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..

Neural Attention Search

Authors: Difan Deng, Marius Lindauer

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on both training a new transformer from scratch and fine-tuning existing large language models show that NAt S can efficiently reduce the KV cache size and the inference costs for the models while maintaining the models performance.
Researcher Affiliation Academia Difan Deng Leibniz University Hannover EMAIL Marius Lindauer Leibniz University Hannover L3S Research Center EMAIL
Pseudocode Yes Algorithm 1 NAt S forward pass on Flash Attention2, we mark the NAt S related operations with blue Algorithm 2 NAt S backward pass on Flash Attention2, we mark the NAt S related operations with blue Algorithm 3 NAt S KV Cache Updating Process
Open Source Code Yes The codes for NAt S implementation and experimental designs can be found under: https://github.com/automl/Neural Attention Search
Open Datasets Yes Yes, all the datasets used in this paper are open-sourced. We further provide a script on how to generate the training set for fine-tuning an LLM with NAt S under: https://github.com/automl/Neural Attention Search
Dataset Splits Yes PG-19 Language Modeling Benchmark [70] with four Nvidia H100-PCIe GPUs and evaluate it on the test sets of PG19 with a context length of 1024. ... We construct a new training dataset that follows the construction rules of Long Bench as a training set to fine-tune LLM models with NAt S. ... Overall, this dataset contains 7 000 instances with a maximal context length of 16 000.
Hardware Specification Yes with four Nvidia H100-PCIe GPUs ... fine-tune two long-context models (Llama-3.1-8B-Instruct [34] and Mistral-7Bv0.3-Instruct [42]) on two Nvidia H100 PCIe GPUs
Software Dependencies No We implement NAt S based on the Flash Attention 2 [18] implementation on triton. ... We adapted Ro PE [74] and RMSNorm kernel from Flash Infer [90] to accelerate the forward process.
Experiment Setup Yes We train all models with a context length of 1024 and a batch size of 480 (using gradient accumulation). We train them for 600 000 iterations ... We fine-tune two long-context models (Llama-3.1-8B-Instruct [34] and Mistral-7Bv0.3-Instruct [42]) on two Nvidia H100 PCIe GPUs for one epoch using Adam W [47, 59] with a learning rate of 2e-3 with a warm-up from 2e-4 in the first 20% steps and reducing back to 2e-4 in the last 20% steps. ... For a fair comparison, we set the sliding window size W as 256, the same as Duo Attention.