Scalable simulation
Build the environments agents learn in. Post-training and self-improvement stop being hard the moment you have the right data and RL environments are how you manufacture it at scale.
Building systems that understand and act in digital worlds.
I am a Research Lead at ServiceNow Research, an Adjunct Professor, and a core industry member at Mila, Montréal. My group works toward generally intelligent machines that understand, reason, and act, grounded in a conviction that the hard part of agency is not raw capability but reliability: getting agents to behave correctly in the messy, stateful digital worlds where they will first do real work.
Any difficulty of real work lives in the state of the environment, not the complexity of the instruction. So rather than authoring ever-harder tasks, we build the world: recovering executable environments from real systems, then letting agents act, fail, adapt, and be measured inside them. This work draws together multimodal perception, computer-use and tool-use agents, RL and world-models.
I earned my Ph.D. at Mila / Université de Montréal under Prof. Aaron Courville, with time as a Research Scientist Intern at Google DeepMind. Earlier, I studied Computer Science at IIT Delhi.
News
Recent releases, papers, and open-source milestones.
- 2026/07Last week, EnterpriseOps-Gym reached a bigger stage: Artificial Analysis launched the EnterpriseOpsGym-AA leaderboard.
- 2026/06EnterpriseOps-Gym accepted to ICML 2026 — work on stateful enterprise agents, long-horizon planning, and realistic tool-use environments.
- 2026/06Open-source enterprise AI artifacts crossed half a million downloads across GroundCUA, BigDocs, UI-Vision, and EnterpriseOps-Gym.
- 2026/02GroundCUA released for grounding computer-use agents on human demonstrations.
Key questions
The first-principles problems behind generally capable digital agents.
Highlighted work
Selected projects and papers. See Google Scholar for the full publication list.
Selected research COLM 2026
SynthDocBench: A Fully Synthetic, Controlled Benchmark for Long-Context Visual Document Understanding. COLM 2026. Paper, Page
We introduce a fully synthetic, controlled benchmark for long-context visual document understanding that varies document length, layout structure, modality composition, and question type as independent axes, enabling systematic diagnosis of VLM failure modes.
Selected research ICML 2026
EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings. ICML 2026. Paper, Page
We study whether agents are ready for real enterprise work.
Reliable automation is still far from solved. We released our largest, most realistic environment to evaluate long-horizon agent planning tool execution under real enterprise constraints.
Selected research
Terminal Agents Suffice for Enterprise Automation. under review. Paper
Terminal-based coding agents with direct API access match or outperform complex MCP and GUI agents, proving that strong foundation models need only simple programmatic interfaces for enterprise automation. Terminal agents achieve the same accuracy as web agents but are 5 to 6x cheaper and outperform MCP agents while staying cost-neutral.
Selected research
Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics. Paper
This work studies whether enterprise agents need learned world models and shows why contextual information is critical for accurately inferring system dynamics in long-horizon digital workflows.
Selected research ICLR 2026
Grounding Computer Use Agents on Human Demonstrations. ICLR 2026. Paper
Building reliable computer-use agents requires accurately connecting natural language instructions to the correct on-screen elements. We introduce GroundCUA, large-scale dataset & GroundNext family of models.
Selected research NeurIPS 2025
AlignVLM: Bridging Vision and Language Latent Spaces for Multimodal Understanding. NeurIPS 2025. Paper
A novel approach to bridging vision and language latent spaces for multimodal understanding in VLMs that maps vision features into a weighted average of LLM text embeddings, ensuring they remain in a space that the LLM can effectively interpret.
Selected research
WebMMU: A Multimodal, Multilingual Benchmark for Website Understanding & Code Generation. EMNLP 2025 (Oral). Page
Address a critical gap in AI evals: how well can models understand and build websites. Unlike many recent works, we collected a real-world, challenging dataset with 117 expert annotators across diverse domains. It is multilingual.
Selected research
BigCharts-R1: Enhanced Chart Reasoning with Visual Reinforcement Finetuning. CoLM 2025. Page
Chart comprehension is crucial for effective human decision-making, yet current VLMs struggle with this task due to limitations in training data and methodologies. We introduce BigCharts-R1, a state-of-the-art chart reasoning model, alongside a novel dataset and training framework.
Selected research ICML 2025
UI-Vision: A Desktop-centric GUI Benchmark for Visual Perception and Interaction. ICML 2025. Page
first comprehensive, license-permissive benchmark for offline, fine-grained evaluation of CUA agents in real-world desktops. We provides: (i) dense, high-quality annotations of human demonstrations, including bounding boxes, UI labels, and action trajectories across 83 software applications.
Selected research
BigDocs: An Open Dataset for Training Multimodal Models on Document and Code Tasks. ICLR 2025. Page
BigDocs is a multimodal dataset effort for advanced document understanding, consisting of two key components: BigDocs-7.5M: A high-quality, open-access, large-scale dataset of 7.5 million multimodal documents spanning 30 tasks BigDocs-Bench: A benchmark suite with 10 real-world-inspired tasks like reasoning over graphical user interfaces (GUI).
Selected research
GenRL: MULTIMODAL FOUNDATION WORLD MODELS FOR GENERALIST EMBODIED AGENTS. NeurIPS 2024
GenRL allows one to specify tasks through vision and/or language prompts, ground them in the embodied domain’s dynamics, and learns the corresponding behaviors in imagination. This exhibits strong multi-task generalization in locomotion and manipulation domains. Read more.
Selected research NeurIPS 2025
Rendering-Aware Reinforcement Learning for Vector Graphics Generation. NeurIPS 2025
RL way to enhances SVG generation in autoregressive VLMs by leveraging feedback from rendered SVG outputs. SVG roll-outs are rendered and compared to original image to compute a reward. This visual fidelity feedback guides model to produce more semantically coherent SVGs.
Selected research
StarVector: Generating Scalable Vector Graphics Code From Images And Text CVPR 2025
GenRL allows one to specify tasks through vision and/or language prompts, ground them in the embodied domain’s dynamics, and learns the corresponding behaviors in imagination. This exhibits strong multi-task generalization in locomotion and manipulation domains. Read more.
Selected research
REPRESENTING POSITIONAL INFORMATION IN GENERATIVE WORLD MODELS FOR OBJECT MANIPULATION
We introduce a general approach that empowers world model-based agents to effectively solve object-positioning tasks. We propose two declinations of this approach for generative world models: position-conditioned (PCP) and latent-conditioned (LCP) policy learning Read more.
Selected research
EQUIVARIANT ADAPTATION OF LARGE PRETRAINED MODELS. NeurIPS 2023
Equivariant networks are specifically designed to ensure consistent behavior with respect to a set of input transformations, leading to higher sample efficiency and more accurate and robust predictions. Read more
Selected research
THE UNSOLVED CHALLENGES OF LLMS AS GENERALIST WEB AGENTS: A CASE STUDY
In this work, we investigate the challenges associated with developing goal-driven AI agents capable of performing novel tasks in a web environment using zero-shot learning. Our primary focus is on harnessing the capabilities of large language models (LLMs) as generalist web agents interacting with HTML-based user interfaces (UIs). Read more
Selected research
MASTERING UNSUPERVISED REINFORCEMENT LEARNING FROM PIXELS. ICML 2023 (Oral)
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 finetuning strategy combined with a new proposed hybrid planner, Dyna-MPC, to adapt the agent for downstream tasks. Read more
Selected research
CHOREOGRAPHER: LEARNING & ADAPTING SKILLS IN IMAGINATION. ICLR 2024 (Spotlight)
a model-based agent that exploits its world model to learn and adapt skills in imagination. We decouple the exploration and skill learning processes, being able to discover skills in the latent state space. During adaptation, the agent uses a meta-controller to evaluate and adapt the learned skills efficiently by deploying them in parallel in imagination. Read more
Selected research
EFFICIENT DYNAMICS MODELING IN INTERACTIVE ENVIRONMENTS WITH KOOPMAN THEORY, ICLR 2024
We approach this problem from the lens of Koopman theory, where the
nonlinear dynamics of the environment can be linearized in a high-dimensional latent space. This allows us to efficiently parallelize the sequential problem of long-range prediction using convolution while accounting for the agent’s action at every time step. Read more
Selected research
Multi-label Iterated Learning for Image Classification with Label Ambiguity, CVPR 2022
Transfer learning from large-scale pre-trained models has become essential for many computer vision tasks. Recent studies have shown that datasets like ImageNet are weakly labeled since images with multiple object classes present are assigned a single label. Inspired by language emergence literature, we propose MILe to incorporate the inductive biases of multi-label learning from single labels using iterated learning. Read more
Selected research
Touch based Curiosity for Sparse-Reward Tasks, CoRL 2023
we leverage surprise from mismatches in touch feedback to guide exploration in hard sparse-reward reinforcement learning tasks. Our approach, Touch-based Curiosity (ToC), learns what visible objects interactions are supposed to “feel” like. We encourage exploration by rewarding interactions where the experience do not match. We test on a range of touch-intensive robot tasks (e.g. pushing objects, opening doors)
Collaboration: I am excited to work with students, interns, and collaborators interested in enterprise agents, terminal/computer-use agents, digital-world simulations, and world models for reliable intelligence.