On Subquadratic Architectures: From Applications to Principles

Hugging Face Daily Papers Papers

Summary

This paper compares xLSTM, Mamba-2, and Gated DeltaNet on complex sequence modeling tasks and finds xLSTM superior due to its enhanced state tracking and memory dynamics, validated on synthetic length-generalization tasks.

Transformers dominate modern sequence modeling, but their quadratic attention incurs substantial computational cost. Subquadratic architectures offer a scalable alternative. However, it remains unclear which designs yield the most effective sequence models. We compare three leading approaches: xLSTM, Mamba-2, and Gated DeltaNet. We evaluate these models on tasks with complex dependencies: (1) code-model pre-training, (2) distillation of code models from large language models, and (3) pre-training of time-series foundation models. Across these settings, xLSTM delivers the strongest overall performance. To explain xLSTM's advantage, we present a unified formulation and analyze the underlying architectural mechanisms, focusing on state tracking and memory dynamics. Our results show that xLSTM enables more flexible and stable memory correction via its gating scheme. We corroborate these findings on controlled synthetic length-generalization tasks. Overall, our findings indicate that xLSTM's gains on complex tasks stem from robust state tracking and accumulation.
Original Article
View Cached Full Text

Cached at: 06/11/26, 01:36 PM

Paper page - On Subquadratic Architectures: From Applications to Principles

Source: https://huggingface.co/papers/2606.12364

Abstract

xLSTM demonstrates superior performance in sequence modeling tasks compared to Mamba-2 and Gated DeltaNet due to enhanced state tracking and memory dynamics.

Transformersdominate modernsequence modeling, but their quadraticattentionincurs substantial computational cost.Subquadratic architecturesoffer a scalable alternative. However, it remains unclear which designs yield the most effective sequence models. We compare three leading approaches:xLSTM,Mamba-2, andGated DeltaNet. We evaluate these models on tasks with complex dependencies: (1)code-model pre-training, (2)distillationof code models from large language models, and (3) pre-training oftime-series foundation models. Across these settings,xLSTMdelivers the strongest overall performance. To explainxLSTM’s advantage, we present a unified formulation and analyze the underlying architectural mechanisms, focusing onstate trackingandmemory dynamics. Our results show thatxLSTMenables more flexible and stable memory correction via itsgating scheme. We corroborate these findings on controlled syntheticlength-generalization tasks. Overall, our findings indicate thatxLSTM’s gains on complex tasks stem from robuststate trackingand accumulation.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2606\.12364

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.12364 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.12364 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.12364 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Unlocking Feature Learning in Gated Delta Networks at Scale

arXiv cs.LG

This paper derives scaling rules for Gated Delta Networks using Maximal Update Parametrization (μP), enabling zero-shot hyperparameter transfer across model widths for efficient sub-quadratic LLM architectures. Experiments confirm stable learning-rate transfer under both AdamW and SGD, whereas standard parametrization fails.

Looped State-Space Language Models with Adaptive Exit-State Selection

arXiv cs.AI

This paper explores looped (recurrent) state-space language models using Mamba and hybrid Mamba-Transformer backbones, showing they outperform non-looped baselines on reasoning tasks and remain competitive under iso-parameter and iso-FLOPs pretraining, with adaptive exit-state selection improving intermediate-depth performance.