Enhancing 6G Wireless Intelligence: Do LLMs Work for CSI Prediction?

Abstract: In high-mobility 6G scenarios, rapidly time-varying channels lead to very short coherence times, which makes conventional pilot-based channel state information (CSI) estimation approaches prone to outdated information or excessive pilot overhead. Therefore, channel prediction becomes essential in such dynamic wireless systems. To address this challenge, LLMs are emerging learning frameworks that have recently attracted attention for CSI prediction due to their strong sequence modeling capability and ability to generalize across different environments. This paper proposes an LLM-based framework for channel prediction in high-mobility orthogonal time frequency space (OTFS) communication systems. In this work, we develop a physics-aware LLM-based predictor that learns the temporal evolution of OTFS channel coefficients from historical channel observations while incorporating mobility-related physical descriptors (e.g., maximum Doppler frequency) to achieve accurate prediction of future channel states in rapidly time-varying environments. The effectiveness of the proposed framework is evaluated through extensive simulations under user velocities ranging from 100 to 500 km/h. Numerical results show that the proposed method consistently achieves lower normalized mean square error (NMSE) compared with both classical deep learning predictors and LLM-based predictors without physical channel descriptors. These results demonstrate the advantage of integrating mobility-related channel knowledge with LLM-based sequence modeling for channel prediction in highly dynamic OTFS systems.

• The paper introduces a hybrid LLM framework that integrates explicit mobility descriptors to enhance CSI prediction accuracy in high-mobility OTFS channels.
• It demonstrates that incorporating Doppler frequency data significantly lowers NMSE compared to SNR-based methods across various prediction horizons.
• The method reduces pilot overhead while ensuring robust, real-time CSI updates critical for 6G communication systems.

Physics-Aware LLM-Based CSI Prediction for High-Mobility OTFS Systems

Introduction

The transition to 6G wireless networks brings stringent requirements for high-mobility communications, where orthogonal time frequency space (OTFS) modulation is positioned as a robust alternative to OFDM due to its exploitation of the delay–Doppler (DD) domain. However, extreme temporal variability in such environments induces short coherence times and challenges conventional pilot-based CSI estimation methods, resulting in outdated channel knowledge or excessive pilot overhead. Accurate channel prediction thus becomes essential to sustain spectral efficiency and communication reliability.

Recent advances in AI-based approaches, particularly LLMs, offer promising sequence modeling capabilities for time-varying wireless channels. Nevertheless, the generic adoption of LLMs in OTFS channel prediction remains underexplored, and existing baselines often do not leverage explicit physical channel knowledge. This paper addresses that gap by introducing a physics-aware LLM-based framework, integrating mobility-related physical descriptors as auxiliary input to enhance prediction accuracy.

Background and Related Work

Traditional channel prediction approaches fall into two main categories:

1. Model-based predictors: These utilize analytical models with explicit physical parameters such as Doppler spread, delay spread, and employ methodologies like functional autoregressive (FAR) models [FAROTFS]. Their interpretability is strong, but they are highly sensitive to parameter estimation and model mismatch in fast-varying and heterogeneous settings.
2. AI-based predictors: These include conventional deep learning models and recent LLM-based approaches. For example, CNN-transformer hybrids extract spatial and temporal features [aitwo], while transformer descendants such as LinFormer utilize lightweight temporal mixing for CSI extrapolation [linformer]. Conditional generative models such as CVAEs seek to capture parameter-induced uncertainty [cvae]. Emerging LLM-based methods (e.g., LLM4CP [llmcp], FTAlign-LLM [ftalign], FAS-LLM [fasllm]) demonstrate benefits in sequence modeling and some degree of generalization, but typically ignore or underutilize physical dynamics, which limits their robustness in highly dynamic OTFS deployment scenarios.

Cross-modal fusion solutions (e.g., CSI-ALM [bridge]) further attempt to align semantic and physical representations, though these designs incur appreciable complexity overhead.

Proposed Physics-Aware LLM Framework

This study advances a hybrid LLM-based predictor for OTFS channels in non-stationary, high-mobility environments. The key innovation is the direct fusion of historical OTFS channel coefficients and explicit mobility-related physical descriptors, particularly Doppler frequency, into the LLM input pipeline. The architecture consists of four main modules:

• Preprocessing: Complex channel coefficients from OTFS frames are separated into real and imaginary parts, batched, normalized, and concatenated with descriptors capturing channel environment dynamics (e.g., maximum Doppler frequency, user velocity).
• Feature Embedding: Both CSI and dynamic descriptors are projected into a common high-dimensional embedding space via learned fully connected layers, followed by feature fusion (additive) and the addition of positional encodings for temporal order retention.
• LLM Backbone: A pretrained transformer-based LLM, e.g., GPT-2, processes the fused embeddings as input tokens. The majority of LLM weights are frozen to leverage broad sequence modeling, with task-specific fine-tuning applied to select modules.
• Output and Reconstruction: A final fully connected layer projects model outputs to the desired future frame dimension. The real-valued predictions are then recombined to recover complex channel coefficients for each future OTFS frame.

The optimization objective for supervised training is the minimization of normalized mean square error (NMSE) between predicted and ground-truth channel matrices.

Numerical Results

Simulation experiments use QuaDRiGa-generated synthetic OTFS channels under 3GPP UMa NLOS conditions, considering user velocities from 100 to 500 km/h, $N_P=16$ past frames for context, and a prediction horizon of $N_F=4$ future frames. The physical descriptor is primarily the maximum Doppler frequency.

Key findings include:

• Performance vs. Velocity: As expected, all predictors degrade as velocity increases due to diminishing correlation between past and future states. Nonetheless, the proposed physics-aware LLM achieves consistently lower NMSE. At 450 km/h, the Doppler-informed approach yields NMSE ≈ $4\times10^{-3}$, significantly outperforming SNR-based variants (NMSE ≈ $1.1\times10^{-2}$) and ablation baselines that lack physical descriptor input (NMSE ≈ $4\times10^{-2}$).
• Prediction Horizon Impact: The NMSE increases with longer prediction horizons (lower channel temporal correlation), but the advantage of mobility descriptor integration remains clear. At a 10-frame prediction horizon, the physics-aware approach attains NMSE ≈ $3\times10^{-3}$ versus ≈ $3\times10^{-2}$ for the descriptor-agnostic ablation.

The numerical superiority of Doppler descriptors over SNR is explicitly demonstrated, underscoring the importance of including environment-specific, mobility-relevant features for robust CSI prediction under high dynamism.

Theoretical and Practical Implications

The integration of mobility-aware physical descriptors into LLM-based channel prediction frameworks facilitates substantial improvements in prediction accuracy and generalization in OTFS scenarios with rapid channel time-variations. This methodology bridges the strengths of physics-based interpretability with the expressive power of large-scale pre-trained sequence models. In contrast to traditional deep learning approaches that are purely statistical, the explicit conditioning on physical parameters allows for better adaptation to specific propagation environments and user mobility conditions.

Moreover, this design reduces the dependency on high pilot overhead for time-critical CSI updates, enabling increased spectral efficiency and reliability, which are foundational for 6G applications such as vehicular communications, UAVs, and satellite links.

Future Directions

Anticipated trajectories include the extension of the physics-aware LLM prediction framework to multi-user and multi-cell OTFS systems, diverse heterogeneous propagation environments, and the incorporation of additional physical-awareness modalities (e.g., environment maps, sensor readings, or cross-domain data fusion). Transfer learning protocols leveraging broad LLM pretraining for rapid adaptation to new environments or hardware platforms present further promising avenues.

The observed strong gap between SNR-based and mobility descriptor-based performance also motivates deeper investigation into descriptor selection, feature engineering, and their joint representation learning with LLMs for next-generation wireless systems.

Conclusion

The proposed physics-aware LLM-based CSI prediction framework demonstrates significant gains over both traditional AI and LLM baselines by leveraging mobility-related physical descriptors. The results highlight that Doppler-aware conditioning of LLMs is essential for robust CSI forecasting in high-mobility OTFS channels, achieving the lowest NMSE across a broad range of velocities and prediction horizons. Practical deployment in 6G systems can benefit from the reduced overhead and improved adaptation that physics-aware AI brings to the wireless physical layer, marking a notable advance in the alignment of data-driven and physics-based wireless intelligence.

---

Reference:

"Enhancing 6G Wireless Intelligence: Do LLMs Work for CSI Prediction?" (2604.04028)