<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>TAAF on ARC3 Dev Journal</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/tags/taaf/</link><description>Recent content in TAAF on ARC3 Dev Journal</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Tue, 01 Sep 2026 03:38:00 +0800</lastBuildDate><atom:link href="https://shih-yu-yeh.github.io/arc3-dev-agi-journal/tags/taaf/index.xml" rel="self" type="application/rss+xml"/><item><title>ARC3 V35.2 (LB 1.59): 修對了 Syntax，LB 沒變 — Modules 是 Net-Negative</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-09-01-arc3-v35-2-syntax-fix/</link><pubDate>Tue, 01 Sep 2026 03:38:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-09-01-arc3-v35-2-syntax-fix/</guid><description>修了 UnboundLocalError（0 個錯誤）。6 個 modules 正確跑（10 個 experiment events）。LB 1.59，跟 V35.1 broken 分數相同。Modules 確認是淨負面。</description></item><item><title>ARC3 V35.1 (LB 1.59): 6 Modules + 2443 UnboundLocalError — Closure Scope Bug</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-31-arc3-v35-1-closure-bug/</link><pubDate>Mon, 31 Aug 2026 01:24:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-31-arc3-v35-1-closure-bug/</guid><description>加 6 個 custom modules（ReasoningMemory、ExplorationTracker、HypothesisEngine、TransferMechanism、ReflectionRecovery）。step_env 因缺少 nonlocal 宣告產生 2443 個 UnboundLocalError。LB 1.59，比 V34 -0.65。</description></item><item><title>ARC3 V34 (LB 2.24): NOOA v3 — V31 Done Right，但仍輸 V24</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-30-arc3-v34-nooa-v3/</link><pubDate>Sun, 30 Aug 2026 05:11:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-30-arc3-v34-nooa-v3/</guid><description>修 V31 的 wiring（selective timeout 900 → 1200、每 level 最多 2 extensions）。456 events 提到 NOOA。LB 2.24，比 V33 +0.73 但仍比 V24 的 2.56 低 -0.32。</description></item><item><title>ARC3 V33 (LB 1.51): NOOA v2 — Memory + Supervisor Hysteresis</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-29-arc3-v33-nooa-v2/</link><pubDate>Sat, 29 Aug 2026 11:31:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-29-arc3-v33-nooa-v2/</guid><description>加 NOOA v2 modules（two-tier memory + supervisor hysteresis 15/20/3）。536 events 提到 NOOA。LB 1.51，比 V24 的 2.56 -1.05。Modules 跑了但 LB 退步。</description></item><item><title>ARC3 V28 (LB 1.72): Text-Only Ablation — 證明 Hidden Games 偏好 Image</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-27-arc3-v28-textonly-ablation/</link><pubDate>Thu, 27 Aug 2026 05:45:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-27-arc3-v28-textonly-ablation/</guid><description>V24 config，把 MULTIMODAL_CONTEXT 從 image 改成 text。LB 1.72，比 V24 的 2.56 -0.84。證明 hidden games 偏好 image modality。</description></item><item><title>ARC3 V25 (LB 1.42): 7 Grafts + MM8 — 為何加了 Grafts 反而退步</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-26-arc3-v25-grafts-regression/</link><pubDate>Wed, 26 Aug 2026 07:02:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-26-arc3-v25-grafts-regression/</guid><description>在 V24 上加 7 個 TAAF grafts（winframe、goalkeep、clockwatch、hudmask、clickmap、searchmap、lawbook）。但過程中掉了 FP8 KV、WBC、RE cap、NG。LB 1.42，比 V24 -1.14。</description></item><item><title>ARC3 V24 (LB 2.56): MULTIMODAL_UPSCALE=8 — 從 1.53 到 2.56 的決定性一擊</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-25-arc3-v24-mm-upscale8-winner/</link><pubDate>Tue, 25 Aug 2026 04:12:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-25-arc3-v24-mm-upscale8-winner/</guid><description>單一變數改動：MULTIMODAL_UPSCALE 4 → 8（256x256 → 512x512 vision）。Local mean 4.984（比 V23 +66%）。LB 2.56（比 V23 +1.03）。個人最佳。</description></item><item><title>ARC3 V23 (LB 1.53): Anim Bundle + Qwen3.8 Wheelhouse Fix</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-24-arc3-v23-anim-bundle/</link><pubDate>Mon, 24 Aug 2026 01:43:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-24-arc3-v23-anim-bundle/</guid><description>切到 jakobbrggen anim bundle，修 wheelhouse owner，退回 1-pass。LB 1.53，從 V21 的 0.00 恢復並創新個人最佳。</description></item><item><title>ARC3 V21 RESUBMIT (LB 0.00): 同一坑踩兩次 — Dummy Submission 追蹤</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-23-arc3-v21-resubmit/</link><pubDate>Sun, 23 Aug 2026 01:12:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-23-arc3-v21-resubmit/</guid><description>同一份 V21 kernel 重 submit。LB 還是 0.00。確認 dummy submission 問題不是 transient，是 2-pass 實作的真 bug。</description></item><item><title>ARC3 V21 (LB 0.00): 2-Pass Visible Updates — Dummy 提交事故調查</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-22-arc3-v21-dummy-submission/</link><pubDate>Sat, 22 Aug 2026 13:36:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-22-arc3-v21-dummy-submission/</guid><description>9 個 patch 全部 fire，加了 2-pass visible updates。LB 回 0.00，因為 submission.parquet 是 3411 byte 的 dummy。Pipeline 問題，不是 solver 問題。</description></item><item><title>ARC3 V19 (LB 1.17): P2+P3 Patches — 9 Patches 全 fire 但 LB 退步</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-18-arc3-v19-p2p3-patches/</link><pubDate>Tue, 18 Aug 2026 13:31:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-18-arc3-v19-p2p3-patches/</guid><description>加了 9 個 patch：FP8 KV cache、writable bundle copy、reasoning effort cap、noop guard 驗證、MULTIMODAL_UPSCALE=8。9 個全部在 stdout 驗證 fire，LB 卻從 1.43 掉到 1.17。</description></item><item><title>ARC3 V17 (LB 1.43): Qwen3.8-27B-FP8 + Visual Updates — 突破 1.10 天花板</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-17-arc3-v17-qwen38-visupd/</link><pubDate>Mon, 17 Aug 2026 07:12:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-17-arc3-v17-qwen38-visupd/</guid><description>從 Qwen3.6 升級到 Qwen3.8-27B-FP8。加 anim bundle 的 animation-awareness 與 noop guard。LB 1.43，比 V15 +0.63。首次突破 1.10 天花板。</description></item><item><title>ARC3 V15 (LB 0.80): DVM Fork — 從別人的 Fork 學到的細節</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-15-arc3-v15-dvm-fork/</link><pubDate>Sat, 15 Aug 2026 16:12:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-15-arc3-v15-dvm-fork/</guid><description>Fork dvm 的 kernel 觀察 vLLM 啟動處理。LB 0.80，跟 V13 在 hidden rerun variance 內。沒加新 patch，純觀察。</description></item><item><title>ARC3 V14 (ERROR): Temperature 0.3 Patch — 一個靜默失敗的紀錄</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-11-arc3-v14-temperature-patch-failed/</link><pubDate>Tue, 11 Aug 2026 10:25:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-11-arc3-v14-temperature-patch-failed/</guid><description>聲稱把 temperature 從 0.6 降到 0.3。Patch 驗證顯示 temperature 還是 0.6。Kernel 因其他原因 ERROR。Temperature 實驗白做了。</description></item><item><title>ARC3 V13 (LB 0.87): Tanaka Safety v1 + Registration Mode — 跳過 9 小時 Commit 的投機技巧</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-10-arc3-v13-registration-mode/</link><pubDate>Mon, 10 Aug 2026 12:48:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-10-arc3-v13-registration-mode/</guid><description>用 1 列 placeholder submission.parquet 跳過 9 小時 commit run。Hidden rerun 產出真實 LB 0.87。這是 pipeline 技巧，不是 solver 改善。</description></item><item><title>ARC3 V10 (LB 0.79): Rokaiya 純 Baseline — 故意退回原點的 Calibration Run</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-09-arc3-v10-rokaiya-baseline/</link><pubDate>Sun, 09 Aug 2026 01:09:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-09-arc3-v10-rokaiya-baseline/</guid><description>刻意拿掉所有 patch、graft、prompt addendum。LB 0.79。用來當作絕對地板，後續每個 patch 的貢獻都對這個基準量測。</description></item><item><title>ARC3 V9 (LB 1.10): YW8837 LB-1.17 Replica + AGI_8/AGI_9 Patches</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-08-arc3-v9-yw8837-replica/</link><pubDate>Sat, 08 Aug 2026 11:24:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-08-arc3-v9-yw8837-replica/</guid><description>重製 yw8837 的 LB-1.17 config，拿掉 program synthesis prompt。LB 1.10，比 V6 +0.35。確認 AGI_8/AGI_9 patches 單獨就能提供 V4 大部分的改善。</description></item><item><title>ARC3 V6 (LB 0.75): Context Budget 32768 → 49152 — 過頭的代價 -0.31</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-04-arc3-v6-context-budget/</link><pubDate>Tue, 04 Aug 2026 02:19:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-04-arc3-v6-context-budget/</guid><description>把 analyzer context window 從 32k 加到 49k tokens（+50%）。Local mean 0.58 預測會改善，LB 卻從 1.06 掉到 0.75（-0.31）。第一次具體看到 local mean 跟 LB score 沒有相關。</description></item><item><title>ARC3 V4 (LB 1.06): TAAF + Program Synthesis Prompt Triple-Patch</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-02-arc3-v4-triple-patch/</link><pubDate>Sun, 02 Aug 2026 02:01:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-02-arc3-v4-triple-patch/</guid><description>疊了三個 patch：program synthesis prompt、offline env_dir 修復、AGI_8/AGI_9。LB 1.06，比 V2 +0.19。首次改善。</description></item><item><title>ARC3 V2 (LB 0.87): Tufa Labs Duck Harness Fork — 我的 ARC3 起點</title><link>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-01-arc3-v2-tufa-duck-harness-fork/</link><pubDate>Sat, 01 Aug 2026 06:52:00 +0800</pubDate><guid>https://shih-yu-yeh.github.io/arc3-dev-agi-journal/posts/2026-08-01-arc3-v2-tufa-duck-harness-fork/</guid><description>Fork 自 samrishb/sam-solver（Tufa Labs milestone 1.21）。Stock Qwen3.6-27B-FP8，無任何 patch。LB 0.87 落在 Tufa 自報的 0.5-1.5 範圍內。</description></item></channel></rss>