship to reduce the communication/update cost within the social network and enhance the privacy guarantee at the same time; ( 2 ) designing a strong privacy-preserving algorithm for graph publication, which addresses an open concern proposed recently; ( 3 ) presenting an efficient invitee-selection algorithm, which outperforms the existing ones. Our simulation results show that the proposed framework has good performance. In our framework, the server is assumed to be untrustworthy but can nonetheless help users organize group activities intelligently and efficiently. Moreover, the new definition of the friendship allows the social network to be described by a directed graph. To the best of our knowledge, it is the first work to publish a directed graph in a differentially private manner with an untrustworthy server.