Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
Authors: Zihan Qiu, Zekun Wang, Bo Zheng, Zeyu Huang, Kaiyue Wen, Songlin Yang, Rui Men, Le Yu, Fei Huang, Suozhi Huang, Dayiheng Liu, Jingren Zhou, Junyang Lin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we conduct comprehensive experiments to systematically investigate gating-augmented softmax attention variants. Specifically, we perform a comprehensive comparison over 30 variants of 15B Mixture-of-Experts (Mo E) models and 1.7B dense models trained on a 3.5 trillion token dataset. Our central finding is that a simple modification applying an head-specific sigmoid gate after the Scaled Dot-Product Attention (SDPA) consistently improves performance. |
| Researcher Affiliation | Collaboration | 1Qwen Team, Alibaba Group 2University of Edinburgh 3Stanford University 4MIT 5Tsinghua University |
| Pseudocode | No | The paper only contains mathematical equations and diagrams (e.g., Figure 1) to describe the methodology, but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps. |
| Open Source Code | Yes | we also release related codes and models to facilitate future research. ... We provide a Pytorch version implementation for the main method in the paper. ... Any released code implementing our proposed gated attention mechanism will be released under the MIT License, allowing for open research and further development. |
| Open Datasets | Yes | We test the few-shots results on popular benchmarks, including, Hellaswag [31] for English, MMLU [32] for general knowledge, GSM8k [33] for math reasoning, Human Eval [34] for coding, C-eval [35] and CMMLU [36] for Chinese proficiency. |
| Dataset Splits | No | The paper mentions training models on 'subsets of a 4T high-quality tokens' and evaluates on 'diverse held-out test sets' and 'few-shots results on popular benchmarks'. However, it does not provide specific dataset split percentages or sample counts for its primary training corpus. While standard benchmarks have predefined splits, the paper itself does not detail these splits for its own experiments. |
| Hardware Specification | No | The paper mentions 'BF16 training' and discusses 'computational resources' in the Neur IPS checklist but does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for the experiments. |
| Software Dependencies | No | The paper states 'We provide a Pytorch version implementation for the main method in the paper.' but does not specify the version number for PyTorch or any other software libraries used. |
| Experiment Setup | Yes | We conduct experiments on both Mo E models (15B total parameters with 2.54B activated, 15A2B) and dense models (1.7B total parameters). ... All models use a scheduler that warms up to a maximum LR of 2e-3 in 1k steps and decays using cosine to 3e-5. We use a global bsz of 1024, comprising 100k optimization steps. ... For the 1.7B model trained on 400B tokens, we use a maximum LR of 4e-3 and a bsz of 1024. For training on 3.5T tokens, we increase the maximum LR to 4.5e-3 and the bsz to 2048. More detailed model architecture configurations are discussed in Appendix A.2. |