Yasmin Moslem

I build efficient multilingual, multimodal systems.

Senior Researcher at the ADAPT Centre, School of Computer Science, Trinity College Dublin. My work covers multi-LLM routing and cascades, iterative layer pruning, knowledge distillation, and quantization, applied across text, speech, and translation. Research funded by Huawei and Microsoft Research. Projects shipped in production at Big Language Solutions, Bering Lab, and Future Group.

Dublin, Ireland

Projects

Efficient inference · with Huawei Research, 2026

CRE-Router, cost-aware multi-LLM serving

CRE-Router is a two-stage framework that first routes clusters of queries to the most cost-effective model in a pool, and then escalates queries with low-quality outputs to a stronger model, saving costs while retaining quality.

ClusteringCost-aware routingEscalation

Code Paper

Speech · IWSLT 2025

Halving an 8B speech translation model

Compressed Qwen2-Audio to half its parameters and storage by combining iterative layer pruning, QLoRA, and knowledge distillation, while holding on to almost all of the teacher model's quality.

50% fewer parameters 97 to 100% of teacher quality

Code Paper

Speech · IWSLT 2026

Cross-lingual voice cloning for scientific speech

Multi-model distillation for cross-lingual voice cloning applied to scientific speech, alongside synthetic audio to improve low-resource speech translation.

Voice cloningTTSDistillation

Paper Findings

Model compression · WMT 2025

Pruning a multilingual LLM without losing quality

Iterative layer pruning applied to Aya-Expanse 8B. Removing between a fifth and nearly half of the model left translation quality almost intact and made it substantially faster to serve, with a further speedup once deployed through vLLM.

22 to 44% smaller 99% of COMET retained up to 84% higher throughput 4.3× faster serving

Code Paper

Open models · AfricaNLP 2026, Spotlight

AfriNLLB, faster open translation for African languages

Compressed translation models covering 15 language pairs in both directions, released openly alongside a curated training set. They run considerably faster than the NLLB-200 baseline at equal or better quality.

30 directions 2.4× faster inference 1.6M open segments

Code Models and dataset Paper

Models and datasets

Most of what I build ships publicly. Compressed translation and speech models, the AfriNLLB releases and their 1.6M-segment dataset, and the models and data produced through the SpeechT programme are all openly available.

Hugging Face GitHub

Publications

  1. Cluster, route, escalate: Cascaded framework for cost-aware LLM serving arXiv, 2026 Link
  2. Dynamic model routing and cascading for efficient LLM inference: A survey Transactions on Machine Learning Research (TMLR), 2026 Link
  3. Iterative layer pruning for efficient translation inference WMT 2025 Link
  4. Efficient speech translation through model compression and knowledge distillation IWSLT 2025 Link
  5. SEACrowd: A multilingual multimodal data hub and benchmark suite for Southeast Asian languages EMNLP 2024 Link
  6. Leveraging synthetic audio data for end-to-end low-resource speech translation IWSLT 2024 Link
  7. Adaptive machine translation with large language models EAMT 2023 300+ citations Link
  8. Domain-specific text generation for machine translation AMTA 2022 Best Presentation Award Link

Full list on Google Scholar

Community leadership

Background

Ten years of building NLP systems across more than thirty languages, in research labs and in production.

Education