{"slug": "staple-static-sparse-upcycling-of-pretrained-checkpoints-via-embedding-modules", "title": "STAPLE: Static Sparse Upcycling of Pretrained Checkpoints via Embedding Modules", "summary": "STAPLE, a recipe built on STEM for integrating static sparse memory into pretrained LLMs, improves over continued dense training on OLMo-2 1B by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP, according to the arXiv paper arXiv:2610.12599v1. STAPLE retains FFN up-projections and adds token-indexed embeddings as a residual pathway, outperforming static sparse baselines STEM and an Engram-style model and beating an MoE upcycling baseline on average score while decoding 2x faster with less than half its GPU model memory. A host-offloaded implementation keeps generation throughput within 1.5% and training throughput within 3% of the dense baseline while caching only the ~5% most frequent embeddings on GPU.", "body_md": "arXiv:2610.12599v1 Announce Type: new \nAbstract: Sparse computation can scale the parametric capacity of LLMs without proportionally increasing active computation, but pretraining sparse architectures remains expensive and data-hungry. Existing sparse upcycling methods mainly target Mixture-of-Experts models, where expert homogeneity, sample inefficiency, and system overheads can limit practical scalability. Static sparse memory offers predictable access and token-specific updates that can help address these challenges, but naively integrating it into pretrained computation yields poor returns. In this work, we introduce STAPLE, a recipe built upon STEM for integrating static sparse memory into pretrained models. Whereas STEM replaces FFN up-projections with token-indexed embeddings, STAPLE retains them and adds the embeddings as a residual pathway, preserving pretrained computation while learning complementary token-specific information in a scalable manner. On OLMo-2 1B, STAPLE improves over continued dense training by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP. Average performance increases with static capacity, and improvements persist on the larger OLMo-2 7B backbone. STAPLE also outperforms static sparse baselines such as STEM and an Engram-style model, and achieves a higher average score than an MoE upcycling baseline while decoding 2x faster with less than half its GPU model memory and measured GPU energy per token in our throughput-oriented decoding evaluation. Importantly, despite introducing billions of parameters late in training, our host-offloaded implementation keeps generation throughput within 1.5% and training throughput within 3% of the dense baseline, while caching only the ~5% most frequent embeddings on GPU. These results show that static sparse memory can provide a sample-efficient, scalable, and sustainable path for expanding pretrained models without sparse pretraining from scratch.", "url": "https://wpnews.pro/news/staple-static-sparse-upcycling-of-pretrained-checkpoints-via-embedding-modules", "canonical_source": "https://www.machinebrief.com/news/staple-static-sparse-upcycling-of-pretrained-checkpoints-via-5a75", "published_at": "2026-10-12 04:00:00+00:00", "updated_at": "2026-10-12 04:06:50.433794+00:00", "lang": "en", "topics": ["large-language-models", "machine-learning", "ai-research", "ai-infrastructure"], "entities": ["STAPLE", "STEM", "OLMo-2 1B", "OLMo-2 7B", "arXiv", "HumanEval", "MBPP", "GSM8K"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/staple-static-sparse-upcycling-of-pretrained-checkpoints-via-embedding-modules", "markdown": "https://wpnews.pro/news/staple-static-sparse-upcycling-of-pretrained-checkpoints-via-embedding-modules.md", "text": "https://wpnews.pro/news/staple-static-sparse-upcycling-of-pretrained-checkpoints-via-embedding-modules.txt", "jsonld": "https://wpnews.pro/news/staple-static-sparse-upcycling-of-pretrained-checkpoints-via-embedding-modules.jsonld"}}