[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-llm-benchmarks-not-enough-2026-zh":3,"article-related-llm-benchmarks-not-enough-2026-zh":31,"series-research-f25ed4f5-db61-4d8c-bc59-e80c93e27927":77},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":23,"views":27,"created_at":28,"published_at":29,"topic_cluster_id":30},"f25ed4f5-db61-4d8c-bc59-e80c93e27927","llm-benchmarks-not-enough-2026-zh","2026 年挑 LLM，別再把 benchmark 當答案","\u003Cp data-speakable=\"summary\">1548 code Elo 不足以在 2026 年選定一個 LLM。\u003C\u002Fp>\u003Cp>2026 年選 LLM，\u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 只能當篩選工具，不能當主要決策依據。Iternal 2026 年 7 月的整理把差異寫得很清楚：\u003Ca href=\"\u002Ftag\u002Fclaude\">Claude\u003C\u002Fa> Opus 4.6 以 1548 的 Arena code Elo 領跑 coding，GPT-5.4 在 OSWorld 拿到 75% 的 structured reasoning 與 computer use 表現，\u003Ca href=\"\u002Ftag\u002Fgemini\">Gemini\u003C\u002Fa> 3.1 Pro 在 GPQA 以 94.3% 領先抽象推理與多模態科學，Grok 4 則在 HLE 以 50.7% 取勝。這種分裂不是噪音，而是證據：你選到的往往只是某一張榜單的冠軍，不是整個工作流程的最佳解。\u003C\u002Fp>\u003Ch2>第一個論點\u003C\u002Fh2>\u003Cp>benchmark 最有價值的地方，是幫你先排除不合格的\u003Ca href=\"\u002Fnews\u002Fgpt-56-turns-openai-into-a-model-menu-zh\">模型\u003C\u002Fa>，而不是直接替你做決策。若你的場景有法規限制、內網部署、低延遲要求或資料隔離需求，很多 frontier 模型在比較分數之前就已經出局。Iternal 的排序也把這件事放在前面：先看 privacy、latency、cost，再看 intelligence tier，最後才讀 benchmark。這個順序是對的，因為一個違反硬限制的模型，再高分也不能上線。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783818161840-t0n4.png\" alt=\"2026 年挑 LLM，別再把 benchmark 當答案\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>還有一個更實際的原因，讓 benchmark 不該被神化。Iternal 指出，MMLU 上差 2 到 3 個百分點的\u003Ca href=\"\u002Fnews\u002Fhalf-price-ai-real-frontier-smarter-models-zh\">模型\u003C\u002Fa>，往往在這個指標上已經接近不可區分。換到真實工作流，prompt、工具調用、檢索、人工覆核一加進來，這點差距常常消失。也就是說，benchmark 能幫你縮小候選名單，但不能告訴你哪個模型真的能替團隊省時間。\u003C\u002Fp>\u003Ch2>第二個論點\u003C\u002Fh2>\u003Cp>真正決定成敗的，是任務匹配，不是抽象智力。2026 年的模型版圖已經把這件事攤開：Claude Opus 4.6 在 coding 領先，但 GPT-5.4 更適合 computer use 與 structured reasoning，Gemini 3.1 Pro 則更強於多模態科學與抽象推理。這不是矛盾，而是證明「最佳模型」本來就是一個錯誤問法，除非你先把任務定義清楚。寫 \u003Ca href=\"\u002Ftag\u002Fcode-review\">code review\u003C\u002Fa>、做研究助理、操作軟體介面，三者需要的 trade-off 本來就不同。\u003C\u002Fp>\u003Cp>開源陣營也一樣。Iternal 提到 MiniMax M2.5\u002FM2.7、GLM-5\u002F5.1、Kimi K2.5 這些模型，已經在 \u003Ca href=\"\u002Ftag\u002Fswe-bench\">SWE-bench\u003C\u002Fa> 等任務上逼近 frontier proprietary 模型。對要私有部署的團隊來說，真正要買的不是「最高分」，而是「在硬體、治理與資料駐留限制下最好的結果」。在這種情境裡，一個分數略低、但能本地跑、能控資料的模型，往往才是唯一合理選項。\u003C\u002Fp>\u003Ch2>反方可能怎麼說\u003C\u002Fh2>\u003Cp>先看 steelman。benchmark-first 的支持者並\u003Ca href=\"\u002Fnews\u002Fchina-winning-ai-cold-war-building-stack-zh\">不是在\u003C\u002Fa>偷懶，而是在追求可比較、可重現、可向管理層說明的決策方式。當一個模型在 SWE-bench、OSWorld 或 GPQA 上明顯領先，這確實代表能力更強，不只是行銷話術。對沒有時間做大量內部評測的團隊來說，選 benchmark 領先者是務實的預設值；而且，一個 benchmark 表現很差的模型，通常也不會在 production 突然變神。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783818161688-n20z.png\" alt=\"2026 年挑 LLM，別再把 benchmark 當答案\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這個論點有一半是對的：benchmark 必要。但它錯在把必要條件當成充分條件。production 成功還取決於 privacy、latency、成本、工具鏈、失敗模式與人工作業流程，這些 benchmark 幾乎都不測。Iternal 建議先 shortlist 3 到 5 個模型，再在自己的環境裡跑測試，這正是對 benchmark-first 最直接的反駁。benchmark 只能縮小範圍，不能替你簽約。\u003C\u002Fp>\u003Ch2>你能做什麼\u003C\u002Fh2>\u003Cp>如果你是工程師，先做 routing layer，再用一個真實任務測每個層級，不要拿單一 leaderboard 當總裁判；如果你是 PM，先寫下不可妥協條件，再把驗收標準綁到業務結果、延遲與失敗模式；如果你是創辦人，別再問「哪個 LLM 最強」，改問「哪個模型組合在品質、隱私、擴展性與成本之間最划算」。2026 年的正解不是押一個模型，而是建立一個有規則的模型組合。","2026 年選 LLM，benchmark 只能當篩選工具，不能當主要決策依據；真正該看的是任務匹配、部署限制與路由策略。","iternal.ai","https:\u002F\u002Fiternal.ai\u002Fllm-selection-guide",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783818161840-t0n4.png","research","zh","f15d1c6d-fdb6-4fe0-a671-f0450c038250",[17,18,19,20,21,22],"LLM","benchmark","model routing","task fit","production AI","2026",[24,25,26],"benchmark 只能用來篩選，不該直接決定 LLM 選型","任務匹配、部署限制與成本，往往比排行榜名次更重要","生產環境最好的架構通常是多模型路由，而不是單一模型押注",0,"2026-07-12T01:02:19.419242+00:00","2026-07-12T01:02:19.4+00:00","1581d9f8-c7f3-4316-ab55-3014f2589380",{"tags":32,"relatedLang":36,"relatedPosts":40},[33,34],{"name":18,"slug":18},{"name":17,"slug":35},"llm",{"id":15,"slug":37,"title":38,"language":39},"llm-benchmarks-not-enough-2026-en","Benchmarks should not pick your LLM in 2026","en",[41,47,53,59,65,71],{"id":42,"slug":43,"title":44,"cover_image":45,"image_url":45,"created_at":46,"category":13},"35378c9f-bc39-4cc0-b9e1-1ce4a746ba5b","rust-breaks-into-tiobe-top-10-zh","Rust 進入 TIOBE 前十的判讀筆記","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783816365723-zjl1.png","2026-07-12T00:32:23.969578+00:00",{"id":48,"slug":49,"title":50,"cover_image":51,"image_url":51,"created_at":52,"category":13},"48d2bd0a-8140-4bda-8a04-8523145b3197","ai-ransomware-still-needs-a-human-bottleneck-zh","AI勒索仍卡在人手瓶頸","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783800203353-1ntk.png","2026-07-11T20:02:52.10043+00:00",{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":13},"7773aa86-8276-49ff-8fce-a58639ce180c","benchmark-scientific-lineage-reasoning-zh","IG-Bench：測科學譜系推理","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783666980843-7o76.png","2026-07-10T07:02:31.047117+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":13},"7be813ff-524d-4445-a924-5c11002c87cf","opencof-video-generation-reasoning-zh","OpenCoF 讓影片模型逐幀推理","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783665174981-siq1.png","2026-07-10T06:32:29.458194+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":13},"5c1a4cfe-93d9-4619-8f04-67a10de880ce","uniclawbench-proactive-agents-live-tasks-zh","UniClawBench：活體任務測主動式代理","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783663371728-73eg.png","2026-07-10T06:02:24.502805+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":13},"a5119837-3405-44e3-86fd-cf3923096cb2","webassembly-to-c-rivals-native-runtimes-2026-zh","WebAssembly-to-C 仍能打平原生執行環境","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783647172829-3d39.png","2026-07-10T01:32:30.98686+00:00",[78,83,88,93,98,103,108,113,118,123],{"id":79,"slug":80,"title":81,"created_at":82},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":84,"slug":85,"title":86,"created_at":87},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":89,"slug":90,"title":91,"created_at":92},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":94,"slug":95,"title":96,"created_at":97},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00",{"id":99,"slug":100,"title":101,"created_at":102},"53a0dc54-0371-4e40-8d5e-74e94a73840c","geometry-aware-similarity-metrics-for-neural-representations-zh","超越距離測量：用微分幾何重新理解神經網路","2026-03-31T06:01:01.241968+00:00",{"id":104,"slug":105,"title":106,"created_at":107},"fee7d472-a775-4b1d-bbc2-1e8bca1bbf8b","on-the-fly-repulsion-in-the-contextual-space-for-rich-divers-zh","讓AI繪圖更有創意：用排斥力提升生成多樣性","2026-03-31T06:01:25.439673+00:00",{"id":109,"slug":110,"title":111,"created_at":112},"a9901203-d69b-447b-8854-15d14eab32b4","vision-aided-beam-prediction-cnn-eca-zh","影像輔助波束預測升級 CNN","2026-04-01T10:00:25.8073+00:00",{"id":114,"slug":115,"title":116,"created_at":117},"b55e7dd4-0a24-4b3d-804d-b0309a03f498","triple-band-fss-mimo-antenna-sub-6-ghz-zh","三頻 FSS MIMO 天線瞄準 sub-6 GHz","2026-04-01T13:18:36.857305+00:00",{"id":119,"slug":120,"title":121,"created_at":122},"f68290bd-e7f3-4b30-ba22-dcd4e0130a66","openclaw-1299-repos-eight-weeks-analysis-zh","OpenClaw 1299 個 Repo 的資料解讀","2026-04-02T05:03:45.208411+00:00",{"id":124,"slug":125,"title":126,"created_at":127},"ed9f80eb-eb02-4d35-8ad4-0ddf428751dd","beam-coherence-aware-combining-mmwave-mimo-zh","毫米波 MIMO 的雙階合併法","2026-04-02T05:27:26.897188+00:00"]