← Theses

Language Models as Decoders for Very Slow Wireless Links

Open topic · MA · Advisors: Bruno Rodrigues

A language model knows what a message is likely to say. Put that knowledge in the receiver, and a garbled emergency text that no error-correcting code can repair may still be read correctly. But can we trust what the model reads?

One testbed: our Quiet Link prototype, which sends short text messages between phones through sound. Every extra decibel of decoding gain is extra range.
One testbed: our Quiet Link prototype, which sends short text messages between phones through sound. Every extra decibel of decoding gain is extra range.

Why this matters

When networks fail, the links that remain are very slow: sound between phones, LoRa radios, satellite SOS. They carry a few dozen bits per second, so every message is short and every bit is precious. Today, the receiver treats a message as random bits. But human messages are not random: “two people trapped, one injured” is far more likely than a random string of letters. A language model captures exactly that knowledge.

The idea

A language model knows which messages are likely. This thesis uses that knowledge when decoding: the receiver keeps several candidate readings of a damaged frame and ranks them by how well each fits the received signal and how plausible it is as a message [1, 2]. A frame that the error-correcting code alone cannot repair may still decode.

Research questions

  1. The gain. How many decibels does a language prior at the receiver buy on real, slow links, compared with decoding without it? How close does it come to the limit set by how predictable the messages are?
  2. Real channels. How should the prior work with the soft information of a real decoder, under echoes, noise bursts and a moving sender?
  3. Trust. A model that “repairs” an SOS into a plausible but wrong message is dangerous. How can the receiver tell a correct repair from a confident invention, and what false-repair rate can it guarantee?
  4. Offline and embedded. How much of the gain survives when the model shrinks to fit offline devices, from a language model on a laptop, to a small model on a phone, to a character model of a few megabytes on a microcontroller?

What you will do

  • Build the decoder. Add language-model list decoding to our Quiet Link receiver, starting with small models that run on a laptop or phone.
  • Test on real links. Measure over the air with phones in rooms, and on a second slow link such as LoRa, using a corpus of real disaster messages collected after earthquakes and floods.
  • Bound it. Relate the measured gain to the entropy of the messages, which limits how much any prior can help.
  • Make it safe. Design and evaluate a confidence test that keeps false repairs rare, even when the model is sure of itself.
  • Shrink it. Measure the decoding gain against model size and memory, down to a model that runs offline on an ESP32-class board.

Why you will enjoy it

  • Language models meet the physical layer, one of the most active frontiers in communications.
  • Real impact: more range for messages when everything else fails.
  • A running start: the Quiet Link modem, the over-the-air setup and a corpus of disaster messages exist. Background in either machine learning or communications is enough, and you will learn the other.

References

[1] Z. Li, C. Yue, J. Hao, B. Vucetic, Y. Li. LLM-Viterbi: Semantic-Aware Decoding for Convolutional Codes. arXiv:2604.19035, 2026. arxiv.org/abs/2604.19035

[2] Z. Wang, T. Ren, R. Li, Z. Zhao, H. Zhang. In-Context Source and Channel Coding. arXiv:2601.10267, 2026. arxiv.org/abs/2601.10267

Interested? Email the advisors: Bruno Rodrigues.