{"id":7832,"date":"2025-11-27T06:25:00","date_gmt":"2025-11-26T21:25:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7832"},"modified":"2025-11-27T06:20:40","modified_gmt":"2025-11-26T21:20:40","slug":"tidar-think-in-diffusion-talk-in-autoregression","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7832","title":{"rendered":"TiDAR: Think in Diffusion, Talk in Autoregression"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>TiDAR: Think in Diffusion, Talk in Autoregression\u00a0<\/strong>[59.9]<br>TiDAR\u306f\u3001Diffusion\u3067\u30c8\u30fc\u30af\u30f3(Thinking)\u3092\u30c9\u30e9\u30d5\u30c8\u3057\u3001\u6700\u7d42\u7684\u306a\u51fa\u529b(Talking)\u3092AutoRegressively\u306b\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3059\u308b\u30b7\u30fc\u30b1\u30f3\u30b9\u30ec\u30d9\u30eb\u306e\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3067\u3042\u308b\u3002 TiDAR\u306fAR\u30e2\u30c7\u30eb\u3068\u54c1\u8cea\u30ae\u30e3\u30c3\u30d7\u3092\u57cb\u3081\u308b\u6700\u521d\u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3067\u3042\u308a\u3001\u6bce\u79d24.71\u500d\u304b\u30895.91\u500d\u306e\u30c8\u30fc\u30af\u30f3\u3092\u63d0\u4f9b\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2511.08923v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2511.08923v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Thu, 13 Nov 2025 01:18:11 GMT)<\/li>\n\n\n\n<li>Diffusion model\u3068Auto regressive\u306e\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u300cWe introduce TiDAR, a sequence-level hybrid architecture that drafts tokens (Thinking) in Diffusion and samples final outputs (Talking) AutoRegressively &#8211; all within a single forward pass using specially designed structured attention masks.\u300d<\/li>\n\n\n\n<li>\u300cWe extensively evaluate TiDAR against AR models, speculative decoding, and diffusion variants across generative and likelihood tasks at 1.5B and 8B scales. Thanks to the parallel drafting and sampling as well as exact KV cache support, TiDAR outperforms speculative decoding in measured throughput and surpasses diffusion models like Dream and Llada in both efficiency and quality. Most notably, TiDAR is the first architecture to close the quality gap with AR models while delivering 4.71\u00d7 to 5.91\u00d7 more tokens per second.\u300d\u3068\u30b9\u30b1\u30fc\u30eb\u3059\u308b\u3053\u3068\u304c\u78ba\u8a8d\u3067\u304d\u3066\u3044\u308b\u306e\u304c\u3059\u3054\u3044\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":[114],"class_list":["post-7832","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-diffusion-model"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7832","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=7832"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7832\/revisions"}],"predecessor-version":[{"id":7833,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7832\/revisions\/7833"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7832"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7832"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7832"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}