Researchers have introduced X-AuT, a progressive compression framework designed to reduce the depth of audio encoders in speech large language models (LLMs) without compromising decoder-side embedding stability. To mitigate issues like premature end-of-sequence errors caused by block removal, the method employs a multi-stage approach involving layer combination selection via behavioral probes, representation alignment, and cross-scale distillation. During training, the LLM backbone is kept frozen, while adaptation is achieved through attention LoRA adapters and tied output embeddings. The framework utilizes a transcript-consistency pipeline to select high-agreement tiers for training, followed by source reweighting during fine-tuning. Experimental results on ten Chinese-English benchmarks demonstrate the efficacy of this progressive pruning: compressing the Qwen3-ASR-0.6B model from 18 to 16 layers improved the macro-average error from 5.61% to 5.27%. Furthermore, a 14-layer configuration achieved a 5.75% error rate while reducing audio-tower parameters by 20.7%. The study highlights that progressive pruning (18 to 14 layers) is superior to direct pruning, which yielded a 6.73% error rate. Additionally, using a 1.7B teacher model through the matched recipe resulted in a 5.55% mean error, significantly outperforming the 8.45% error rate observed with self-distillation.