{"id":7315,"date":"2025-08-19T05:27:00","date_gmt":"2025-08-18T20:27:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7315"},"modified":"2025-08-17T22:32:32","modified_gmt":"2025-08-17T13:32:32","slug":"provable-in-context-vector-arithmetic-via-retrieving-task-concepts","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7315","title":{"rendered":"Provable In-Context Vector Arithmetic via Retrieving Task Concepts\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Provable In-Context Vector Arithmetic via Retrieving Task Concepts\u00a0<\/strong>[53.7]<br>\u30af\u30ed\u30b9\u30a8\u30f3\u30c8\u30ed\u30d4\u30fc\u640d\u5931\u306b\u5bfe\u3059\u308b\u52fe\u914d\u964d\u4e0b\u306b\u3088\u308b\u975e\u7dda\u5f62\u6b8b\u5dee\u5909\u5727\u5668\u306e\u8a13\u7df4\u306f,\u30d9\u30af\u30c8\u30eb\u6f14\u7b97\u306b\u3088\u308b\u5b9f-\u30ea\u30b3\u30fc\u30ebICL\u30bf\u30b9\u30af\u3092\u3044\u304b\u306b\u884c\u3046\u304b\u3092\u793a\u3059\u3002 \u3053\u308c\u3089\u306e\u7d50\u679c\u306f\u3001\u9759\u7684\u57cb\u3081\u8fbc\u307f\u524d\u99c6\u4f53\u3088\u308a\u3082\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\u306e\u5229\u70b9\u3092\u89e3\u660e\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2508.09820v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2508.09820v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 13 Aug 2025 13:54:44 GMT)<\/li>\n\n\n\n<li>\u300cWe develop an optimization theory demonstrating that transformers with nonlinear softmax attention, MLP, layer normalization, and residual connections\u2014trained via Gradient Descent (GD) with cross- entropy loss\u2014can effectively perform factual-recall ICL in a vector arithmetic manner, grounded in empirically motivated data modeling. Our analysis shows that the transformer retrieves the high-level task\/function concept through attention-MLP, which, when combined with any embedded query vector within the same high- level task concept, yields the correct corresponding answer vector.\u300d\u3068task vector\u3092\u60f3\u5b9a\u3057\u305f\u7406\u8ad6\u7684\u7814\u7a76\u3002<\/li>\n\n\n\n<li>\u4e0d\u660e\u70b9\u306f\u307e\u3060\u307e\u3060\u591a\u305d\u3046\u306b\u601d\u3046\u304c\u3001\u7406\u8ad6\u7684\u7814\u7a76\u304c\u9032\u3080\u3053\u3068\u306b\u671f\u5f85\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[200],"class_list":["post-7315","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-in-context-learning"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7315","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7315"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7315\/revisions"}],"predecessor-version":[{"id":7316,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7315\/revisions\/7316"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7315"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7315"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7315"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}