AI-aided FEC code design and decoding – AI4CODE

How can we learn to design better codes?

Traditional expert-based design methods offer limited guarantees regarding the optimality of the codes they produce. Typically, designing a good error-correcting code involves two steps:

1) Constructing a family of candidate codes adapted to a given channel model and decoder, and

2) Selecting, within that family, a code of the desired length that achieves the best trade-off between performance and complexity.

Each of these steps involves searching within an extremely high-dimensional space, where exhaustive exploration is impossible. As a result, heuristic optimization and partial search techniques are commonly used, but they provide little assurance that the resulting codes are anywhere near optimal. Another source of uncertainty lies in the fact that the parameters guiding these heuristic searches are often poor proxies for the actual performance of the decoder. Consequently, expert-driven design methods may fail to produce codes that meet the stringent requirements of modern communication or storage applications, simply due to the lack of suitable heuristics. Machine learning offers a new way forward. It can improve the alignment between a code and its target decoder by identifying and optimizing code attributes that better capture true decoder performance than conventional metrics. Learning-based methods can also make the design process more efficient by enabling a faster and more intelligent exploration of the vast space of possible codes.

How can we learn to decode better?

In theory, the optimal decoder that minimizes the probability of decision error is well known. However, its exponential computational complexity makes it impractical for real-world systems. In practice, approximate decoders are used instead, trading off performance for computational feasibility. Unfortunately, their performance typically degrades as their implementation becomes simpler. The central question is therefore whether learning can help bridge this performance gap at a manageable computational cost — by discovering structures or strategies that human designers have not yet envisioned. The goal is to develop smarter decoders: either by augmenting existing algorithms with learned components, or by replacing them altogether with AI-based models trained to correct errors directly.

What can we learn from the machine?

The codes and models discovered by learning-based approaches will, in turn, be analyzed to understand why and how they perform. This dual objective is to (i) refine our expert algorithms based on these insights, and (ii) extract general lessons about how to best adapt machine learning architectures and practices to the specific challenges of channel coding.

Key Results

Among the most significant outcomes of the project, the use of machine learning — in various forms — has enabled us to:

- Discover a new modulation strategy for codewords, reducing the residual error rate after decoding by a factor of 1000 in a specific use case.

- Identify new families of cyclic spreading sequences that are robust to truncation and ideally suited for the transmission of short IoT packets. We have proven that no better sequences exist in this class, and they have been patented.

- Design new polar codes (used on the 5G control channel) that preserve strong performance even under severe constraints on decoder computational capacity.

- Outperform state-of-the-art expert algorithms by identifying unreliable bits more efficiently in the decoding of LDPC codes (used for data transmission in 5G), thereby increasing the likelihood of successful re-decoding attempts.

- Train generic decoder models capable of approximating the performance of the optimal decoder with remarkable precision — an important first step toward the long-term goal of developing universal, code-agnostic decoders.

- Improve multi-antenna (MIMO) demodulation for wireless communications (e.g., WiFi or 5G), approaching the performance of the theoretically optimal, yet computationally infeasible, detector.

Key Insights and Lessons Learned

The project also provided several broader insights:

- Machine learning can uncover solutions and strategies beyond expert intuition, especially when allowed to operate autonomously through unsupervised or reinforcement learning.

- Existing generic decoder models do not scale well; their architectures need to be reconsidered, and greater attention must be paid to the selection and quality of training data.

- Fully leveraging deep learning requires embracing best practices from this still highly experimental discipline.

- Some of these best practices themselves may inspire new ideas for channel coding theory and methods.

- For the decoding problem, we could not yet determine whether it is preferable to augment expert algorithms with an AI layer or to replace them entirely with learned models. The former approach yields simple and interpretable solutions but remains limited by the original algorithm’s structure; the latter offers greater potential but leads to complex, less interpretable models.

After four years of research on this topic, our main conclusion is that deep learning still holds tremendous potential for channel coding research. While it remains uncertain whether it will ever become a practically viable solution for real-world receivers, AI models can already serve as powerful tools for probing the theoretical limits of what is achievable in challenging problems of coding, information theory, and signal processing.

Regarding the use of AI for code construction, it is striking that generative AI methods remain largely underexplored, even though they appear to be ideally suited for this purpose. Given the spectacular advances in this field over the past three years—particularly in discrete diffusion models—there is a vast and promising area of research to be explored. The main challenge will be to design models that are sufficiently creative to discover codes fundamentally different from those on which they were trained.

As emphasized in our overall assessment, one key conclusion of the project is that the architectures that made AI successful in computer vision (convolutional networks) and natural language processing (recurrent networks and transformers) do not perform well enough when applied directly to error-correction decoding. They do not scale: increasing model size or dataset size does not improve performance. The architecture that will make AI the game changer for decoding still remains to be found.

Channel coding relies on algorithms that are, for the most part, heuristic or approximations of intractable optimal solutions. The AI4CODE project explored ways to compensate for these limitations, either by augmenting existing algorithms with learning modules or by replacing them entirely with AI-based models. However, in many cases, neither approach proved fully satisfactory. Meanwhile, large language models (LLMs) have made extraordinary progress. We now have models capable of advanced reasoning, complex mathematical problem solving, and even programming. We believe these tools could profoundly transform the way scientific research is conducted. This vision forms the foundation of our new proposal, submitted to the AAPG 2026 call. The project aims to discover new algorithms through program synthesis guided by generative AI (LLMs), in the spirit of Google DeepMind’s AlphaEvolve system. The advantage of program synthesis lies in its production of explicit, interpretable, and verifiable code. In this approach, AI acts as a guide, introducing creativity into the code-generation process. We believe that such a methodology could accelerate the emergence of new ideas and solutions in disciplines such as channel coding, where algorithms play such a central role.

The AI4CODE project brings together 6 research team with strong expertise in the design, decoding and standardization of forward-error-correction codes. The aim is to develop skills in artificial intelligence and machine learning, and to explore how learning techniques can contribute to the improvement of code design methods (by using less parameters, more relevant heuristics, producing stronger codes) and decoders (better performance, reduced complexity or energy consumption), on selected scenarios of practical interest for which a full theoretical understanding is still lacking. The proposed methodology is to augment legacy design methods and decoders with learning capabilities or decision support systems wherever relevant, rather than replacing them by a generic, black-box neural network, so that we can inspect the trained solutions and try to infer why they work better. Our ultimate goal is to obtain new theoretical hindsight that could translate into better codes and decoders.