Qualcomm Innovation Fellowship Europe Rewards Excellent Research in AI and Cybersecurity
Qualcomm Technologies, Inc. today unveiled the latest recipients of the Qualcomm Innovation Fellowship (QIF) Europe, an award now marking its 17th edition. This year's five fellows are Mar Gonzàlez I Català (University of Cambridge), Jiajun He (University of Cambridge), Abhinandan Pal (University of Birmingham), Naila Sebastián Esandi (INRIA / ENSAE Paris), and Christopher Wewer (Max Planck Institute for Informatics).
Held each year across Europe, India, and the United States, QIF exists to spotlight, celebrate, and nurture the engineering PhD students producing the field's boldest work. In Europe, the fellowship turns its focus to promising early-career talent in artificial intelligence and cybersecurity, giving every recipient a $40,000 award alongside one-on-one mentorship from the Qualcomm Technologies research team.
“Submissions climbed to an all-time high this year—close to 50 percent above last year's total—which says a great deal about how fast machine learningAI is moving, and how vital it is to keep our rapidly evolving hardware and software secure,” said Michael Hofmann, Senior Director of Engineering at Qualcomm Technologies Netherlands B.V. “What stood out was the sheer range of the proposals, stretching from multimodal generation, trustworthy agents, world models, robotics, and generative AI safety to hardware verification, secure systems, privacy, communications, edge AI, and machine learning for scientific discovery. We're honored to continue mentoring each of the winners as their research unfolds.”
This year's five fellows emerged from a shortlist of eighteen finalists, drawn from PhD programs at CISPA, ETH Zurich, INRIA, KU Leuven, Max Planck Institute for Informatics, Max Planck Institute for Intelligent Systems, Max Planck Institute for Security and Privacy, Technical University of Munich, University of Amsterdam, University of Birmingham, University of Cambridge, University of Oxford, and University of Tübingen.
After careful review, the following five fellows were selected for their outstanding proposals:
“A Geometric Theory of Autoregressive Reasoning” - Mar Gonzàlez I Català (University of Cambridge)
Reasoning models increasingly expose chain-of-thought traces that make parts of their problem-solving process observable, yet the structure and significance of these traces remain poorly understood. Recent work suggests that entropy dynamics within reasoning traces correlate with correctness and can serve as useful diagnostic or intervention signals. This project asks whether such signals reflect a deeper structure: an underlying geometry of reasoning trajectories that could explain both when models become uncertain and how they converge, stabilize, derail, or recover. To make this hypothesis operational, we propose to model autoregressive reasoning as a stochastic dynamical system over latent reasoning states, with chain-of-thought traces interpreted as trajectories that terminate in answer attractors. From this perspective, reasoning quality can be characterized not only by the final answer produced, but also by the geometry of the trajectory that leads to it. The framework further suggests new forms of intervention: training objectives that encourage desirable trajectory geometries, and inference-time methods that redirect trajectories converging to incorrect answers.
“Robust Diffusion Model Control for Multiple-Constraint Inverse Problems with Replica Exchange” - Jiajun He (University of Cambridge)
Diffusion models are powerful generative priors across images, molecules, proteins, and scientific domains, but real applications increasingly require test-time control under measurements, preferences, and multiple evolving constraints. Existing methods such as guidance, sequential Monte Carlo, and search can become biased, tuning-sensitive, or prone to diversity collapse under strong constraints. This proposal develops a replica-exchange framework for robust diffusion model control. The first key innovation is path-space exchange: by exchanging and reweighting trajectories, the method avoids unreliable marginal-density approximations and remains accurate under imperfect learned or non-diffusive dynamics. The second is online adaptation: the method can diagnose poor mixing, refine the replica ladder, improve proposal dynamics, and continue inference as constraints harden or evolve. Together, these advances aim to deliver a stable, diverse, adaptive, and plug-and-play control procedure, with wide applications in generation, simulation, and scientific discovery.
“Neural Model Checking for Hardware Verification” - Abhinandan Pal (University of Birmingham)
Before hardware is manufactured, engineers must show that unwanted behaviour never occurs, and that required behaviour eventually does. This task is becoming harder as circuits grow more complex, and AI tools begin to generate hardware-design code more rapidly. My research develops Neural Model Checking (NMC), which learns a small neural network from sample executions as a candidate correctness proof. A mathematical solver then checks this proof against every possible behaviour of the circuit. If the solver does not find a violation, it proves the required property rather than merely testing selected examples. On SystemVerilog benchmarks, NMC is substantially faster than leading fully automated verification tools. This fellowship will move NMC from academic promise to industry-relevant hardware verification. The project focuses on building a test suite of realistic hardware designs and correctness requirements, identifying the tasks that defeat existing automated tools, and extending the theory and implementation of NMC to handle real-world designs. The project will deliver a reproducible evaluation pipeline and measure how far automated, solver-checked proofs can scale in practical hardware development. Its goal is to automate verification tasks that currently require extensive manual proof effort or rely on testing that may miss critical corner cases.
“A Theoretical Framework for Zero-Shot Reinforcement Learning” - Naila Sebastián Esandi (INRIA / ENSAE Paris)
While Reinforcement Learning (RL) has successfully addressed complex control problems without the need for high-fidelity world models, a significant bottleneck remains: the requirement to solve a planning problem for each objective independently. This redundancy is computationally unfeasible for endpoint devices, such as autonomous vehicles, surgical robots, or mobile phones. Zero-Shot RL (ZSRL) has emerged as a candidate paradigm to satisfy these hardware constraints; however, its progress is currently restricted by a lack of rigorous theoretical foundations and unified evaluation methods. This project aims to bridge this gap by establishing a formal framework for ZSRL and developing a new class of theoretically grounded, efficient algorithms.
“Towards Scene State Tokenization for World Models” - Christopher Wewer (Max Planck Institute for Informatics)
Recent models like Genie 3 or LingBot-World are impressive video generators, but they fall short as world models for three reasons. Since they operate on video frames, (1) the same environment is not guaranteed to behave the same way under two different actions, (2) long-horizon simulation forces expensive history caching, and (3) inference cost is fixed regardless of how complex the scene actually is. In this project, I argue that the main bottleneck is the representation, not the architecture. Videos are only observations of a world, not the world itself. I propose to learn compact latent scene states that describe the physical configurations of environments, including geometry, object identity, and the attributes needed to predict what happens next. On top of these states, I introduce state-space world models that update a single world state over time, rather than accumulating a growing history of video frames. Observations are then decoded from these states. I will outline how such latent states can be learned: first by training an encoder-decoder on complete synthetic scenes, then by treating the encoder as a conditional generative model that infers full states from partial observations, and finally by bridging the gap to real video data. This builds directly on my prior work on compressed 3D scene tokens (SceneTok), generative 3D reconstruction, and spatial reasoning with diffusion models.
About Qualcomm
Qualcomm is a global computing leader at the center of the AI era, enabling intelligence to scale from the most personal devices to large‑scale infrastructure. Building on more than four decades of innovation, we develop platforms and solutions that bring together advanced AI, high‑performance, low power computing and industry‑leading connectivity—powering products and services used around the world. At Qualcomm, we are engineering human progress.
Qualcomm Incorporated includes our licensing business, QTL, and the vast majority of our patent portfolio. Qualcomm Technologies, Inc., a subsidiary of Qualcomm Incorporated, operates, along with its subsidiaries, substantially all of our engineering and research and development functions and substantially all of our products and services businesses, including our QCT semiconductor business. Snapdragon and Qualcomm branded products are products of Qualcomm Technologies, Inc. and/or its subsidiaries. Qualcomm patents are licensed by Qualcomm Incorporated. Qualcomm, Snapdragon, Qualcomm Dragonwing and Qualcomm Dragonfly are trademarks or registered trademarks of Qualcomm Incorporated.
