{
  "id": 7052425,
  "title": "When a Zero-Parameter Cache Overtakes a Transformer",
  "url": "https://urgent.news/2026/09/13/when-a-zero-parameter-cache-overtakes-a-transformer",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-13T03:00:00.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/megapixel99/when-a-zero-parameter-cache-overtakes-a-transformer-37c6"
  },
  "original_language": "en",
  "account": "A zero-parameter cache overtakes a transformer in terms of performance when dealing with longer documents. An experiment was conducted with six systems, each with identical positions, using text cut into documents of varying lengths. The systems were compared on their ability to predict the next word in a sequence. The results showed that a zero-parameter cache, which does not require any trained parameters, outperformed a 1.43 million-parameter transformer at different document lengths. Specifically, at 60 tokens, the cache achieved 0.103 top-1 accuracy, while the transformer achieved 0.152. As the document length increased to 250 tokens, the cache improved to 0.172 top-1 accuracy, while the transformer dropped to 0.143. At 1000 tokens, the cache had a significant advantage with 0.213 top-1 accuracy, compared to the transformer's 0.149. This demonstrates that a count table over the current document, with no parameters and minimal construction cost, can surpass a large transformer model when dealing with longer documents. The crossover point where the cache outperforms the transformer is early, occurring below roughly 100 tokens. The results indicate that the transformer's parameters do not contribute significantly to performance beyond a certain document length, while the cache continues to improve with increasing document size.",
  "summary": "The question this series exists to answer, put in a form that can be measured: a transformer sees a fixed 64-token window, a count table over the current document sees the whole document, and as documents get longer, how much of the trained model does the free mechanism replace? An earlier experiment had established that the document cache is worth up to +0.106 top-1 on a long coherent document,…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}