pith. machine review for the scientific record.
sign in

arxiv: 2509.24552 · v3 · submitted 2025-09-29 · 💻 cs.LG · cs.AI

Short window attention enables long-term memorization

Lo\"ic Cabannes , Maximilian Beck , Gergely Szilvasy , Matthijs Douze , Maria Lomeli , Jade Copet , Pierre-Emmanuel Mazar\'e , Gabriel Synnaeve
show 1 more author
Herv\'e J\'egou
This is my paper

Pith reviewed 2026-05-18 12:07 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords sliding window attentionxLSTMhybrid architectureslong-context modelingstochastic window sizelong-term memorylocal-global attention
0
0 comments X

The pith

Short sliding windows strengthen long-term memory in hybrid attention-xLSTM models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper studies hybrid architectures that combine local sliding-window attention layers with global xLSTM linear RNN layers. It finds that smaller window sizes improve long-context performance because the model can no longer rely on attention for distant retrieval and must instead strengthen its xLSTM memory. The same pattern appears when alternating short-window and full-attention layers: the short layers must stay small to keep the full layers useful. Overly small fixed windows hurt short-context tasks, so the authors train models with randomly varying window sizes, which improves results on both short and long sequences.

Core claim

In the SWAX hybrid of sliding-window attention and xLSTM layers, larger sliding windows reduce long-context performance while shorter windows improve it by forcing the xLSTM to handle long-term retrieval that local attention can no longer cover. The same holds for local-global attention stacks, where short layers must remain small. Training with stochastic window sizes lets the model use both short-term local information and long-term memory, outperforming fixed-window baselines on short- and long-context problems.

What carries the argument

The sliding-window attention mechanism whose length controls how much retrieval is offloaded to the xLSTM linear RNN layers.

If this is right

  • Shorter fixed windows improve long-context results by increasing dependence on xLSTM memory.
  • In alternating local-global attention stacks, keeping short layers small preserves the value of full attention layers.
  • Stochastic variation of window size during training yields gains on both short- and long-context tasks over any fixed window.
  • Excessively small fixed windows degrade short-context performance that moderate windows could handle.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Architects of long-context systems may improve global memory by deliberately restricting local context windows.
  • The same short-window pressure could be tested on other recurrent or stateful modules paired with attention.
  • Measuring internal long-range retrieval accuracy before and after short-window training would directly test the claimed mechanism.

Load-bearing premise

Gains from shorter windows come from forcing greater use of xLSTM long-term memory rather than from incidental changes in gradients or regularization.

What would settle it

Train the same short-window model but add an auxiliary long-range retrieval path that bypasses the xLSTM; if long-context gains disappear, the memory-forcing account is supported.

read the original abstract

Recent works show that hybrid architectures combining local sliding window attention layers and global attention layers outperform either of these architectures taken separately. However, the impact of the window length and the interplay between local layers and global layers remain under-studied. In this work, we first analyze the interaction between short and long term memory by considering SWAX: a hybrid architecture consisting of sliding-window attention and xLSTM linear RNN layers. A counter-intuitive finding is that larger sliding windows hurts the long-context performance. In fact, short window attention encourages the model to better train the long-term memory of the xLSTM as it cannot rely on the local softmax attention mechanism for long context-retrieval. We also validate our findings on local-global architectures alternating short window and full attention layers: the short layers should be small in order not to hinder the usefulness of the long layers. However, employing too small sliding windows is detrimental even for short-context tasks, which could be solved with information from moderately larger sliding windows otherwise. Therefore, we train hybrid architectures by stochastically changing the sliding window size, forcing the model to leverage both the short term window and the long-term memory. Training with stochastic window sizes significantly outperforms regular window attention both on short and long-context problems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces SWAX, a hybrid architecture combining sliding-window attention layers with xLSTM linear RNN layers. It reports the counter-intuitive result that larger sliding windows degrade long-context performance, attributing this to short windows forcing greater reliance on and better training of the xLSTM long-term memory. The finding is extended to alternating short-window and full-attention hybrids, and a stochastic window-size training procedure is proposed that improves results on both short- and long-context tasks.

Significance. If the central mechanism holds, the work supplies a lightweight, parameter-free training intervention (stochastic window sizing) that could improve long-term memorization in hybrid attention-RNN models without increasing compute. The result would be practically useful for scaling context length in resource-constrained settings and would motivate further study of how local attention interacts with recurrent memory.

major comments (3)
  1. [Abstract and experimental validation sections] The interpretation that short windows improve long-context performance specifically by compelling the xLSTM to learn better long-term memory (rather than through incidental effects on gradient flow, regularization, or effective capacity) is load-bearing for the central claim yet unsupported by isolating experiments. No memory-state ablations, gradient-norm measurements, or matched-capacity controls are described that would separate the proposed mechanism from these confounds.
  2. [Validation on local-global architectures] The statement that 'short layers should be small in order not to hinder the usefulness of the long layers' is presented as a general guideline, but the manuscript provides no quantitative analysis of the interaction (e.g., performance curves versus window size for fixed long-layer capacity) or statistical significance of the reported gains.
  3. [Training with stochastic window sizes] The stochastic window-size training method is claimed to outperform regular window attention on both short- and long-context problems, but the abstract and description lack details on the distribution from which window sizes are sampled, the frequency of resampling, and whether the improvement survives when total training compute is matched.
minor comments (2)
  1. [Introduction / Architecture description] Notation for the hybrid layer ordering and the precise definition of 'short' versus 'long' context lengths should be clarified with a diagram or explicit equations early in the manuscript.
  2. [Experimental setup] The manuscript would benefit from an explicit statement of the baseline models and hyper-parameter search protocol used for all reported comparisons.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and insightful comments, which have helped us identify areas where the manuscript can be strengthened. We address each major comment below and commit to revisions that provide additional experimental support and details without altering the core findings.

read point-by-point responses

  1. Referee: [Abstract and experimental validation sections] The interpretation that short windows improve long-context performance specifically by compelling the xLSTM to learn better long-term memory (rather than through incidental effects on gradient flow, regularization, or effective capacity) is load-bearing for the central claim yet unsupported by isolating experiments. No memory-state ablations, gradient-norm measurements, or matched-capacity controls are described that would separate the proposed mechanism from these confounds.

    Authors: We agree that isolating the proposed mechanism from potential confounds such as gradient flow or regularization effects would strengthen the central claim. Our existing results demonstrate that shorter windows consistently yield better long-context performance in the hybrid SWAX architecture, which we interpret as evidence of increased reliance on xLSTM long-term memory. However, we acknowledge the value of direct ablations. In the revised manuscript we will add memory-state analyses (e.g., inspecting or intervening on xLSTM hidden states), gradient-norm comparisons across window sizes, and matched-capacity controls that adjust for effective model capacity or regularization strength. revision: yes

  2. Referee: [Validation on local-global architectures] The statement that 'short layers should be small in order not to hinder the usefulness of the long layers' is presented as a general guideline, but the manuscript provides no quantitative analysis of the interaction (e.g., performance curves versus window size for fixed long-layer capacity) or statistical significance of the reported gains.

    Authors: We appreciate this feedback on the need for more rigorous quantification. The guideline is drawn from our experiments showing that larger short-window layers can diminish the contribution of the full-attention layers in alternating local-global setups. To address the concern, the revision will include performance curves of task metrics versus short-window size under fixed long-layer capacity, along with statistical significance testing (multiple random seeds and appropriate hypothesis tests) for the reported improvements. revision: yes

  3. Referee: [Training with stochastic window sizes] The stochastic window-size training method is claimed to outperform regular window attention on both short- and long-context problems, but the abstract and description lack details on the distribution from which window sizes are sampled, the frequency of resampling, and whether the improvement survives when total training compute is matched.

    Authors: We concur that these methodological details are essential for reproducibility and for confirming that gains are not artifacts of unequal compute. The revised manuscript will specify the exact sampling distribution (e.g., uniform over a defined range of window sizes), the resampling frequency (e.g., per batch or per epoch), and will include controlled experiments that match total training compute (by equating FLOPs or step counts) to verify that the stochastic-window approach retains its advantages on both short- and long-context benchmarks. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical claims without derivation chain

full rationale

The paper reports experimental results comparing hybrid sliding-window attention and xLSTM architectures across different window sizes and stochastic training regimes. No equations, first-principles derivations, or fitted parameters are presented that reduce any claimed prediction to its own inputs by construction. Central findings (e.g., short windows improving long-context performance) rest on direct benchmark measurements rather than self-referential definitions or self-citation load-bearing steps. The work is self-contained against external replication and does not invoke uniqueness theorems or ansatzes from prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on the assumption that observed gains arise from memory specialization rather than confounding factors, plus standard assumptions about attention and RNN layers functioning as described in prior literature.

axioms (1)
  • domain assumption Hybrid local-global attention architectures can be stably trained and compared when window size is varied.
    Invoked when claiming that short windows improve long-term memory training without destabilizing optimization.

pith-pipeline@v0.9.0 · 5788 in / 1174 out tokens · 32831 ms · 2026-05-18T12:07:23.023850+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 22 internal anchors

  1. [1]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton

    URL https://arxiv.org/abs/2402.18668. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. Program synthesis with large language models,

    work page arXiv

  2. [2]

    Program Synthesis with Large Language Models

    URL https://arxiv.org/abs/2108.07732. Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Layer Normalization

    URL https://arxiv.org/abs/ 1607.06450. Maximilian Beck, Korbinian Pöppel, Markus Spanring, Andreas Auer, Oleksandra Prudnikova, Michael Kopp, Günter Klambauer, Johannes Brandstetter, and Sepp Hochreiter. xlstm: Extended long short-term memory,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    10 Maximilian Beck, Korbinian Pöppel, Phillip Lippe, and Sepp Hochreiter

    URL https://arxiv.org/abs/2405.04517. 10 Maximilian Beck, Korbinian Pöppel, Phillip Lippe, and Sepp Hochreiter. Tiled Flash Linear Attention: More efficient linear rnn and xlstm kernels.arXiv, 2503.14376, 2025a. URLhttps://arxiv.org/abs/2503.14376. Maximilian Beck, Korbinian Pöppel, Phillip Lippe, Richard Kurle, Patrick M. Blies, Günter Klambauer, Sebasti...

    work page arXiv

  5. [5]

    Longformer: The Long-Document Transformer

    URL https: //arxiv.org/abs/2004.05150. Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning about physical commonsense in natural language,

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  6. [6]

    URLhttps://arxiv.org/abs/1911.11641. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Po...

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  7. [7]

    Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

    URLhttps://arxiv.org/abs/1412.3555. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge

    URL https://arxiv.org/abs/ 1803.05457. Soham De, Samuel L. Smith, Anushan Fernando, Aleksandar Botev, George Cristian-Muraru, Albert Gu, Ruba Haroun, Leonard Berrada, Yutian Chen, Srivatsan Srinivasan, Guillaume Desjardins, Arnaud Doucet, David Budden, Yee Whye Teh, Razvan Pascanu, Nando De Freitas, and Caglar Gulcehre. Griffin: Mixing gated linear recurr...

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models

    URLhttps://arxiv.org/abs/2402.19427. DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    URL https: //arxiv.org/abs/2501.12948. Xin Dong, Yonggan Fu, Shizhe Diao, Wonmin Byeon, Zijia Chen, Ameya Sunil Mahabaleshwarkar, Shih-Yang Liu, Matthijs Van Keirsbilck, Min-Hung Chen, Yoshi Suhara, Yingyan Lin, Jan Kautz, and Pavlo Molchanov. Hymba: A hybrid-head architecture for small language models,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Hymba: A hybrid-head architecture for small language models.arXiv preprint arXiv:2411.13676,

    URLhttps://arxiv.org/abs/2411.13676. Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,

    work page arXiv

  12. [12]

    Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning

    URLhttps://arxiv.org/abs/1702.03118. Lizhe Fang, Yifei Wang, Zhaoyang Liu, Chenheng Zhang, Stefanie Jegelka, Jinyang Gao, Bolin Ding, and Yisen Wang. What is wrong with perplexity for long-context language modeling?,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Quentin Carbonneaux, Taco Cohen, and Gabriel Synnaeve

    URLhttps://arxiv.org/abs/2212.14052. Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Quentin Carbonneaux, Taco Cohen, and Gabriel Synnaeve. Rlef: Grounding code llms in execution feedback with reinforcement learning,

    work page arXiv

  14. [14]

    Google DeepMind Gemma Team

    URL https://arxiv.org/ abs/2410.02089. Google DeepMind Gemma Team. Gemma 3 technical report,

    work page arXiv

  15. [15]

    Gemma 3 Technical Report

    URLhttps://arxiv.org/abs/2503.19786. Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    URL https: //arxiv.org/abs/2312.00752. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.Neural Computation, 9:1735–1780, 11

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Long Short-Term memory,

    doi: 10.1162/neco.1997.9.8.1735. Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What’s the real context size of your long-context language models?,

    work page doi:10.1162/neco.1997.9.8.1735 1997

  18. [18]

    RULER: What's the Real Context Size of Your Long-Context Language Models?

    URL https://arxiv. org/abs/2404.06654. 11 Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Weld, and Luke Zettlemoyer

    Association for Computational Linguistics. doi: 10.18653/v1/P17-1147. URL https://aclanthology.org/ P17-1147/. Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention,

    work page doi:10.18653/v1/p17-1147

  20. [20]

    URLhttps://arxiv.org/abs/2006.16236. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural questions: A benchmark for...

    work page arXiv 2006

  21. [21]

    URL https://aclanthology.org/Q19-1026/

    doi: 10.1162/tacl_a_00276. URL https://aclanthology.org/Q19-1026/. Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. Race: Large-scale reading comprehension dataset from examinations,

    work page doi:10.1162/tacl_a_00276

  22. [22]

    RACE: Large-scale ReAding Comprehension Dataset From Examinations

    URLhttps://arxiv.org/abs/1704.04683. Bo Liu, Rui Wang, Lemeng Wu, Yihao Feng, Peter Stone, and Qiang Liu. Longhorn: State space models are amortized online learners,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Hanxiao Liu, Zihang Dai, David R

    URLhttps://arxiv.org/abs/2407.14207. Hanxiao Liu, Zihang Dai, David R. So, and Quoc V . Le. Pay attention to mlps,

    work page arXiv

  24. [24]

    Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang

    URL https://arxiv.org/abs/ 2105.08050. Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by chatGPT really correct? rigorous evaluation of large language models for code generation. InThirty-seventh Conference on Neural Information Processing Systems,

    work page arXiv

  25. [25]

    URL https://arxiv.org/abs/2504.03624. OpenAI. gpt-oss-120b & gpt-oss-20b model card,

    work page arXiv

  26. [26]

    gpt-oss-120b & gpt-oss-20b Model Card

    URLhttps://arxiv.org/abs/2508.10925. Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. Yarn: Efficient context window extension of large language models,

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    YaRN: Efficient Context Window Extension of Large Language Models

    URLhttps://arxiv.org/abs/2309.00071. Liliang Ren, Yang Liu, Yadong Lu, Yelong Shen, Chen Liang, and Weizhu Chen. Samba: Simple hybrid state space models for efficient unlimited context language modeling,

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Ricardo Buitrago Ruiz and Albert Gu

    URLhttps://arxiv.org/abs/2406.07522. Ricardo Buitrago Ruiz and Albert Gu. Understanding and improving length generalization in recurrent models,

    work page arXiv

  29. [29]

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi

    URLhttps://arxiv.org/abs/2507.02782. Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale,

    work page arXiv

  30. [30]

    WinoGrande: An Adversarial Winograd Schema Challenge at Scale

    URLhttps://arxiv.org/abs/1907.10641. Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. Socialiqa: Commonsense reasoning about social interactions,

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  31. [31]

    SocialIQA: Commonsense Reasoning about Social Interactions

    URLhttps://arxiv.org/abs/1904.09728. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958,

    work page internal anchor Pith review Pith/arXiv arXiv 1904

  32. [32]

    RoFormer: Enhanced Transformer with Rotary Position Embedding

    URLhttps://arxiv.org/abs/2104.09864. Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, and Carlos Guestrin. Learning to (learn at test time): Rnns with expressive hidden states,

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Learning to (Learn at Test Time): RNNs with Expressive Hidden States

    URLhttps://arxiv.org/abs/2407.04620. Dustin Wang, Rui-Jie Zhu, Steven Abreu, Yong Shan, Taylor Kergan, Yuqi Pan, Yuhong Chou, Zheng Li, Ge Zhang, Wenhao Huang, and Jason Eshraghian. A systematic analysis of hybrid linear attention,

    work page internal anchor Pith review Pith/arXiv arXiv

  34. [34]

    org/abs/2507.06457

    URL https://arxiv. org/abs/2507.06457. 12 Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models,

    work page arXiv

  35. [35]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    URL https://arxiv.org/abs/ 2201.11903. Guangxuan Xiao. Why stacking sliding windows can’t see very far. https://guangxuanx.com/blog/ stacking-swa.html,

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    Gated Linear Attention Transformers with Hardware-Efficient Training

    URLhttps://arxiv.org/abs/2312.06635. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence?,

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    HellaSwag: Can a Machine Really Finish Your Sentence?

    URLhttps://arxiv.org/abs/1905.07830. Biao Zhang and Rico Sennrich. Root mean square layer normalization,

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  38. [38]

    URL https://arxiv.org/abs/1910. 07467. Jianyu Zhang and Léon Bottou. Memory mosaics at scale.arXiv preprint arXiv:2507.03285,

    work page arXiv 1910

  39. [39]

    URLhttps://arxiv.org/abs/2505.19488. 13 SUPPLEMENTARY MATERIAL A Results of pure SWA models In section 4.3 we hypothesize that the worse performance of SWAX models with long windows comes from the model utilizing the SWA layers instead of the xLSTM layers. To further confirm this hypothesis, we train a 1.4B pure SWA model with a window size of 2048 and co...

    work page arXiv 2048

  40. [40]

    benchmark is an extension of HumanEval (Chen et al., 2021), which is designed to evaluate the functional correctness of code generated by AI models. 14 model xLSTM SWAX SWAX parameters 7B 7B 1.4B train-time window NA 128p=0.9p=0.75p 90%=0.75p=0.5 2048 p=0.5p 90%=0.5 test-time window NA 128 2048 2048 2048 2048 2048 2048 2048 niah_single 61.20 62.43 58.9963...

    work page arXiv 2021

  41. [41]

    NIAH single and multikey results are the average overall all 3 sub-tasks for each

    p90% indicates annealing, i.e., only doing the stochastic window size for the first 90% of the training and then using a fixed window size of 2048 for the rest of training. NIAH single and multikey results are the average overall all 3 sub-tasks for each. • MBPP (Austin et al.,

    work page 2048

  42. [42]

    is designed to evaluate the code generation abilities of AI models, particularly for Python programming tasks. Common sense and general reasoning.We use benchmarks consisting of question-answer or multiple- choice questions designed to evaluate the commonsense reasoning abilities of AI models, particularly in the context of natural language understanding:...

    work page 2019

  43. [43]

    and TQA (Joshi et al., 2017). 15

    work page 2017