Research map / Large language models / Language, safety and interpretability
Effective Distillation to Hybrid xLSTM Architectures
Lukas Hauzenberger, Niklas Schmidinger, Thomas Schmied, Anamaria-Roberta Hartl, David Stap, Pieter-Jan Hoedt, Sebastian Böck, Günter Klambauer, Sepp Hochreiter
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| merge four domain experts instead of training a single generalist (ablation) | xLSTM-Llama3.1-8B-IT · MATH500 | score: 0.37 -> 0.54 | Improved |
| merge four domain experts instead of a generalist (ablation) | xLSTM-Llama3.1-8B-IT · IFEval | score: 0.49 -> 0.69 | Improved |
| merge four domain experts instead of a generalist (ablation) | xLSTM-Qwen2.5-7B-IT · GPQA-D | score: 0.19 -> 0.26 | Improved |
| merge domain experts into one student (ablation) | xLSTM IT students · GPQA / GPQA-Diamond (STEM) | score | Hurt |
| merge domain experts into one student (ablation) | xLSTM IT students · IFEval (Llama) | score | Improved |
| use mLSTM gating instead of pure linear attention (ablation) | hybrid student · distillation loss, Llama3.1-8B 2.5B tokens | loss | Improved |
| add sliding-window attention to mLSTM (ablation) | mLSTM student · distillation loss, Llama3.1-8B | loss | Improved |
| add attention sink tokens (ablation) | mLSTM+SWA student · distillation loss / downstream | loss | Improved |
| train with CE+KL mix (gamma=0.9, beta=0.1) instead of pure KL (ablation) | stage II distillation · Llama3.1-8B | performance | Improved |
| use full fine-tuning instead of LoRA (ablation) | linearization · Llama3.1-8B | performance | Improved |
| add normalization layers before output projections in mLSTM (ablation) | mLSTM student · student-teacher alignment | alignment | Hurt |
| use concatenated [q k v] as input to the output gate instead of x_t (ablation) | mLSTM output gate · Llama3.1-8B | signal | Improved |
Closest papers by meaning
- On-the-Fly Adaptive Distillation of Transformer to Dual-State Linear Attention for Long-Context LLM Serving (ICML 2025 · similarity 0.88)
- DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.88)
- Joint Fine-tuning and Conversion of Pretrained Speech and Language Models towards Linear Complexity (ICLR 2025 · similarity 0.88)
- Strong Teacher Not Needed? On Distillation in LLM Pretraining (NeurIPS 2026 · similarity 0.88)
- xLSTM 7B: A Recurrent LLM for Fast and Efficient Inference (ICML 2025 · similarity 0.88)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
