Arash Behboodi

I am a research scientist and Director of Engineering at Qualcomm AI Research, working on the theoretical foundations of information and computational systems.

My interests include information theory, learning theory, machine learning for inverse problems, compressed sensing, geometric deep learning, differentiable simulation, and agentic AI.

Information Theory Learning Theory Inverse Problems

View Publications Read Bio Visit Blog

News and Highlights

Recent papers, talks, and research updates.

New paper on our NeurIPS 2025 demo: Efficient Reasoning on the Edge and Hugging Face page
New paper: LaneRoPE: Positional Encoding for Collaborative Parallel Reasoning and Generation
New paper: Reasoning as Compression: Unifying Budget Forcing via the Conditional Information Bottleneck
New paper: LUMINA: Long-horizon Understanding for Multi-turn Interactive Agents
See our NeurIPS 2025 expo demo on: Efficient Reasoning on the Edge.
New paper: Fundamental bounds on efficiency-confidence trade-off for transductive conformal prediction
I give a keynote talk about Information Theoretic Perspective on Conformal Prediction at 14th Symposium on Conformal and Probabilistic Prediction with Applications, COPA 2025
Accepted paper at Globecom 2025: ReQuestNet: A Foundational Learning model for Channel Estimation
Accepted paper at ACL 2025: Local Look-Ahead Guidance via Verifier-in-the-Loop for Automated Theorem Proving, also presented at ICLR 2025 Workshop on Reasoning and Planning for Large Language Models
Accepted paper at ICLR 2025: Multi-Draft Speculative Sampling: Canonical Architectures and Theoretical Limits (Spotlight)
Accepted paper at ICLR 2025: Probabilistic and Differentiable Wireless Simulation with Geometric Transformers
Accepted paper at TMLR: On the Sample Complexity of One Hidden Layer Networks with Equivariance, Locality and Weight Sharing
Accepted paper at TMLR: Neural Lattice Reduction: A Self-Supervised Geometric Deep Learning Approach

Selected Recent Papers

A quick tour through recent work across theory, reasoning, and simulation.

Yelysei Bondarenko et al., Efficient Reasoning on the Edge, 2026
Gabriele Cesa, Thomas Hehn, Aleix Torres-Camps, Àlex Batlle Casellas, Jordi Ros-Giralt, Arash Behboodi, Tribhuvanesh Orekondy, LaneRoPE: Positional Encoding for Collaborative Parallel Reasoning and Generation, Preprint 2026
Fabio Valerio Massoli, Andrey Kuzmin, Arash Behboodi, Reasoning as Compression: Unifying Budget Forcing via the Conditional Information Bottleneck, Preprint 2026
Amin Rakhsha, Thomas Hehn, Pietro Mazzaglia, Fabio Valerio Massoli, Arash Behboodi, Tribhuvanesh Orekondy, LUMINA: Long-horizon Understanding for Multi-turn Interactive Agents, Preprint 2026
Arash Behboodi, Alvaro H.C. Correia, Fabio Valerio Massoli, Christos Louizos, Fundamental bounds on efficiency-confidence trade-off for transductive conformal prediction, Preprint, 2025
Julian Suk, Thomas Hehn, Arash Behboodi, Gabriele Cesa, ViNE-GATr: scaling geometric algebra transformers with virtual nodes embeddings, ICLR 2025 Workshop on Machine Learning Multiscale Processes (Best Poster Award)
Sara Rajaee, Kumar Pratik, Gabriele Cesa, Arash Behboodi, Local Look-Ahead Guidance via Verifier-in-the-Loop for Automated Theorem Proving, ACL 2025, and ICLR 2025 Workshop on Reasoning and Planning for Large Language Models
Ashish Khisti, M.Reza Ebrahimi, Hassan Dbouk, Arash Behboodi, Roland Memisevic, Christos Louizos, Multi-Draft Speculative Sampling: Canonical Architectures and Theoretical Limits, ICLR 2025 (Spotlight), also appeared at ICML 2024 Workshop on Theoretical Foundations of Foundation Models
Thomas Hehn, Markus Peschl, Tribhuvanesh Orekondy, Arash Behboodi, Johann Brehmer, Probabilistic and Differentiable Wireless Simulation with Geometric Transformers, ICLR 2025, also appeared at ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling 2024
Arash Behboodi, Gabriele Cesa, On the Sample Complexity of One Hidden Layer Networks with Equivariance, Locality and Weight Sharing, TMLR, 2025
Giovanni Luca Marchetti, Gabriele Cesa, Kumar Pratik, Arash Behboodi, Neural Lattice Reduction: A Self-Supervised Geometric Deep Learning Approach, TMLR 2025, also appeared in Symmetry and Geometry in Neural Representations - NeurReps Workshop, NeurIPS 2023

Misc.

Workshops, conversations, and related projects.

Workshop D3S3: Data-driven and Differentiable Simulations, Surrogates, and Solvers at NeurIPS 2024.
My chat with Sam Charrington at TWIML AI podcast around Differentiable Simulation, Conformal Prediction, and AI at the Edge.