From Text to Beams: Understanding the Role of Representation, Generation, and Linear Decodability for LLM-Based mmWave Beam Prediction

Large language models (LLMs) are increasingly
proposed as flexible reasoning components for 6G wireless
systems, including integrated sensing and communication (ISAC)
pipelines. In these systems, beam management decisions may
depend on heterogeneous and unstructured contextual information.
Previous work has integrated LLMs into beam prediction
through fine-tuning or trainable embedding adapters. However,
an important question remains open: Does a frozen generalpurpose
LLM lack the reasoning ability for this task, or can it
understand the task but fail to express the correct answer as text?
This work addresses this question directly. We study mmWave
beam prediction from natural-language scene descriptions using
DeepMIMO ray-traced channels. We compare two approaches:
directly prompting a frozen LLM to generate a beam index
and extracting its frozen internal hidden-state representation
for beam prediction using a lightweight classifier. We evaluate
two independent model families, Qwen2.5 and Llama, with
models ranging from 1.5B to 8B parameters. We test two
DeepMIMO scenarios and three random seeds. We also compare
our methods with geometric, majority-class, and MLP baselines.
The embedding-probe approach outperforms direct generation
by 2–6× in Top-1 beam accuracy. Increasing model size also
improves generative accuracy, while probe accuracy remains
almost unchanged. This pattern appears consistently in both
model families. These results indicate that the frozen LLM
representations encode substantially more beam relevant information
than is expressed through direct text generation. Frozen
LLMs can capture useful information for beam prediction, but
they often struggle to express the correct beam index through
generated text. This limitation is also not specific to a single
model architecture or training recipe. Therefore, we propose a
clear design guideline for LLM-assisted ISAC systems: extract
features from frozen LLM representations instead of relying on
generated text output.