X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
Updated 1 h ago · first seen 11 Sept 2026
paper_01M294WYDTQJF309BA9AX08EJW
- Published
- 10 Sept 2026
- T2 · 5 h ago
- arXiv
- 2609.11412
- T2 · 5 h ago
Abstract
Reducing audio-encoder depth lowers the inference cost of speech large language models, but removing complete blocks perturbs the embeddings consumed by the decoder and can cause deletion and premature end-of-sequence errors. We introduce X-AuT, a progressive framework that selects layer combinations through short behavioral probes and restores the pruned model through representation alignment, cross-scale distillation, scheduled student-policy supervision, and LoRA finetuning. The language-model backbone remains frozen, while attention LoRA adapters and the tied output embedding adapt during distillation. Training uses the highest-agreement tier from a transcript-consistency pipeline, followed by source reweighting during finetuning. On ten public Chinese--English benchmarks, compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers reduces macro-average error from 5.61% to 5.27%. The 14-layer model reaches 5.75% with 20.7% fewer audio-tower parameters. Under the matched recipe, the 1.7B teacher yields 5.55% mean error, compared with 8.45% for self-distillation, and progressive 18rightarrow14 pruning outperforms direct pruning (5.75% vs. 6.73%). These single-run results establish two practical operating points and show that the accuracy effects vary across benchmarks. Project website: https://xpeng-ai.github.io/x-aut
Authors 10
Haojun Zhang, Yi Zou, Min Chen, Qize Yu, Lianrui Fan, Xini Ding, Hao Li, Shuchang Zhou, Xianming Liu, Shiyu Huang
Specification
- Official page
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- arXiv id
- 2609.11412
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Github repo
- XPENG-AI/X-AuT
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Hf paper url
- https://huggingface.co/papers/2609.11412
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Github stars
- 7
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 1 h agomedium
- Hf comments
- 2
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 1 h agomedium
- Upvotes
- 13
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 1 h agomedium
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Published
- 10 Sept 2026
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
Each value shows its source, tier and observation time. Conflicting claims are kept side by side and flagged — never averaged. How AI Atlas records facts →
Provenance
Attributed facts
11
Source tiers
T211
Freshest observation
1 h ago
Conflicts
None
No models linked to this paper yet.
No relations recorded.
As of
Rewind the record: see this entity's attributes exactly as AI Atlas knew them on a given day.
Claim history · Abstract
Abstractabstract1
| Value | Valid from → to | Status | Source | Confidence | Extractor |
|---|---|---|---|---|---|
| Reducing audio-encoder depth lowers the inference cost of speech large language models, but removing complete blocks perturbs the embeddings consumed by the decoder and can cause deletion and premature end-of-sequence errors. We introduce X-AuT, a progressive framework that selects layer combinations through short behavioral probes and restores the pruned model through representation alignment, cross-scale distillation, scheduled student-policy supervision, and LoRA finetuning. The language-model backbone remains frozen, while attention LoRA adapters and the tied output embedding adapt during distillation. Training uses the highest-agreement tier from a transcript-consistency pipeline, followed by source reweighting during finetuning. On ten public Chinese--English benchmarks, compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers reduces macro-average error from 5.61% to 5.27%. The 14-layer model reaches 5.75% with 20.7% fewer audio-tower parameters. Under the matched recipe, the 1.7B teacher yields 5.55% mean error, compared with 8.45% for self-distillation, and progressive 18rightarrow14 pruning outperforms direct pruning (5.75% vs. 6.73%). These single-run results establish two practical operating points and show that the accuracy effects vary across benchmarks. Project website: https://xpeng-ai.github.io/x-aut | → current | current | Hugging Face Hub (public pages, model cards, papers)T2 | medium | deterministic |
Claims are temporal and append-only: a new observation closes the previous claim (valid_to) instead of overwriting it. Conflicting claims from different sources are kept side by side and flagged — never averaged. Methodology →
- New paperPaperX-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
New paper: X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
huggingface
| Source | Document | Type | Tier | Last observed | Snapshots |
|---|---|---|---|---|---|
| Hugging Face Hub (public pages, model cards, papers) | huggingface.co/papers | listing | T2· Quality secondary | 1 h ago | 4 |
Tier 1 = official/primary, 2 = quality secondary, 3 = community, 4 = unverified. Every snapshot is archived; see all sources and the methodology.