ANR research project · 2026–2030
Letting AI evolve the algorithms of tomorrow’s networks
Televolve combines large language models with mechanisms inspired by natural evolution to discover new, interpretable algorithms for telecommunications.
ANR ANR-26-CE25-3930
Oct 2026 – Sep 2030
Coordinated by IMT Atlantique · Lab-STICC
🧬 Generation 0. This temporary page was evolved with the help of AI. The humans have been notified and will soon step in to select, mutate and add their grain of salt. Stay tuned!
Context
Faster, more reliable, more energy-efficient networks
5G, 6G, the Internet of Things: every new generation of networks raises the bar. At their core lie countless algorithms, most of them hand-crafted, often closer to expert heuristics than to optimal solutions, which are either unknown or too complex to compute.
Machine learning offers new tools, but neural networks remain black boxes that are hard to analyse and deploy. In 2025, Google DeepMind showed with AlphaEvolve that a large language model coupled with an evolutionary search can discover better algorithms by writing and refining code itself.
Televolve brings this approach to telecommunications, a field full of hard, well-defined algorithmic problems.
The result: high-performing algorithms that remain readable, and therefore interpretable and verifiable by humans.
Related open initiatives: ShinkaEvolve (Sakana AI), The AI Telco Engineer (NVIDIA).
Approach
Darwin, applied to code
A language model writes programs, an evaluator tests them, and only the best are kept and modified further.
01
Seed
Start from a simple program that gives a first answer to the problem.
02
Mutate
An LLM proposes modified versions of the code: the “mutations”.
03
Evaluate
Each candidate is automatically scored on its performance.
04
Select
Only the best survive and give rise to the next generation.
05
Repeat
The loop runs over many generations, until progress stalls or the budget runs out.
What Televolve adds
Self-steered exploration
The agent adapts its search strategy and breaks complex problems into steps.
Expert knowledge
It draws on a knowledge library curated by researchers.
Memory
It keeps track of past attempts, successful or not, to guide new ones.
Human in the loop
Researchers set the goals, steer the search and validate the results.
Objectives
AI as an accelerator of scientific discovery for telecom
Push the state of the art
Show on concrete open problems (error-correcting codes, radio resource allocation, information theory) that this approach can improve on the best known solutions.
Build an open-source agent
An agent that runs locally on open, reasonably sized models, draws on expert knowledge and learns from its own trial and error.
Assess the approach
A rigorous evaluation of its strengths and limits: what can be achieved with academic resources, compared with the frontier models of major AI labs?
Expected outcomes
New algorithms
Better algorithms for future communication systems, delivered as readable, verifiable code.
Open agentic system
Built on local AI models and released open-source, so that any lab can run it.
Open vs. frontier
A comparison of proprietary and open AI models, in support of digital sovereignty and independent research.
Consortium
Project partners
Coordinator
Lab-STICC
IMT Atlantique · Brest
Coordinates the project, leads the development of the AI agent and the work on error-correcting codes.
Partner
CEA-Leti
Grenoble
Expertise in wireless communications and experience with radio demonstrators.
Partner
ISAE-SUPAERO
Toulouse
Information theory and evolutionary algorithms; leads the evaluation of the approach.
Funding
ANR-26-CE25-3930
Televolve, Leveraging large language models and evolutionary program search to design better algorithms for communication systems and networks, is funded by the French National Research Agency (Agence Nationale de la Recherche, ANR) from October 2026 to September 2030.