Qualcomm Innovation Fellowship Europe: 5 PhD Students Win 0,000 Each for AI and Cybersecurity Research

What is the Qualcomm Innovation Fellowship Europe?

The Qualcomm Innovation Fellowship (QIF) runs in three regions — Europe, India, and the United States — and is now in its 17th year. The Europe edition specifically targets PhD students working in artificial intelligence and cybersecurity, fields Qualcomm considers central to the next generation of its Snapdragon platforms and edge-AI ambitions.

Each winner receives $40,000 in unrestricted research funding — at current exchange rates, that is roughly €35,200 — plus a dedicated mentor from Qualcomm Technologies' research division. The selection process is genuinely competitive: this year, five fellows emerged from a shortlist of 18 finalists drawn from thirteen European research institutions, according to the announcement.

Michael Hofmann, Senior Director of Engineering at Qualcomm Technologies Netherlands B.V., noted that submissions climbed to "an all-time high this year — close to 50 percent above last year's total." The breadth of proposals impressed the committee: topics ranged from multimodal generation and trustworthy agents to edge AI, secure systems, and machine learning for scientific discovery.

The five winners and their research

Here is a breakdown of what each fellow proposed — and why it matters.

Mar González I Català — University of Cambridge

Project: "A Geometric Theory of Autoregressive Reasoning"

González I Català tackles a problem anyone who has watched a reasoning model "think out loud" will recognise: chain-of-thought traces often look convincing but can derail silently. She proposes modelling autoregressive reasoning as a stochastic dynamical system over latent reasoning states, where chain-of-thought traces are interpreted as trajectories terminating in answer "attractors." The idea is that reasoning quality can be judged not just by the final answer but by the geometry of the trajectory that leads there — opening the door to training objectives that reward desirable trajectory shapes and inference-time interventions that redirect failing reasoning paths.

Jiajun He — University of Cambridge

Project: "Robust Diffusion Model Control for Multiple-Constraint Inverse Problems with Replica Exchange"

Diffusion models are everywhere — images, molecules, proteins — but real applications increasingly demand test-time control under multiple evolving constraints. Existing methods (guidance, sequential Monte Carlo, search) can become biased or collapse in diversity under strong constraints. He proposes a replica-exchange framework that exchanges and reweights trajectories to avoid unreliable marginal-density approximations. The approach also includes online adaptation that can diagnose poor mixing and refine the replica ladder as constraints evolve — aiming for a plug-and-play control procedure usable across generation, simulation, and scientific discovery.

Abhinandan Pal — University of Birmingham

Project: "Neural Model Checking for Hardware Verification"

Before a chip is manufactured, engineers must formally prove that unwanted behaviour never occurs. As circuits grow more complex — and as AI tools begin generating hardware-design code — this verification bottleneck is worsening. Pal's Neural Model Checking (NMC) learns a small neural network from sample executions as a candidate correctness proof, then uses a mathematical solver to check that proof against every possible circuit behaviour. On SystemVerilog benchmarks, NMC is already substantially faster than leading automated verification tools. The fellowship will fund scaling NMC to industry-relevant designs, building a test suite of realistic hardware, and measuring how far automated solver-checked proofs can go in practice.

Naila Sebastián Esandi — INRIA / ENSAE Paris

Project: "A Theoretical Framework for Zero-Shot Reinforcement Learning"

Reinforcement learning solves control problems well, but it requires solving a planning problem for each objective independently — computationally unfeasible for endpoint devices like autonomous vehicles, surgical robots, or mobile phones. Zero-Shot RL (ZSRL) has emerged as a candidate to satisfy these hardware constraints, but it lacks rigorous theoretical foundations. Sebastián Esandi aims to establish a formal framework for ZSRL and develop a new class of theoretically grounded, efficient algorithms — research that could directly influence how Qualcomm bakes AI into Snapdragon-powered edge devices.

Christopher Wewer — Max Planck Institute for Informatics

Project: "Towards Scene State Tokenization for World Models"

Models like Genie 3 and LingBot-World generate impressive videos, but they fall short as world models: operating on video frames means the same environment can behave differently under two actions, long-horizon simulation requires expensive history caching, and inference cost is fixed regardless of scene complexity. Wewer's core argument: the bottleneck is representation, not architecture. He proposes learning compact latent scene states that describe physical configurations — geometry, object identity, predictive attributes — and building state-space world models that update a single world state over time instead of accumulating video frames. This builds on his prior work on compressed 3D scene tokens (SceneTok) and generative 3D reconstruction.

The European AI research ecosystem in microcosm

The list of finalist institutions reads like a map of European AI research strength: CISPA Helmholtz Center, ETH Zurich, INRIA, KU Leuven, three Max Planck institutes, TU Munich, University of Amsterdam, Birmingham, Cambridge, Oxford, and Tübingen. That is seven EU member-state institutions plus three in the UK and two in Switzerland — a reminder that European AI talent remains concentrated in a handful of well-funded clusters.

For context, $40,000 (€35,200) is meaningful but modest compared with other European funding instruments. An ERC Starting Grant provides up to €1.5 million over five years; a Marie Skłodowska-Curie postdoctoral fellowship runs around €200,000 for two years. Where QIF distinguishes itself is the industry mentorship — direct access to Qualcomm engineers working on production Snapdragon silicon, edge AI deployment, and on-device security — something no academic grant provides.

The 50 % surge in submissions year-on-year is worth dwelling on. It mirrors what we see across the AI field more broadly: the number of people trying to push beyond scaling language models into areas like world models, formal verification, and reasoning geometry is expanding fast. Qualcomm's selection suggests the company sees these as strategic research directions — unsurprising for a chipmaker that needs AI to run efficiently on battery-powered devices, not just in data centres.

What this means for European AI

Industry-funded PhD fellowships are not new, but QIF Europe's longevity (17 editions) and specific focus on AI and cybersecurity make it one of the more established programmes connecting European academic research with semiconductor-industry priorities. For the winners, the real prize is arguably not the $40,000 cheque but the mentorship pipeline into Qualcomm's research organisation — a foot in the door at one of the world's largest fabless chip designers, whose Snapdragon processors ship in hundreds of millions of devices annually.

For European readers wondering about practical impact: the hardware verification work (Pal), zero-shot RL for edge devices (Sebastián Esandi), and scene-state tokenization for world models (Wewer) all have direct relevance to on-device AI — the kind that runs locally on a phone or laptop rather than in the cloud. With the EU AI Act now in force and GDPR governing data processing, on-device inference is increasingly attractive from both a privacy and regulatory standpoint. Research that makes local AI more capable feeds directly into that European regulatory reality.