Research map / Architectures and efficiency / RNNs and normalization
xLSTM Scaling Laws: Competitive Performance with Linear Time-Complexity
Maximilian Beck, Kajetan Schweighofer, Sebastian Böck, Sebastian Lehner, 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 |
|---|---|---|---|
| use xLSTM instead of Transformer | LLM pretraining · validation loss vs FLOPs, ~5 orders of magnitude of compute | cross-entropy loss | Improved |
| use xLSTM instead of Transformer | LLM inference · 16k prefill | time to first token: 30-50% lower TTFT for the same model size | Improved |
| use xLSTM instead of Transformer | LLM inference · 16k prefill | step time | Improved |
| increase token-to-parameter ratio into over-training (ablation) | xLSTM · loss vs compute power law | exponent eta: 0.047 (M=22) -> 0.047 (M=1100) | No clear effect |
| increase token-to-parameter ratio into over-training (ablation) | Transformer · loss vs compute power law | exponent eta: 0.050 (M=22) -> 0.049 (M=550) | No clear effect |
| use xLSTM instead of Transformer | compute-optimal model size · IsoFLOP | optimal parameters N* | Improved |
| increase context length from 2048 to 16384 (ablation) | Transformer · IsoFLOP | compute-optimal model size | Hurt |
| increase context length from 2048 to 16384 (ablation) | xLSTM · IsoFLOP | compute-optimal model size | No clear effect |
| include gamma as a free parameter in the loss surface fit (ablation) | parametric loss fit · Token/Param runs | fit quality | Improved |
Closest papers by meaning
- LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws (ICML 2025 · similarity 0.91)
- xLSTM 7B: A Recurrent LLM for Fast and Efficient Inference (ICML 2025 · similarity 0.91)
- Scaling Inference-Efficient Language Models (ICML 2025 · similarity 0.91)
- Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.91)
- Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving (ICLR 2025 · similarity 0.91)
- Provable Scaling Laws for the Test-Time Compute of Large Language Models (NeurIPS 2025 · similarity 0.90)
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.
