Effective Distillation to Hybrid xLSTM Architectures
2026-03-16 • Machine Learning
Machine Learning
AI summaryⓘ
The authors tried to create smaller, faster language models that still work as well as bigger teacher models. They focused on making sure the smaller models don't lose performance by carefully measuring how often they match or tie with the bigger models on different tasks. To do this, they designed a new step where small expert parts are combined into one efficient model. Testing their approach on several popular models, their smaller versions often kept up with or even did better than the originals. This work helps make language models more energy-efficient and cheaper to use.
large language modelsdistillationquadratic attentionlinearized architecturesxLSTMinstruction tuningWin-and-Tie ratemodel mergingtransformersenergy efficiency
Authors
Lukas Hauzenberger, Niklas Schmidinger, Thomas Schmied, Anamaria-Roberta Hartl, David Stap, Pieter-Jan Hoedt, Maximilian Beck, Sebastian Böck, Günter Klambauer, Sepp Hochreiter
Abstract
There have been numerous attempts to distill quadratic attention-based large language models (LLMs) into sub-quadratic linearized architectures. However, despite extensive research, such distilled models often fail to match the performance of their teacher LLMs on various downstream tasks. We set out the goal of lossless distillation, which we define in terms of tolerance-corrected Win-and-Tie rates between student and teacher on sets of tasks. To this end, we introduce an effective distillation pipeline for xLSTM-based students. We propose an additional merging stage, where individually linearized experts are combined into a single model. We show the effectiveness of this pipeline by distilling base and instruction-tuned models from the Llama, Qwen, and Olmo families. In many settings, our xLSTM-based students recover most of the teacher's performance, and even exceed it on some downstream tasks. Our contributions are an important step towards more energy-efficient and cost-effective replacements for transformer-based LLMs.