Alexandre Lacoste

Gaurav Sahu, Abhay Puri, Juan Rodriguez et al. · mila

Data analytics is essential for extracting valuable insights from data that can assist organizations in making effective decisions. We introduce InsightBench, a benchmark dataset with three key features. First, it consists of 100 datasets representing diverse business use cases such as finance and incident management, each accompanied by a carefully curated set of insights planted in the datasets. Second, unlike existing benchmarks focusing on answering single queries, InsightBench evaluates agents based on their ability to perform end-to-end data analytics, including formulating questions, interpreting answers, and generating a summary of insights and actionable steps. Third, we conducted comprehensive quality assurance to ensure that each dataset in the benchmark had clear goals and included relevant and meaningful questions and analysis. Furthermore, we implement a two-way evaluation mechanism using LLaMA-3 as an effective, open-source evaluator to assess agents' ability to extract insights. We also propose AgentPoirot, our baseline data analysis agent capable of performing end-to-end data analytics. Our evaluation on InsightBench shows that AgentPoirot outperforms existing approaches (such as Pandas Agent) that focus on resolving single queries. We also compare the performance of open- and closed-source LLMs and various evaluation strategies. Overall, this benchmark serves as a testbed to motivate further development in comprehensive automated data analytics and can be accessed here: https://github.com/ServiceNow/insight-bench.

Léo Boisvert, Megh Thakkar, Maxime Gasse et al.

The ability of large language models (LLMs) to mimic human-like intelligence has led to a surge in LLM-based autonomous agents. Though recent LLMs seem capable of planning and reasoning given user instructions, their effectiveness in applying these capabilities for autonomous task solving remains underexplored. This is especially true in enterprise settings, where automated agents hold the promise of a high impact. To fill this gap, we propose WorkArena++, a novel benchmark consisting of 682 tasks corresponding to realistic workflows routinely performed by knowledge workers. WorkArena++ is designed to evaluate the planning, problem-solving, logical/arithmetic reasoning, retrieval, and contextual understanding abilities of web agents. Our empirical studies across state-of-the-art LLMs and vision-language models (VLMs), as well as human workers, reveal several challenges for such models to serve as useful assistants in the workplace. In addition to the benchmark, we provide a mechanism to effortlessly generate thousands of ground-truth observation/action traces, which can be used for fine-tuning existing models. Overall, we expect this work to serve as a useful resource to help the community progress toward capable autonomous agents. The benchmark can be found at https://github.com/ServiceNow/WorkArena.

Alexandre Lacoste, Nils Lehmann, Pau Rodriguez et al.

Recent progress in self-supervision has shown that pre-training large neural networks on vast amounts of unsupervised data can lead to substantial increases in generalization to downstream tasks. Such models, recently coined foundation models, have been transformational to the field of natural language processing. Variants have also been proposed for image data, but their applicability to remote sensing tasks is limited. To stimulate the development of foundation models for Earth monitoring, we propose a benchmark comprised of six classification and six segmentation tasks, which were carefully curated and adapted to be both relevant to the field and well-suited for model evaluation. We accompany this benchmark with a robust methodology for evaluating models and reporting aggregated results to enable a reliable assessment of progress. Finally, we report results for 20 baselines to gain information about the performance of existing models. We believe that this benchmark will be a driver of progress across a variety of Earth monitoring tasks.

Nasim Rahaman, Martin Weiss, Frederik Träuble et al.

Geospatial Information Systems are used by researchers and Humanitarian Assistance and Disaster Response (HADR) practitioners to support a wide variety of important applications. However, collaboration between these actors is difficult due to the heterogeneous nature of geospatial data modalities (e.g., multi-spectral images of various resolutions, timeseries, weather data) and diversity of tasks (e.g., regression of human activity indicators or detecting forest fires). In this work, we present a roadmap towards the construction of a general-purpose neural architecture (GPNA) with a geospatial inductive bias, pre-trained on large amounts of unlabelled earth observation data in a self-supervised manner. We envision how such a model may facilitate cooperation between members of the community. We show preliminary results on the first step of the roadmap, where we instantiate an architecture that can process a wide variety of geospatial data modalities and demonstrate that it can achieve competitive performance with domain-specific architectures on tasks relating to the U.N.'s Sustainable Development Goals.

Pietro Mazzaglia, Tim Verbelen, Bart Dhoedt et al.

Unsupervised skill learning aims to learn a rich repertoire of behaviors without external supervision, providing artificial agents with the ability to control and influence the environment. However, without appropriate knowledge and exploration, skills may provide control only over a restricted area of the environment, limiting their applicability. Furthermore, it is unclear how to leverage the learned skill behaviors for adapting to downstream tasks in a data-efficient manner. We present Choreographer, a model-based agent that exploits its world model to learn and adapt skills in imagination. Our method decouples the exploration and skill learning processes, being able to discover skills in the latent state space of the model. During adaptation, the agent uses a meta-controller to evaluate and adapt the learned skills efficiently by deploying them in parallel in imagination. Choreographer is able to learn skills both from offline data, and by collecting data simultaneously with an exploration policy. The skills can be used to effectively adapt to downstream tasks, as we show in the URL benchmark, where we outperform previous approaches from both pixels and states inputs. The learned skills also explore the environment thoroughly, finding sparse rewards more frequently, as shown in goal-reaching tasks from the DMC Suite and Meta-World. Website and code: https://skillchoreographer.github.io/

Sai Rajeswar, Pietro Mazzaglia, Tim Verbelen et al.

Controlling artificial agents from visual sensory data is an arduous task. Reinforcement learning (RL) algorithms can succeed but require large amounts of interactions between the agent and the environment. To alleviate the issue, unsupervised RL proposes to employ self-supervised interaction and learning, for adapting faster to future tasks. Yet, as shown in the Unsupervised RL Benchmark (URLB; Laskin et al. 2021), whether current unsupervised strategies can improve generalization capabilities is still unclear, especially in visual control settings. In this work, we study the URLB and propose a new method to solve it, using unsupervised model-based RL, for pre-training the agent, and a task-aware fine-tuning strategy combined with a new proposed hybrid planner, Dyna-MPC, to adapt the agent for downstream tasks. On URLB, our method obtains 93.59% overall normalized performance, surpassing previous baselines by a staggering margin. The approach is empirically evaluated through a large-scale empirical study, which we use to validate our design choices and analyze our models. We also show robust performance on the Real-Word RL benchmark, hinting at resiliency to environment perturbations during adaptation. Project website: https://masteringurlb.github.io/

Nishanth Madhusudhan, Vikas Yadav, Alexandre Lacoste

Effective abstention (EA), recognizing evidence insufficiency and refraining from answering, is critical for reliable multimodal systems. Yet existing evaluation paradigms for vision-language models (VLMs) and multi-agent systems (MAS) assume answerability, pushing models to always respond. Abstention has been studied in text-only settings but remains underexplored multimodally; current benchmarks either ignore unanswerability or rely on coarse methods that miss realistic failure modes. We introduce MM-AQA, a benchmark that constructs unanswerable instances from answerable ones via transformations along two axes: visual modality dependency and evidence sufficiency. Evaluating three frontier VLMs spanning closed and open-source models and two MAS architectures across 2079 samples, we find: (1) under standard prompting, VLMs rarely abstain; even simple confidence baselines outperform this setup, (2) MAS improves abstention but introduces an accuracy-abstention trade-off, (3) sequential designs match or exceed iterative variants, suggesting the bottleneck is miscalibration rather than reasoning depth, and (4) models abstain when image or text evidence is absent, but attempt reconciliation with degraded or contradictory evidence. Effective multimodal abstention requires abstention-aware training rather than better prompting or more agents.

Rishika Bhagwatkar, Kevin Kasa, Abhay Puri et al.

AI agents are vulnerable to indirect prompt injection attacks, where malicious instructions embedded in external content or tool outputs cause unintended or harmful behavior. Inspired by the well-established concept of firewalls, we show that a simple, modular, and model-agnostic defense operating at the agent--tool interface achieves perfect security with high utility across all four public benchmarks: AgentDojo, Agent Security Bench, InjecAgent and tau-Bench, while achieving a state-of-the-art security--utility tradeoff compared to prior results. Specifically, we employ two firewalls: a Tool-Input Firewall (Minimizer) and a Tool-Output Firewall (Sanitizer). Unlike prior complex approaches, this defense makes minimal assumptions about the agent and can be deployed out of the box. This makes it highly generalizable while maintaining strong performance without compromising utility. Our analysis also reveals critical limitations in these existing benchmarks, including flawed success metrics, implementation bugs, and most importantly, weak attacks, hindering progress. To address this, we present targeted fixes to these issues for AgentDojo and Agent Security Bench, and propose best practices for more robust benchmark design. Moreover, we introduce a three-stage attack strategy that cascades standard prompt injection attacks, second-order attacks, and adaptive attacks to evaluate the robustness beyond existing attacks. Overall, our work shows that existing agentic security benchmarks are easily saturated by a simple approach and highlights the need for stronger benchmarks with carefully chosen evaluation metrics and strong adaptive attacks.

Alexandre Lacoste, Nicolas Gontier, Oleh Shliazhko et al. · ibm-research

The proliferation of agent benchmarks has created critical fragmentation that threatens research productivity. Each new benchmark requires substantial custom integration, creating an "integration tax" that limits comprehensive evaluation. We propose CUBE (Common Unified Benchmark Environments), a universal protocol standard built on MCP and Gym that allows benchmarks to be wrapped once and used everywhere. By separating task, benchmark, package, and registry concerns into distinct API layers, CUBE enables any compliant platform to access any compliant benchmark for evaluation, RL training, or data generation without custom integration. We call on the community to contribute to the development of this standard before platform-specific implementations deepen fragmentation as benchmark production accelerates through 2026.

Muhammad Sohail Danish, Muhammad Akhtar Munir, Syed Roshaan Ali Shah et al.

While numerous recent benchmarks focus on evaluating generic Vision-Language Models (VLMs), they do not effectively address the specific challenges of geospatial applications. Generic VLM benchmarks are not designed to handle the complexities of geospatial data, an essential component for applications such as environmental monitoring, urban planning, and disaster management. Key challenges in the geospatial domain include temporal change detection, large-scale object counting, tiny object detection, and understanding relationships between entities in remote sensing imagery. To bridge this gap, we present GEOBench-VLM, a comprehensive benchmark specifically designed to evaluate VLMs on geospatial tasks, including scene understanding, object counting, localization, fine-grained categorization, segmentation, and temporal analysis. Our benchmark features over 10,000 manually verified instructions and spanning diverse visual conditions, object types, and scales. We evaluate several state-of-the-art VLMs to assess performance on geospatial-specific challenges. The results indicate that although existing VLMs demonstrate potential, they face challenges when dealing with geospatial-specific tasks, highlighting the room for further improvements. Notably, the best-performing LLaVa-OneVision achieves only 41.7% accuracy on MCQs, slightly more than GPT-4o, which is approximately double the random guess performance. Our benchmark is publicly available at https://github.com/The-AI-Alliance/GEO-Bench-VLM .

Xiaoqiang Wang, Chao Wang, Hadi Nekoei et al.

We present Mem-$π$, a framework for adaptive memory in large language model (LLM) agents, where useful guidance is generated on demand rather than retrieved from external memory stores. Existing memory-augmented agents typically rely on similarity-based retrieval from episodic memory banks or skill libraries, returning static entries that often misalign with the current context. In contrast, Mem-$π$ uses a dedicated language or vision-language model with its own parameters, separate from the downstream agent, to generate context-specific guidance for complex tasks. Conditioned on the current agent context, the model jointly decides when to produce guidance and what guidance to produce. We train it with a decision-content decoupled reinforcement learning (RL) objective, enabling it to abstain when generation would not help and otherwise produce concise, useful guidance. Across diverse agentic benchmarks spanning web navigation, terminal-based tool use, and text-based embodied interaction, Mem-$π$ consistently outperforms retrieval-based and prior RL-optimized memory baselines, achieving over 30% relative improvement on web navigation tasks.

Dheeraj Vattikonda, Santhoshi Ravichandran, Emiliano Penaloza et al. · mila

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents. To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B student to imitate a Llama 3.3 70B teacher via supervised fine-tuning (SFT), followed by on-policy reinforcement learning. We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Alexandre Drouin, Maxime Gasse, Massimo Caccia et al.

We study the use of large language model-based agents for interacting with software via web browsers. Unlike prior work, we focus on measuring the agents' ability to perform tasks that span the typical daily work of knowledge workers utilizing enterprise software systems. To this end, we propose WorkArena, a remote-hosted benchmark of 33 tasks based on the widely-used ServiceNow platform. We also introduce BrowserGym, an environment for the design and evaluation of such agents, offering a rich set of actions as well as multimodal observations. Our empirical evaluation reveals that while current agents show promise on WorkArena, there remains a considerable gap towards achieving full task automation. Notably, our analysis uncovers a significant performance disparity between open and closed-source LLMs, highlighting a critical area for future exploration and development in the field.

Hadi Nekoei, Aman Jaiswal, Patrice Bechard et al.

Large language model (LLM) agents perform well in sequential decision-making tasks, but improving them on unfamiliar domains often requires costly online interactions or fine-tuning on large expert datasets. These strategies are impractical for closed-source models and expensive for open-source ones, with risks of catastrophic forgetting. Offline trajectories offer reusable knowledge, yet demonstration-based methods struggle because raw traces are long, noisy, and tied to specific tasks. We present Just-in-time Episodic Feedback Hinter (JEF Hinter), an agentic system that distills offline traces into compact, context-aware hints. A zooming mechanism highlights decisive steps in long trajectories, capturing both strategies and pitfalls. Unlike prior methods, JEF Hinter leverages both successful and failed trajectories, extracting guidance even when only failure data is available, while supporting parallelized hint generation and benchmark-independent prompting. At inference, a retriever selects relevant hints for the current state, providing targeted guidance with transparency and traceability. Experiments on MiniWoB++, WorkArena-L1, and WebArena-Lite show that JEF Hinter consistently outperforms strong baselines, including human- and document-based hints.

Pau Rodriguez, Massimo Caccia, Alexandre Lacoste et al.

Explainability for machine learning models has gained considerable attention within the research community given the importance of deploying more reliable machine-learning systems. In computer vision applications, generative counterfactual methods indicate how to perturb a model's input to change its prediction, providing details about the model's decision-making. Current methods tend to generate trivial counterfactuals about a model's decisions, as they often suggest to exaggerate or remove the presence of the attribute being classified. For the machine learning practitioner, these types of counterfactuals offer little value, since they provide no new information about undesired model or data biases. In this work, we identify the problem of trivial counterfactual generation and we propose DiVE to alleviate it. DiVE learns a perturbation in a disentangled latent space that is constrained using a diversity-enforcing loss to uncover multiple valuable explanations about the model's prediction. Further, we introduce a mechanism to prevent the model from producing trivial explanations. Experiments on CelebA and Synbols demonstrate that our model improves the success rate of producing high-quality valuable explanations when compared to previous state-of-the-art methods. Code is available at https://github.com/ElementAI/beyond-trivial-explanations.

Issam Laradji, Pau Rodriguez, Freddie Kalaitzis et al.

Cattle farming is responsible for 8.8\% of greenhouse gas emissions worldwide. In addition to the methane emitted due to their digestive process, the growing need for grazing areas is an important driver of deforestation. While some regulations are in place for preserving the Amazon against deforestation, these are being flouted in various ways, hence the need to scale and automate the monitoring of cattle ranching activities. Through a partnership with \textit{Global Witness}, we explore the feasibility of tracking and counting cattle at the continental scale from satellite imagery. With a license from Maxar Technologies, we obtained satellite imagery of the Amazon at 40cm resolution, and compiled a dataset of 903 images containing a total of 28498 cattle. Our experiments show promising results and highlight important directions for the next steps on both counting algorithms and the data collection process for solving such challenges. The code is available at \url{https://github.com/IssamLaradji/cownter_strike}.

Parmida Atighehchian, Frédéric Branchaud-Charron, Alexandre Lacoste

Active learning is able to reduce the amount of labelling effort by using a machine learning model to query the user for specific inputs. While there are many papers on new active learning techniques, these techniques rarely satisfy the constraints of a real-world project. In this paper, we analyse the main drawbacks of current active learning techniques and we present approaches to alleviate them. We do a systematic study on the effects of the most common issues of real-world datasets on the deep active learning process: model convergence, annotation error, and dataset imbalance. We derive two techniques that can speed up the active learning loop such as partial uncertainty sampling and larger query size. Finally, we present our open-source Bayesian active learning library, BaaL.

Pau Rodríguez, Issam Laradji, Alexandre Drouin et al.

Few-shot classification is challenging because the data distribution of the training set can be widely different to the test set as their classes are disjoint. This distribution shift often results in poor generalization. Manifold smoothing has been shown to address the distribution shift problem by extending the decision boundaries and reducing the noise of the class representations. Moreover, manifold smoothness is a key factor for semi-supervised learning and transductive learning algorithms. In this work, we propose to use embedding propagation as an unsupervised non-parametric regularizer for manifold smoothing in few-shot classification. Embedding propagation leverages interpolations between the extracted features of a neural network based on a similarity graph. We empirically show that embedding propagation yields a smoother embedding manifold. We also show that applying embedding propagation to a transductive classifier achieves new state-of-the-art results in mini-Imagenet, tiered-Imagenet, Imagenet-FS, and CUB. Furthermore, we show that embedding propagation consistently improves the accuracy of the models in multiple semi-supervised learning scenarios by up to 16\% points. The proposed embedding propagation operation can be easily integrated as a non-parametric layer into a neural network. We provide the training code and usage examples at https://github.com/ElementAI/embedding-propagation.

Boris N. Oreshkin, Pau Rodriguez, Alexandre Lacoste

Few-shot learning has become essential for producing models that generalize from few examples. In this work, we identify that metric scaling and metric task conditioning are important to improve the performance of few-shot algorithms. Our analysis reveals that simple metric scaling completely changes the nature of few-shot algorithm parameter updates. Metric scaling provides improvements up to 14% in accuracy for certain metrics on the mini-Imagenet 5-way 5-shot classification task. We further propose a simple and effective way of conditioning a learner on the task sample set, resulting in learning a task-dependent metric space. Moreover, we propose and empirically test a practical end-to-end optimization procedure based on auxiliary task co-training to learn a task-dependent metric space. The resulting few-shot learning model based on the task-dependent scaled metric achieves state of the art on mini-Imagenet. We confirm these results on another few-shot dataset that we introduce in this paper based on CIFAR100. Our code is publicly available at https://github.com/ElementAI/TADAM.

Thibault Le Sellier De Chezelles, Maxime Gasse, Alexandre Drouin et al. · mila

The BrowserGym ecosystem addresses the growing need for efficient evaluation and benchmarking of web agents, particularly those leveraging automation and Large Language Models (LLMs). Many existing benchmarks suffer from fragmentation and inconsistent evaluation methodologies, making it challenging to achieve reliable comparisons and reproducible results. In an earlier work, Drouin et al. (2024) introduced BrowserGym which aims to solve this by providing a unified, gym-like environment with well-defined observation and action spaces, facilitating standardized evaluation across diverse benchmarks. We propose an extended BrowserGym-based ecosystem for web agent research, which unifies existing benchmarks from the literature and includes AgentLab, a complementary framework that aids in agent creation, testing, and analysis. Our proposed ecosystem offers flexibility for integrating new benchmarks while ensuring consistent evaluation and comprehensive experiment management. As a supporting evidence, we conduct the first large-scale, multi-benchmark web agent experiment and compare the performance of 6 state-of-the-art LLMs across 6 popular web agent benchmarks made available in BrowserGym. Among other findings, our results highlight a large discrepancy between OpenAI and Anthropic's latests models, with Claude-3.5-Sonnet leading the way on almost all benchmarks, except on vision-related tasks where GPT-4o is superior. Despite these advancements, our results emphasize that building robust and efficient web agents remains a significant challenge, due to the inherent complexity of real-world web environments and the limitations of current models.

Andrew Robert Williams, Arjun Ashok, Étienne Marcotte et al.

Forecasting is a critical task in decision-making across numerous domains. While historical numerical data provide a start, they fail to convey the complete context for reliable and accurate predictions. Human forecasters frequently rely on additional information, such as background knowledge and constraints, which can efficiently be communicated through natural language. However, in spite of recent progress with LLM-based forecasters, their ability to effectively integrate this textual information remains an open question. To address this, we introduce "Context is Key" (CiK), a time-series forecasting benchmark that pairs numerical data with diverse types of carefully crafted textual context, requiring models to integrate both modalities; crucially, every task in CiK requires understanding textual context to be solved successfully. We evaluate a range of approaches, including statistical models, time series foundation models, and LLM-based forecasters, and propose a simple yet effective LLM prompting method that outperforms all other tested methods on our benchmark. Our experiments highlight the importance of incorporating contextual information, demonstrate surprising performance when using LLM-based forecasting models, and also reveal some of their critical shortcomings. This benchmark aims to advance multimodal forecasting by promoting models that are both accurate and accessible to decision-makers with varied technical expertise. The benchmark can be visualized at https://servicenow.github.io/context-is-key-forecasting/v0/.

Sébastien Lachapelle, Pau Rodríguez López, Yash Sharma et al.

This work introduces a novel principle for disentanglement we call mechanism sparsity regularization, which applies when the latent factors of interest depend sparsely on observed auxiliary variables and/or past latent factors. We propose a representation learning method that induces disentanglement by simultaneously learning the latent factors and the sparse causal graphical model that explains them. We develop a nonparametric identifiability theory that formalizes this principle and shows that the latent factors can be recovered by regularizing the learned causal graph to be sparse. More precisely, we show identifiablity up to a novel equivalence relation we call "consistency", which allows some latent factors to remain entangled (hence the term partial disentanglement). To describe the structure of this entanglement, we introduce the notions of entanglement graphs and graph preserving functions. We further provide a graphical criterion which guarantees complete disentanglement, that is identifiability up to permutations and element-wise transformations. We demonstrate the scope of the mechanism sparsity principle as well as the assumptions it relies on with several worked out examples. For instance, the framework shows how one can leverage multi-node interventions with unknown targets on the latent factors to disentangle them. We further draw connections between our nonparametric results and the now popular exponential family assumption. Lastly, we propose an estimation procedure based on variational autoencoders and a sparsity constraint and demonstrate it on various synthetic datasets. This work is meant to be a significantly extended version of Lachapelle et al. (2022).

Diego Velazquez, Pau Rodriguez López, Sergio Alonso et al.

This paper presents EarthView, a comprehensive dataset specifically designed for self-supervision on remote sensing data, intended to enhance deep learning applications on Earth monitoring tasks. The dataset spans 15 tera pixels of global remote-sensing data, combining imagery from a diverse range of sources, including NEON, Sentinel, and a novel release of 1m spatial resolution data from Satellogic. Our dataset provides a wide spectrum of image data with varying resolutions, harnessed from different sensors and organized coherently into an accessible HuggingFace dataset in parquet format. This data spans five years, from 2017 to 2022. Accompanying the dataset, we introduce EarthMAE, a tailored Masked Autoencoder, developed to tackle the distinct challenges of remote sensing data. Trained in a self-supervised fashion, EarthMAE effectively processes different data modalities such as hyperspectral, multispectral, topographical data, segmentation maps, and temporal structure. This model helps us show that pre-training on Satellogic data improves performance on downstream tasks. While there is still a gap to fill in MAE for heterogeneous data, we regard this innovative combination of an expansive, diverse dataset and a versatile model adapted for self-supervised learning as a stride forward in deep learning for Earth monitoring.

Emiliano Penaloza, Dheeraj Vattikonda, Nicolas Gontier et al.

Training-time privileged information (PI) can enable language models to succeed on tasks they would otherwise fail, making it a powerful tool for reinforcement learning in hard, long-horizon settings. However, transferring capabilities learned with PI to policies that must act without it at inference time remains a fundamental challenge. We study this problem in the context of distilling frontier models for multi-turn agentic environments, where closed-source systems typically hide their internal reasoning and expose only action trajectories. This breaks standard distillation pipelines, since successful behavior is observable but the reasoning process is not. For this, we introduce π-Distill, a joint teacher-student objective that trains a PI-conditioned teacher and an unconditioned student simultaneously using the same model. Additionally, we also introduce On-Policy Self-Distillation (OPSD), an alternative approach that trains using Reinforcement Learning (RL) with a reverse KL-penalty between the student and the PI-conditioned teacher. We show that both of these algorithms effectively distill frontier agents using action-only PI. Specifically we find that π-Distill and in some cases OPSD, outperform industry standard practices (Supervised finetuning followed by RL) that assume access to full Chain-of-Thought supervision across multiple agentic benchmarks, models, and forms of PI. We complement our results with extensive analysis that characterizes the factors enabling effective learning with PI, focusing primarily on π-Distill and characterizing when OPSD is competitive.

Issam Laradji, Perouz Taslakian, Sai Rajeswar et al.

The extraction of a small number of relevant insights from vast amounts of data is a crucial component of data-driven decision-making. However, accomplishing this task requires considerable technical skills, domain expertise, and human labor. This study explores the potential of using Large Language Models (LLMs) to automate the discovery of insights in data, leveraging recent advances in reasoning and code generation techniques. We propose a new evaluation methodology based on a "capture the flag" principle, measuring the ability of such models to recognize meaningful and pertinent information (flags) in a dataset. We further propose two proof-of-concept agents, with different inner workings, and compare their ability to capture such flags in a real-world sales dataset. While the work reported here is preliminary, our results are sufficiently interesting to mandate future exploration by the community.

Imene Kerboua, Sahar Omidi Shayegan, Megh Thakkar et al. · mila

Web agents powered by large language models (LLMs) must process lengthy web page observations to complete user goals; these pages often exceed tens of thousands of tokens. This saturates context limits and increases computational cost processing; moreover, processing full pages exposes agents to security risks such as prompt injection. Existing pruning strategies either discard relevant content or retain irrelevant context, leading to suboptimal action prediction. We introduce FocusAgent, a simple yet effective approach that leverages a lightweight LLM retriever to extract the most relevant lines from accessibility tree (AxTree) observations, guided by task goals. By pruning noisy and irrelevant content, FocusAgent enables efficient reasoning while reducing vulnerability to injection attacks. Experiments on WorkArena and WebArena benchmarks show that FocusAgent matches the performance of strong baselines, while reducing observation size by over 50%. Furthermore, a variant of FocusAgent significantly reduces the success rate of prompt-injection attacks, including banner and pop-up attacks, while maintaining task success performance in attack-free settings. Our results highlight that targeted LLM-based retrieval is a practical and robust strategy for building web agents that are efficient, effective, and secure.

Léo Boisvert, Abhay Puri, Chandra Kiran Reddy Evuru et al.

The practice of fine-tuning AI agents on data from their own interactions--such as web browsing or tool use--, while being a strong general recipe for improving agentic capabilities, also introduces a critical security vulnerability within the AI supply chain. In this work, we show that adversaries can easily poison the data collection pipeline to embed hard-to-detect backdoors that are triggerred by specific target phrases, such that when the agent encounters these triggers, it performs an unsafe or malicious action. We formalize and validate three realistic threat models targeting different layers of the supply chain: 1) direct poisoning of fine-tuning data, where an attacker controls a fraction of the training traces; 2) environmental poisoning, where malicious instructions are injected into webpages scraped or tools called while creating training data; and 3) supply chain poisoning, where a pre-backdoored base model is fine-tuned on clean data to improve its agentic capabilities. Our results are stark: by poisoning as few as 2% of the collected traces, an attacker can embed a backdoor causing an agent to leak confidential user information with over 80% success when a specific trigger is present. This vulnerability holds across all three threat models. Furthermore, we demonstrate that prominent safeguards, including two guardrail models and one weight-based defense, fail to detect or prevent the malicious behavior. These findings highlight an urgent threat to agentic AI development and underscore the critical need for rigorous security vetting of data collection processes and end-to-end model supply chains.

Imene Kerboua, Sahar Omidi Shayegan, Megh Thakkar et al. · mila

While large language models have demonstrated impressive capabilities in web navigation tasks, the extensive context of web pages, often represented as DOM or Accessibility Tree (AxTree) structures, frequently exceeds model context limits. Current approaches like bottom-up truncation or embedding-based retrieval lose critical information about page state and action history. This is particularly problematic for adaptive planning in web agents, where understanding the current state is essential for determining future actions. We hypothesize that embedding models lack sufficient capacity to capture plan-relevant information, especially when retrieving content that supports future action prediction. This raises a fundamental question: how can retrieval methods be optimized for adaptive planning in web navigation tasks? In response, we introduce \textit{LineRetriever}, a novel approach that leverages a language model to identify and retrieve observation lines most relevant to future navigation steps. Unlike traditional retrieval methods that focus solely on semantic similarity, \textit{LineRetriever} explicitly considers the planning horizon, prioritizing elements that contribute to action prediction. Our experiments demonstrate that \textit{LineRetriever} can reduce the size of the observation at each step for the web agent while maintaining consistent performance within the context limitations.

Alexandre Lacoste, Evan David Sherwin, Hannah Kerner et al.

Recent progress in self-supervision shows that pre-training large neural networks on vast amounts of unsupervised data can lead to impressive increases in generalisation for downstream tasks. Such models, recently coined as foundation models, have been transformational to the field of natural language processing. While similar models have also been trained on large corpuses of images, they are not well suited for remote sensing data. To stimulate the development of foundation models for Earth monitoring, we propose to develop a new benchmark comprised of a variety of downstream tasks related to climate change. We believe that this can lead to substantial improvements in many existing applications and facilitate the development of new applications. This proposal is also a call for collaboration with the aim of developing a better evaluation process to mitigate potential downsides of foundation models for Earth monitoring.

Philippe Brouillard, Perouz Taslakian, Alexandre Lacoste et al.

Causal discovery from observational data is a challenging task that can only be solved up to a set of equivalent solutions, called an equivalence class. Such classes, which are often large in size, encode uncertainties about the orientation of some edges in the causal graph. In this work, we propose a new set of assumptions that constrain possible causal relationships based on the nature of variables, thus circumscribing the equivalence class. Namely, we introduce typed directed acyclic graphs, in which variable types are used to determine the validity of causal relationships. We demonstrate, both theoretically and empirically, that the proposed assumptions can result in significant gains in the identification of the causal graph. We also propose causal discovery algorithms that make use of these assumptions and demonstrate their benefits on simulated and pseudo-real data.

Sébastien Lachapelle, Pau Rodríguez López, Yash Sharma et al.

This work introduces a novel principle we call disentanglement via mechanism sparsity regularization, which can be applied when the latent factors of interest depend sparsely on past latent factors and/or observed auxiliary variables. We propose a representation learning method that induces disentanglement by simultaneously learning the latent factors and the sparse causal graphical model that relates them. We develop a rigorous identifiability theory, building on recent nonlinear independent component analysis (ICA) results, that formalizes this principle and shows how the latent variables can be recovered up to permutation if one regularizes the latent mechanisms to be sparse and if some graph connectivity criterion is satisfied by the data generating process. As a special case of our framework, we show how one can leverage unknown-target interventions on the latent factors to disentangle them, thereby drawing further connections between ICA and causality. We propose a VAE-based method in which the latent mechanisms are learned and regularized via binary masks, and validate our theory by showing it learns disentangled representations in simulations.

Yashas Annadani, Jonas Rothfuss, Alexandre Lacoste et al.

Learning the causal structure that underlies data is a crucial step towards robust real-world decision making. The majority of existing work in causal inference focuses on determining a single directed acyclic graph (DAG) or a Markov equivalence class thereof. However, a crucial aspect to acting intelligently upon the knowledge about causal structure which has been inferred from finite data demands reasoning about its uncertainty. For instance, planning interventions to find out more about the causal mechanisms that govern our data requires quantifying epistemic uncertainty over DAGs. While Bayesian causal inference allows to do so, the posterior over DAGs becomes intractable even for a small number of variables. Aiming to overcome this issue, we propose a form of variational inference over the graphs of Structural Causal Models (SCMs). To this end, we introduce a parametric variational family modelled by an autoregressive distribution over the space of discrete DAGs. Its number of parameters does not grow exponentially with the number of variables and can be tractably learned by maximising an Evidence Lower Bound (ELBO). In our experiments, we demonstrate that the proposed variational posterior is able to provide a good approximation of the true posterior.

Frédéric Branchaud-Charron, Parmida Atighehchian, Pau Rodríguez et al.

Dataset bias is one of the prevailing causes of unfairness in machine learning. Addressing fairness at the data collection and dataset preparation stages therefore becomes an essential part of training fairer algorithms. In particular, active learning (AL) algorithms show promise for the task by drawing importance to the most informative training samples. However, the effect and interaction between existing AL algorithms and algorithmic fairness remain under-explored. In this paper, we study whether models trained with uncertainty-based AL heuristics such as BALD are fairer in their decisions with respect to a protected class than those trained with identically independently distributed (i.i.d.) sampling. We found a significant improvement on predictive parity when using BALD, while also improving accuracy compared to i.i.d. sampling. We also explore the interaction of algorithmic fairness methods such as gradient reversal (GRAD) and BALD. We found that, while addressing different fairness issues, their interaction further improves the results on most benchmarks and metrics we explored.

Oscar Mañas, Alexandre Lacoste, Xavier Giro-i-Nieto et al.

Remote sensing and automatic earth monitoring are key to solve global-scale challenges such as disaster prevention, land use monitoring, or tackling climate change. Although there exist vast amounts of remote sensing data, most of it remains unlabeled and thus inaccessible for supervised learning algorithms. Transfer learning approaches can reduce the data requirements of deep learning algorithms. However, most of these methods are pre-trained on ImageNet and their generalization to remote sensing imagery is not guaranteed due to the domain gap. In this work, we propose Seasonal Contrast (SeCo), an effective pipeline to leverage unlabeled data for in-domain pre-training of remote sensing representations. The SeCo pipeline is composed of two parts. First, a principled procedure to gather large-scale, unlabeled and uncurated remote sensing datasets containing images from multiple Earth locations at different timestamps. Second, a self-supervised algorithm that takes advantage of time and position invariance to learn transferable representations for remote sensing applications. We empirically show that models trained with SeCo achieve better performance than their ImageNet pre-trained counterparts and state-of-the-art self-supervised learning methods on multiple downstream tasks. The datasets and models in SeCo will be made public to facilitate transfer learning and enable rapid progress in remote sensing applications.

Alexandre Lacoste, Pau Rodríguez, Frédéric Branchaud-Charron et al.

Progress in the field of machine learning has been fueled by the introduction of benchmark datasets pushing the limits of existing algorithms. Enabling the design of datasets to test specific properties and failure modes of learning algorithms is thus a problem of high interest, as it has a direct impact on innovation in the field. In this sense, we introduce Synbols -- Synthetic Symbols -- a tool for rapidly generating new datasets with a rich composition of latent features rendered in low resolution images. Synbols leverages the large amount of symbols available in the Unicode standard and the wide range of artistic font provided by the open font community. Our tool's high-level interface provides a language for rapidly generating new distributions on the latent features, including various types of textures and occlusions. To showcase the versatility of Synbols, we use it to dissect the limitations and flaws in standard learning algorithms in various learning setups including supervised learning, active learning, out of distribution generalization, unsupervised representation learning, and object counting.

Philippe Brouillard, Sébastien Lachapelle, Alexandre Lacoste et al.

Learning a causal directed acyclic graph from data is a challenging task that involves solving a combinatorial problem for which the solution is not always identifiable. A new line of work reformulates this problem as a continuous constrained optimization one, which is solved via the augmented Lagrangian method. However, most methods based on this idea do not make use of interventional data, which can significantly alleviate identifiability issues. This work constitutes a new step in this direction by proposing a theoretically-grounded method based on neural networks that can leverage interventional data. We illustrate the flexibility of the continuous-constrained framework by taking advantage of expressive neural architectures such as normalizing flows. We show that our approach compares favorably to the state of the art in a variety of settings, including perfect and imperfect interventions for which the targeted nodes may even be unknown.

Massimo Caccia, Pau Rodriguez, Oleksiy Ostapenko et al.

Continual learning studies agents that learn from streams of tasks without forgetting previous ones while adapting to new ones. Two recent continual-learning scenarios have opened new avenues of research. In meta-continual learning, the model is pre-trained to minimize catastrophic forgetting of previous tasks. In continual-meta learning, the aim is to train agents for faster remembering of previous tasks through adaptation. In their original formulations, both methods have limitations. We stand on their shoulders to propose a more general scenario, OSAKA, where an agent must quickly solve new (out-of-distribution) tasks, while also requiring fast remembering. We show that current continual learning, meta-learning, meta-continual learning, and continual-meta learning techniques fail in this new scenario. We propose Continual-MAML, an online extension of the popular MAML algorithm as a strong baseline for this scenario. We empirically show that Continual-MAML is better suited to the new scenario than the aforementioned methodologies, as well as standard continual learning and meta-learning approaches.

Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt et al.

From an environmental standpoint, there are a few crucial aspects of training a neural network that have a major impact on the quantity of carbon that it emits. These factors include: the location of the server used for training and the energy grid that it uses, the length of the training procedure, and even the make and model of hardware on which the training takes place. In order to approximate these emissions, we present our Machine Learning Emissions Calculator, a tool for our community to better understand the environmental impact of training ML models. We accompany this tool with an explanation of the factors cited above, as well as concrete actions that individual practitioners and organizations can take to mitigate their carbon emissions.

Chin-Wei Huang, Ahmed Touati, Pascal Vincent et al.

Recent advances in variational inference enable the modelling of highly structured joint distributions, but are limited in their capacity to scale to the high-dimensional setting of stochastic neural networks. This limitation motivates a need for scalable parameterizations of the noise generation process, in a manner that adequately captures the dependencies among the various parameters. In this work, we address this need and present the Kronecker Flow, a generalization of the Kronecker product to invertible mappings designed for stochastic neural networks. We apply our method to variational Bayesian neural networks on predictive tasks, PAC-Bayes generalization bound estimation, and approximate Thompson sampling in contextual bandits. In all setups, our methods prove to be competitive with existing methods and better than the baselines.

David Rolnick, Priya L. Donti, Lynn H. Kaack et al.

Climate change is one of the greatest challenges facing humanity, and we, as machine learning experts, may wonder how we can help. Here we describe how machine learning can be a powerful tool in reducing greenhouse gas emissions and helping society adapt to a changing climate. From smart grids to disaster management, we identify high impact problems where existing gaps can be filled by machine learning, in collaboration with other fields. Our recommendations encompass exciting research questions as well as promising business opportunities. We call on the machine learning community to join the global effort against climate change.

Prudencio Tossou, Basile Dura, Francois Laviolette et al.

Deep kernel learning provides an elegant and principled framework for combining the structural properties of deep learning algorithms with the flexibility of kernel methods. By means of a deep neural network, we learn a parametrized kernel operator that can be combined with a differentiable kernel algorithm during inference. While previous work within this framework has focused on learning a single kernel for large datasets, we learn a kernel family for a variety of few-shot regression tasks. Compared to single deep kernel learning, our algorithm enables the identification of the appropriate kernel for each task during inference. As such, it is well adapted for complex task distributions in a few-shot learning setting, which we demonstrate by comparing against existing state-of-the-art algorithms using real-world, few-shot regression tasks related to the field of drug discovery.

Chin-Wei Huang, Kris Sankaran, Eeshan Dhekane et al.

Importance weighted variational inference (Burda et al., 2015) uses multiple i.i.d. samples to have a tighter variational lower bound. We believe a joint proposal has the potential of reducing the number of redundant samples, and introduce a hierarchical structure to induce correlation. The hope is that the proposals would coordinate to make up for the error made by one another to reduce the variance of the importance estimator. Theoretically, we analyze the condition under which convergence of the estimator variance can be connected to convergence of the lower bound. Empirically, we confirm that maximization of the lower bound does implicitly minimize variance. Further analysis shows that this is a result of negative correlation induced by the proposed hierarchical meta sampling scheme, and performance of inference also improves when the number of samples increases.

Chin-Wei Huang, Shawn Tan, Alexandre Lacoste et al.

Despite the advances in the representational capacity of approximate distributions for variational inference, the optimization process can still limit the density that is ultimately learned. We demonstrate the drawbacks of biasing the true posterior to be unimodal, and introduce Annealed Variational Objectives (AVO) into the training of hierarchical variational methods. Inspired by Annealed Importance Sampling, the proposed method facilitates learning by incorporating energy tempering into the optimization objective. In our experiments, we demonstrate our method's robustness to deterministic warm up, and the benefits of encouraging exploration in the latent space.

Alexandre Lacoste, Boris Oreshkin, Wonchang Chung et al.

Using variational Bayes neural networks, we develop an algorithm capable of accumulating knowledge into a prior from multiple different tasks. The result is a rich and meaningful prior capable of few-shot learning on new tasks. The posterior can go beyond the mean field approximation and yields good uncertainty on the performed experiments. Analysis on toy tasks shows that it can learn from significantly different tasks while finding similarities among them. Experiments of Mini-Imagenet yields the new state of the art with 74.5% accuracy on 5 shot learning. Finally, we provide experiments showing that other existing methods can fail to perform well in different benchmarks.

Chin-Wei Huang, David Krueger, Alexandre Lacoste et al.

Normalizing flows and autoregressive models have been successfully combined to produce state-of-the-art results in density estimation, via Masked Autoregressive Flows (MAF), and to accelerate state-of-the-art WaveNet-based speech synthesis to 20x faster than real-time, via Inverse Autoregressive Flows (IAF). We unify and generalize these approaches, replacing the (conditionally) affine univariate transformations of MAF/IAF with a more general class of invertible univariate transformations expressed as monotonic neural networks. We demonstrate that the proposed neural autoregressive flows (NAF) are universal approximators for continuous probability distributions, and their greater expressivity allows them to better capture multimodal target distributions. Experimentally, NAF yields state-of-the-art performance on a suite of density estimation tasks and outperforms IAF in variational autoencoders trained on binarized MNIST.

Alexandre Lacoste, Thomas Boquet, Negar Rostamzadeh et al.

The recent literature on deep learning offers new tools to learn a rich probability distribution over high dimensional data such as images or sounds. In this work we investigate the possibility of learning the prior distribution over neural network parameters using such tools. Our resulting variational Bayes algorithm generalizes well to new tasks, even when very few training examples are provided. Furthermore, this learned prior allows the model to extrapolate correctly far from a given task's training data on a meta-dataset of periodic signals.

David Krueger, Chin-Wei Huang, Riashat Islam et al.

We study Bayesian hypernetworks: a framework for approximate Bayesian inference in neural networks. A Bayesian hypernetwork $\h$ is a neural network which learns to transform a simple noise distribution, $p(\vecε) = \N(\vec 0,\mat I)$, to a distribution $q(\pp) := q(h(\vecε))$ over the parameters $\pp$ of another neural network (the "primary network")\@. We train $q$ with variational inference, using an invertible $\h$ to enable efficient estimation of the variational lower bound on the posterior $p(\pp | \D)$ via sampling. In contrast to most methods for Bayesian deep learning, Bayesian hypernets can represent a complex multimodal approximate posterior with correlations between parameters, while enabling cheap iid sampling of~$q(\pp)$. In practice, Bayesian hypernets can provide a better defense against adversarial examples than dropout, and also exhibit competitive performance on a suite of tasks which evaluate model uncertainty, including regularization, active learning, and anomaly detection.

Eunsol Choi, Daniel Hewlett, Alexandre Lacoste et al.

We present a framework for question answering that can efficiently scale to longer documents while maintaining or even improving performance of state-of-the-art models. While most successful approaches for reading comprehension rely on recurrent neural networks (RNNs), running them over long documents is prohibitively slow because it is difficult to parallelize over sequences. Inspired by how people first skim the document, identify relevant parts, and carefully read these parts to produce an answer, we combine a coarse, fast model for selecting relevant sentences and a more expensive RNN for producing the answer from those sentences. We treat sentence selection as a latent variable trained jointly from the answer only using reinforcement learning. Experiments demonstrate the state of the art performance on a challenging subset of the Wikireading and on a new dataset, while speeding up the model by 3.5x-6.7x.

Daniel Hewlett, Alexandre Lacoste, Llion Jones et al.

We present WikiReading, a large-scale natural language understanding task and publicly-available dataset with 18 million instances. The task is to predict textual values from the structured knowledge base Wikidata by reading the text of the corresponding Wikipedia articles. The task contains a rich variety of challenging classification and extraction sub-tasks, making it well-suited for end-to-end models such as deep neural networks (DNNs). We compare various state-of-the-art DNN-based architectures for document classification, information extraction, and question answering. We find that models supporting a rich answer space, such as word or character sequences, perform best. Our best-performing model, a word-level sequence to sequence model with a mechanism to copy out-of-vocabulary words, obtains an accuracy of 71.8%.

Pascal Germain, Francis Bach, Alexandre Lacoste et al.

We exhibit a strong link between frequentist PAC-Bayesian risk bounds and the Bayesian marginal likelihood. That is, for the negative log-likelihood loss function, we show that the minimization of PAC-Bayesian generalization risk bounds maximizes the Bayesian marginal likelihood. This provides an alternative explanation to the Bayesian Occam's razor criteria, under the assumption that the data is generated by an i.i.d distribution. Moreover, as the negative log-likelihood is an unbounded loss function, we motivate and propose a PAC-Bayesian theorem tailored for the sub-gamma loss family, and we show that our approach is sound on classical Bayesian linear regression tasks.