Publications

VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization
Andrei Atanov
Jesse Allardice
Roman Bachmann
Oğuzhan Kar
R Devon Hjelm
David Griffiths
Peter Fu
Amir Zamir
Afshin Dehghan
Visual tokenizers map high-dimensional raw pixels into a compressed representation for downstream modeling, e.g., conditional video generati… (see more)on. Beyond compression, tokenizers define what information is preserved and how it is organized. A _de facto_ standard approach is to represent a video with a spatiotemporal 3D grid of tokens, each corresponding to a local patch in the original signal. This requires a downstream model, e.g., a text-to-video model, to learn to predict all low-level details ``pixel-by-pixel'' irrespective of the video's inherent complexity, resulting in high computational cost during training. We present VideoFlexTok, a tokenizer that represents videos with a _variable-length sequence of tokens structured in a coarse-to-fine manner_, where the first tokens capture abstract information like semantics and motion and later tokens provide fine-grained details. The generative flow decoder enables realistic video reconstructions from any token count. This representation structure allows adapting the tokens count to particular downstream needs and encode videos longer than the 3D grid approach under the same budget. We evaluate VideoFlexTok on class-to-video and text-to-video generative tasks and show that it leads to more efficient training compared to 3D grid tokens, e.g., achieving comparable generation quality (gFVD and ViCLIP Score) with a 10x smaller model (0.4B vs 3.6B). Finally, we demonstrate how VideoFlexTok can enable long video generation without prohibitive computational cost by training a text-to-video model on 10-second 81-frame videos with only 672 tokens, 8x fewer than a comparable 3D grid tokenizer.
2025-12-31
International Conference on Machine Learning (Accept (spotlight))
doi.org
openreview.net
“I Was a Young AI”: On Probing the Effectiveness of Intervening on Anthropomorphic AI System Outputs
Su Lin Blodgett
Myra Cheng
Alexandra Olteanu
We see growing concerns about how the increasingly pervasive deployment of AI systems whose outputs appear human-like might impact people.Th… (see more)ese concerns have already motivated work both examining what makes such outputs appear human-like, as well as developing interventions to help reduce perceptions of human-likeness or mitigate adverse impacts.In this paper, we report on an exploratory crowd study we designed to examine challenges for assessing the effectiveness of interventions, including whether interventions intended to minimize perceptions of human-likeness also mitigate adverse impacts.We find variations both in what kinds of outputs different participants deem more human-like, as well as in their preferences for human-like outputs.Even when participants seem to prefer the outputs they deem more human-like, many of them also recognize that such outputs can have adverse impacts.Drawing on these results and prior work, we discuss challenges to and considerations for assessing the effectiveness of interventions.
2025-12-31
Beyond Alignment: Transdisciplinary Conversations on Human-AI Futures @ Neural Information Processing Systems (published)
doi.org
When does predictive inverse dynamics outperform behavior cloning?
Lukas Schäfer
Pallavi Choudhury
Abdelhak Lemkhenter
Chris Lovett
Somjit Nath
Luis França
Matheus de Mendonça
Alex Lamb
Riashat Islam
Siddhartha Sen
John Langford
Katja Hofmann
Sergio Valcarcel Macua
Behavior cloning (BC) is a practical offline imitation learning method, but it often fails when expert demonstrations are limited. Recent wo… (see more)rks have introduced a class of architectures named predictive inverse dynamics models (PIDM) that combine a future state predictor with an inverse dynamics model (IDM). While PIDM often outperforms BC, the reasons behind its benefits remain unclear. In this paper, we provide a theoretical explanation: PIDM introduces a bias-variance tradeoff. While predicting the future state introduces bias, conditioning the IDM on the prediction can significantly reduce variance. We establish conditions on the state predictor bias for PIDM to achieve lower prediction error and higher sample efficiency than BC, with the gap widening when additional data sources are available. We validate the theoretical insights empirically in 2D navigation tasks, where BC requires up to five times (three times on average) more demonstrations than PIDM to reach comparable performance; and in a complex 3D environment in a modern video game with high-dimensional visual inputs and stochastic transitions, where BC requires over 66\% more samples than PIDM.
2025-12-31
International Conference on Machine Learning (Accept (regular))
openreview.net
openreview.net
When Greedy Wins: Emergent Exploitation Bias in Meta-Bandit LLM Training
Sanxing Chen
Xiaoyin Chen
Yukun Huang
Roy Xie
Bhuwan Dhingra
While Large Language Models (LLMs) hold promise to become autonomous agents, they often explore suboptimally in sequential decision-making. … (see more)Recent work has sought to enhance this capability via supervised fine-tuning (SFT) or reinforcement learning (RL), improving regret on the classic multi-armed bandit task. However, it remains unclear how these learning methods shape exploration strategies and how well they generalize. We investigate both paradigms by training LLMs with SFT on expert trajectories and RL with a range of tailored reward signals including a strategic, regret-shaped reward to reduce variance, and an algorithmic reward that enables oracle imitation. The resulting agents outperform pre-trained models and achieve performance comparable to Upper Confidence Bound (UCB) and Thompson Sampling, with robust generalization to 6
2025-12-31
International Conference on Learning Representations (Accept (Poster))
doi.org
openreview.net
Zero-Shot NAS for TinyML Semantic Segmentation via Weight Sharing
Zhuoran Xiong
Warren J. Gross
Brett Meyer
2025-12-31
IEEE Embedded Systems Letters (published)
doi.org
C3PO: Optimized Large Language Model Cascades with Probabilistic Cost Constraints for Reasoning
Antonios Valkanas
Soumyasundar Pal
Pavel Rumiantsev
Yingxue Zhang
Mark J. Coates
2025-12-30
OpenReview (published)
doi.org
openreview.net
Family caregivers' acceptance of Artificial Intelligence-enabled technologies for providing care to older adults
Amanda Yee
Mark J. Yaffe
Tibor Schuster
Sylvie Lambert
Samira Abbasgholizadeh-Rahimi
Artificial intelligence (AI)-enabled technologies hold promise for assisting in the care of an aging population. Few studies have focused on… (see more) exploring family caregivers’ (FCGs) behavioural intention of using such innovation, and even fewer have employed a technology acceptance framework. This study examined FCGs of older adults’ behavioural intention of using AI-enabled technologies for caregiving. We conducted a theory-based cross-sectional quantitative survey. Eligible FCGs for this study were: (1) aged 45–64; (2) residing in Quebec, Canada; (3) providing care for at least one older adult (65+); (4) having access to a computer or smartphone with internet connectivity; and, (5) having proficiency in reading and comprehending English or French. We adapted and expanded the Unified Theory of Acceptance and Use of Technology (UTAUT) framework to measure their behavioural intention of using AI-enabled technologies for caregiving. We used descriptive statistics and a random forest model to assess the most important predictive factors across nine variables and their direction of association with behavioural intention. The Consensus-Based Checklist for Reporting of Survey Studies (CROSS) guidelines was used for reporting the study’s results. Among the polling firm’s 100,000 panelists, 2740 eligible individuals were randomly chosen to receive an email invitation to the study. Of 465 panelists who opened the survey (i.e., unique visitors),199 were eligible and completed the online survey. The random forest model explained between 56% and 86% of the behavioural intention variance of using AI, with social influence demonstrating the highest predictive relevance as indicated by a 35% increase in mean-squared error once removed from the model. Among the nine variables considered, six demonstrated a positive association with behavioural intention. These variables included social influence, effort expectancy, performance expectancy, perceived trust, confidence in healthcare professionals’ advice for the use of AI-enabled technologies, and facilitating connditions. The variables perceived cost and technology anxiety indicated a negative association with behavioural intention. Our extended UTAUT model identified factors associated with FCGs' intention to use AI. While all nine variables contributed, attitudes toward AI within caregivers’ social circles was the strongest predictor. Stakeholders from industry, government, and healthcare can enhance the adoption of AI-enabled technologies in older adult care by leveraging facilitators and addressing barriers experienced by caregivers.
2025-12-30
BMC Geriatrics (published)
doi.org
Learning Task-Agnostic Representations through Multi-Teacher Distillation
Philippe Formont
Maxime Darrin
Banafsheh Karimian
Jackie CK Cheung
Éric Granger
Ismail Ben Ayed
Mohammadhadi Shateri
Pablo Piantanida
Casting complex inputs into tractable representations is a critical step across various fields. Diverse embedding models emerge from differe… (see more)nces in architectures, loss functions, input modalities and datasets, each capturing unique aspects of the input. Multi-teacher distillation leverages this diversity to enrich representations but often remains tailored to specific tasks. In this paper, we introduce a task-agnostic framework based on a ``majority vote" objective function. We demonstrate that this function is bounded by the mutual information between student and teachers' embeddings, leading to a task-agnostic distillation loss that eliminates dependence on task-specific labels or prior knowledge. Our evaluations across text, vision models, and molecular modeling show that our method effectively leverages teacher diversity, resulting in representations enabling better performance for a wide range of downstream tasks such as classification, clustering, or regression. Additionally, we train and release state-of-the-art embedding models, enhancing downstream performance in various modalities.
2025-12-30
OpenReview (published)
doi.org
openreview.net
SKOLR: Structured Koopman Operator Linear RNN for Time-Series Forecasting
Yitian Zhang
Liheng Ma
Antonios Valkanas
Boris N. Oreshkin
Mark Coates
Koopman operator theory provides a framework for nonlinear dynamical system analysis and time-series forecasting by mapping dynamics to a sp… (see more)ace of real-valued measurement functions, enabling a linear operator representation. Despite the advantage of linearity, the operator is generally infinite-dimensional. Therefore, the objective is to learn measurement functions that yield a tractable finite-dimensional Koopman operator approximation. In this work, we establish a connection between Koopman operator approximation and linear Recurrent Neural Networks (RNNs), which have recently demonstrated remarkable success in sequence modeling. We show that by considering an extended state consisting of lagged observations, we can establish an equivalence between a structured Koopman operator and linear RNN updates. Building on this connection, we present SKOLR, which integrates a learnable spectral decomposition of the input signal with a multilayer perceptron (MLP) as the measurement functions and implements a structured Koopman operator via a highly parallel linear RNN stack. Numerical experiments on various forecasting benchmarks and dynamical systems show that this streamlined, Koopman-theory-based design delivers exceptional performance.
2025-12-30
OpenReview (published)
doi.org
proceedings.mlr.press
Sparse Decomposition of Graph Neural Networks
Yaochen Hu
Mai Zeng
Ge Zhang
Pavel Rumiantsev
Liheng Ma
Yingxue Zhang
Mark Coates
2025-12-30
OpenReview (published)
doi.org
openreview.net
Combining Constraint Programming and Machine Learning: From Current Progress to Future Opportunities
Quentin Cappart
Tias Guns
Michele Lombardi
Gilles Pesant
Dimos Tsouros
The integration of constraint programming (CP) together with machine learning (ML) has emerged as a promising direction for tackling complex… (see more) decision-making and combinatorial optimization problems. While CP offers expressive modeling capabilities and formal guarantees, ML provides adaptive methods for learning from data and generalizing across instances. This survey presents a comprehensive overview of recent advances in combining CP and ML. We first show how ML has been used to improve the CP toolbox, both in modeling and in the efficiency of solving. Then, we examine how CP can support ML, particularly in providing structure, guarantees, and symbolic reasoning capabilities. Finally, we identify key open challenges inherent to such hybrid approaches and outline promising directions for future research. This survey provides a first conceptual and structured review of recent advancements in this emerging field, aiming to serve as a resource for practitioners and researchers in both the CP and ML communities. To keep the progress up to date, a curated list of references is hosted on an accompanying repository (https://github.com/corail-research/CPML-paper-list) and is open to community contributions.
2025-12-29
Journal of Artificial Intelligence Research (published)
doi.org
LeCov: Multi-level testing criteria for large language models
Xuan Xie
Jiayang Song
Yuheng Huang
Da Song
Felix Juefei-Xu
Lei Ma
2025-12-29
Journal of Systems and Software (published)
doi.org