Semantic Pilot Design for Data-Aided Channel Estimation Using a Large Language Model

Sojeong Park, Hyun Jong YangarXiv 2026 · DOI · arXiv

Abstract

This paper proposes a semantic pilot design for data-aided channel estimation in text-inclusive data transmission, using a large language model (LLM). In this scenario, channel impairments often appear as typographical errors in the decoded text, which can be corrected using an LLM. The proposed method compares the initially decoded text with the LLM-corrected version to identify reliable decoded symbols. A set of selected symbols, referred to as a semantic pilot, is used as an additional pilot for data-aided channel estimation. To the best of our knowledge, this work is the first to leverage semantic information for reliable symbol selection. Simulation results demonstrate that the proposed scheme outperforms conventional pilot-only estimation, achieving lower normalized mean squared error and phase error of the estimated channel, as well as reduced bit error rate.

Introduction

Accurate Channel State Information (CSI) is fundamental to reliable wireless communication. Traditionally, transmitters send known "pilot" sequences to estimate the channel. However, increasing pilot length to improve accuracy introduces significant bandwidth overhead. Data-aided channel estimation attempts to mitigate this by using decoded data symbols as additional pilots. The primary challenge is that errors in the initial decoding can lead to a negative feedback loop, degrading the channel estimate. This paper introduces a "semantic pilot" design that leverages the contextual understanding of Large Language Models (LLMs) to identify reliable symbols for data-aided estimation, bridging the gap between physical layer processing and application layer semantics.

Proposed Framework

The proposed system architecture integrates an LLM into the base station's receiver chain to refine channel estimation through a multi-step process.

Proposed system architecture for LLM-aided semantic pilot design and channel estimation.

As illustrated in Figure 1, the workflow follows these stages:

  1. Initial Estimation and Decoding: The base station uses a standard, short pilot sequence to obtain an initial Least Squares (LS) channel estimate and decodes the received text payload. Due to noise, this text typically contains typographical errors.
  2. LLM-Based Correction: The decoded text is processed by an LLM (e.g., OpenAI's o4-mini). The model is prompted to correct typos while strictly maintaining the original string length and character positions.
  3. Semantic Pilot Selection: The system compares the initially decoded text with the LLM-corrected version. Characters that remain unchanged are deemed highly reliable. The modulated symbols corresponding to these characters are grouped to form the "semantic pilot."
  4. Two-Step Channel Refinement:
  5. Phase Refinement: The system recalculates the channel phase using both the original pilot and the new semantic pilot.
  6. Magnitude Scaling: Magnitude estimation is performed using the entire decoded payload to average out zero-mean noise, correcting remaining magnitude bias.

Experimental Methodology

The study utilized the Europarl dataset for source text, encoded with 6-bit fixed-length coding and QPSK modulation. Transmission was simulated over a Single-Input Single-Output (SISO) Rician fading channel. OpenAI’s o4-mini model served as the LLM for text correction, tailored via specific prompt engineering to ensure structural consistency.

Results and Discussion

The effectiveness of the semantic pilot approach is evaluated through Normalized Mean Squared Error (NMSE) and phase error metrics across varying Signal-to-Noise Ratios (SNR).

Performance comparison of the proposed method against pilot-only, LLM-corrected, and standard decoded data-aided methods: (a) NMSE vs SNR, (b) Phase Error vs SNR.

As shown in Figure 6, the proposed method consistently achieves the lowest NMSE and phase error compared to conventional pilot-only and standard data-aided methods. Key performance highlights include:

  • High Reliability: At an SNR of 10 dB, the semantic pilot symbols achieved a reliability rate of 99.94%.
  • High Pilot Yield: At 10 dB SNR, the method identified 96.37% of the payload as reliable pilots, effectively turning the message into a pilot sequence with zero bandwidth overhead.
  • BER Reduction: At 9 dB SNR, the proposed scheme reduced the Bit Error Rate (BER) by 7.85% compared to the pilot-only baseline.

Conclusion and Future Work

This work demonstrates that cross-layer synergy between physical layer estimation and application-layer semantic processing is a powerful tool for improving communication reliability. By using LLMs to identify error-free symbols, the semantic pilot design significantly enhances channel estimation accuracy without increasing overhead. Future research will focus on extending this framework to Multiple-Input Multiple-Output (MIMO) systems and deploying lightweight, local LLMs to mitigate latency issues associated with cloud-based models.

Summary updated

Abstract

This paper proposes a semantic pilot design for data-aided channel estimation in text-inclusive data transmission, using a large language model (LLM). In this scenario, channel impairments often appear as typographical errors in the decoded text, which can be corrected using an LLM. The proposed method compares the initially decoded text with the LLM-corrected version to identify reliable decoded symbols. A set of selected symbols, referred to as a semantic pilot, is used as an additional pilot for data-aided channel estimation. To the best of our knowledge, this work is the first to leverage semantic information for reliable symbol selection. Simulation results demonstrate that the proposed scheme outperforms conventional pilot-only estimation, achieving lower normalized mean squared error and phase error of the estimated channel, as well as reduced bit error rate.
BibTeX
@misc{park2026semantic,
  author = {Sojeong Park and Hyun Jong Yang},
  title = {{Semantic Pilot Design for Data-Aided Channel Estimation Using a Large Language Model}},
  year = {2026},
  eprint = {2602.04126},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.04126},
}
Figure