Qualcomm · Filed Jan 28, 2025 · Published Jul 30, 2026 · verified — real USPTO data

Qualcomm Patents a Negotiation System for Phones Training AI Together Over a Network

When dozens of phones train an AI model together without sharing raw data, the math can break down fast if every device compresses its updates differently. Qualcomm's new patent proposes a simple handshake to fix that.

Qualcomm Patent: Gradient Scaling in Federated Learning — figure from US 2026/0222312 A1
Figure from the official USPTO publication.
See all 10 drawings from this filing ↓
Publication number US 2026/0222312 A1
Applicant QUALCOMM Incorporated
Filing date Jan 28, 2025
Publication date Jul 30, 2026
Inventors Amit MOSES, Ronen SHAKED, Gideon Shlomo KUTZ, Yuval BEN HUR, Evgeny LEVITAN, Ariel Yaakov SAGI, Reuven ALPERT, Aviv REGEV
CPC classification 709/224
Grant likelihood Medium
Examiner KHANAL, SANDARVA (Art Unit 2453)
Status Docketed New Case - Ready for Examination (Mar 19, 2025)
Document 20 claims

What Qualcomm's federated learning compression fix actually does

Imagine a school project where ten students each do part of the research and send their notes to a teacher, who combines everything into a final report. Now imagine each student decides independently how much to abbreviate their notes. Some write full sentences; others use shorthand so compressed it's unreadable. The teacher ends up with a mess.

That's roughly the problem Qualcomm is solving here. In federated learning, your phone (and many others) trains an AI model locally on its own data, then sends just the training updates, not your actual data, to a central server. But those updates need to be compressed to fit over a wireless connection, and if each phone picks its own compression level, the server can't combine them reliably.

Qualcomm's patent describes a process where your device proposes a compression level, the network responds with the one it actually wants, and then your device sends the update using that agreed-upon level. It's a short back-and-forth that keeps everyone in sync without the server having to guess.

How the device and network agree on a scaling factor

The patent describes a negotiation protocol between a user device (like a phone) and a network node (like a base station or server) for coordinating gradient scaling. A gradient, in AI training terms, is a set of numbers that describes how a model's internal weights should change after processing some data. Gradients can be large, and sending them over a wireless link requires compression.

The scaling factor is essentially a multiplier that controls how aggressively those gradients are compressed before transmission. Too much compression and the training signal gets distorted; too little and it's too big to send efficiently.

The protocol works in three steps:

  • The device identifies a scaling factor it thinks is appropriate and sends that proposal to the network.
  • The network replies with its own preferred scaling factor, which may differ based on what other devices are sending or what the model currently needs.
  • The device then sends its gradient update using the network's chosen value.

This back-and-forth (called a two-step negotiation) means the server can normalize gradient updates across many devices, which is important when aggregating contributions from phones with different hardware, connection quality, or local data distributions.

What this means for on-device AI training over 5G

Federated learning is the architecture behind privacy-preserving AI, where a model improves from data on millions of devices without that data ever leaving those devices. Getting it to work reliably over real wireless networks is genuinely hard, and poorly coordinated gradient compression is one of the less-discussed reasons it fails in practice.

For Qualcomm, this is squarely in their lane: they make the chips and modems that handle this exact communication layer in Android phones and network infrastructure. A patent here could give them a standards-level advantage as 5G networks start formally supporting AI workloads. For users, the payoff would be AI features that improve over time without your private data leaving your phone.

Editorial take

This is a narrow but real engineering problem, and Qualcomm is in a better position than almost anyone to solve it given they build both the device-side and network-side silicon. It's not a flashy consumer-facing invention, but it's the kind of plumbing that has to work before federated learning becomes a mainstream feature in phones.

The drawings

10 drawing sheets from US 2026/0222312 A1 · click any drawing to enlarge

Patent filing page

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.