Pıer
TidesCurrentsHarbor LightsLabBottlesAshore
Pıer

Navigation

  • Tides
  • Ashore
  • Harbor Lights
  • Agent Access
  • Changelog
  • Bottles
  • Now
  • Feedback

External links

GitHubCloudborne ↗

© 2026 Pier.

WatchingResearchWatching0 independent reports0

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs via Cross-Scale Distillation

First seen · 9/10/2026, 04:00 AMLatest activity · 9/10/2026, 04:00 AM

Trimming audio-encoder depth cuts inference overhead in speech LLMs, but directly dropping blocks often perturbs decoder embeddings, triggering token deletions and premature end-of-sequence stops. X-AuT addresses this with a progressive compression framework that selects layer subsets through behavioral probes and recovers performance via cross-scale distillation and LoRA tuning while keeping the language model backbone frozen. Evaluated on Qwen3-ASR-0.6B across ten Chinese-English benchmarks, pruning the audio encoder from 18 to 14 layers reduced audio-tower parameters by 20.7% with a macro-average error of 5.75%, outpacing direct pruning.

Event heat · last 24 hours

There are 6 persisted snapshots in the last 24 hours. Peak heat was 0 at 9/12, 20:00; latest heat is 0.

There are 6 persisted snapshots in the last 24 hours. Peak heat was 0 at 9/12, 20:00; latest heat is 0.10.509/12, 20:00, event heat 09/12, 23:00, event heat 09/13, 02:00, event heat 09/13, 05:00, event heat 09/13, 08:00, event heat 09/13, 11:00, event heat 024 hours agoNow
  1. 9/12, 20:00, event heat 0
  2. 9/12, 23:00, event heat 0
  3. 9/13, 02:00, event heat 0
  4. 9/13, 05:00, event heat 0
  5. 9/13, 08:00, event heat 0
  6. 9/13, 11:00, event heat 0

Reporting Timeline

  1. AggregatorHuggingFace Daily Papers9/10, 04:00 AMnot independentRepresentative
    X-AuT: Progressive Audio-Encoder Compression for Speech LLMs via Cross-Scale Distillation