{"slug": "periodic-weak-spots-phase-sensitivity-from-chunked-kv-cache-compression", "title": "Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression", "summary": "A paper submitted to arXiv on 28 Sep 2026 reports that chunked KV-cache compression, which shrinks long-context inference memory and attention costs by compressing windows of consecutive tokens at a fixed stride, introduces \"phase sensitivity\" — retrieval accuracy that varies with a token's position relative to compression-window boundaries. In large open-weight models using such compression, long-context retrieval accuracy can differ by up to 40 percentage points across phases, a gap that average benchmark scores conceal. The authors pretrain transformer families from scratch across multiple KV-compression designs to reproduce the effect, and causal interventions show attention components specialize asymmetrically by source phase, leading them to argue models with chunked KV-cache compression must be evaluated across compression phases rather than on averages alone.", "body_md": "# Computer Science > Machine Learning\n\n  [Submitted on 28 Sep 2026]\n\n# Title:Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression\n\n[View PDF](https://arxiv.org/pdf/2609.36322)\n\nAbstract:Chunked KV-cache compression reduces the memory and attention costs of long-context inference by compressing windows of consecutive tokens into fewer cache entries at a fixed stride. Such compression also introduces a new positional coordinate: a token's phase, or its position relative to compression-window boundaries. We uncover a systematic asymmetry in models using such compression: the same information can be easy to retrieve at one phase and difficult at another. We call this periodic variation in retrieval performance phase sensitivity. In large open-weight models with such compression, long-context retrieval accuracy can differ by up to 40 percentage points across phases, revealing periodic weak spots that average benchmark scores can conceal.\n\nTo investigate this behavior, we pretrain a family of transformers from scratch across multiple KV-compression designs, reproducing phase sensitivity across the variants. Mechanistic analysis using causal interventions in these models reveals phase specialization: different attention components contribute asymmetrically to retrieving information at different source phases. We further analyze idealized retrieval models, showing how gradient flow dynamics may favor sharp phase specialization. Evaluating models with chunked KV-cache compression thus requires measuring across compression phases: high average accuracy can coexist with systematic positional failures.\n\n### References & Citations\n\nLoading...\n\n# Bibliographic and Citation Tools\n\nBibliographic Explorer \n\n*(*[What is the Explorer?](https://info.arxiv.org/labs/showcase.html#arxiv-bibliographic-explorer))\nConnected Papers \n\n*(*[What is Connected Papers?](https://www.connectedpapers.com/about))\nLitmaps \n\n*(*[What is Litmaps?](https://www.litmaps.co/))\nscite Smart Citations \n\n*(*[What are Smart Citations?](https://www.scite.ai/))\n# Code, Data and Media Associated with this Article\n\nalphaXiv \n\n*(*[What is alphaXiv?](https://alphaxiv.org/))\nCatalyzeX Code Finder for Papers \n\n*(*[What is CatalyzeX?](https://www.catalyzex.com))\nDagsHub \n\n*(*[What is DagsHub?](https://dagshub.com/))\nGotit.pub \n\n*(*[What is GotitPub?](http://gotit.pub/faq))\nHugging Face \n\n*(*[What is Huggingface?](https://huggingface.co/huggingface))\nScienceCast \n\n*(*[What is ScienceCast?](https://sciencecast.org/welcome))\n# Demos\n\n# Recommenders and Search Tools\n\nInfluence Flower \n\n*(*[What are Influence Flowers?](https://influencemap.cmlab.dev/))\nCORE Recommender \n\n*(*[What is CORE?](https://core.ac.uk/services/recommender))\nIArxiv Recommender\n\n*(*[What is IArxiv?](https://iarxiv.org/about))\n# arXivLabs: experimental projects with community collaborators\n\narXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.\n\nBoth individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.\n\nHave an idea for a project that will add value for arXiv's community? [**Learn more about arXivLabs**](https://info.arxiv.org/labs/index.html).", "url": "https://wpnews.pro/news/periodic-weak-spots-phase-sensitivity-from-chunked-kv-cache-compression", "canonical_source": "https://arxiv.org/abs/2609.36322", "published_at": "2026-10-17 15:11:13+00:00", "updated_at": "2026-10-17 15:32:57.954232+00:00", "lang": "en", "topics": ["machine-learning", "large-language-models", "ai-research", "ai-infrastructure"], "entities": ["arXiv", "KV-cache", "transformers"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/periodic-weak-spots-phase-sensitivity-from-chunked-kv-cache-compression", "markdown": "https://wpnews.pro/news/periodic-weak-spots-phase-sensitivity-from-chunked-kv-cache-compression.md", "text": "https://wpnews.pro/news/periodic-weak-spots-phase-sensitivity-from-chunked-kv-cache-compression.txt", "jsonld": "https://wpnews.pro/news/periodic-weak-spots-phase-sensitivity-from-chunked-kv-cache-compression.jsonld"}}