2026-11-22#kaggleメダルリサーチ

第1部:人間向けの要約

1. 5コンペ横断・日次リソース配分表

優先度コンペティション残り期間 / 締切現在の公開上位水準推奨投稿配分本日の最重要フォーカス
1(最優先)Google – The Gemma 4 Developer Agent Competition残り11日(11/25ルール受諾・12/2最終)Public 0.29 / 銅圏 0.20(ResearchForge)1 / 1枠(上限1回/日)timeout_seconds=240 固定、ASTチャンク検索導入、未使用グラフツール削除
2Enveda CASMI 2026 – Molecule ID From Mass Spectra残り23日Public MRR 0.492+(CASMI26: Anchored Spectral Network)3〜5枠リーク回避2ランカーブレンド(0.65/0.35)+Adduct-shifted類似度+ICEBERG検証枠
3Google Quantum AI – Verifiable Quantum Circuit Compilation約2ヶ月(2027/1/20)Public 28.20〜33.07(Kaggle Code) / 手元 25.093〜5枠62-qubit invsqrt、17-route GL(2) dqi、Unlookup削減と20 mHa境界較正
4Name That LLM約2ヶ月(2027/1/13)Public 0.684〜0.702(0.702 Apex Style + GLM + Qwen LoRA)2〜3枠prompt_id 5-fold GroupKFold厳守、Prompt-centered twin+Balanced Sinkhorn割当
5Santa 2026 – The Elf Calculator約2ヶ月Public 132,954〜143,381 steps(Kaggle Code)2〜3枠(CPU並列)Addition Sequence / Cross Window初期化+DAG-LNS / CP-SAT部分再結合

2. コンペ別・最新上位解法とarXiv知見(一次ソース付き)

第2部:Antigravity / Geminiチャット コピペ用YAMLブロック

YAML

date: "2026-11-22"
priority_order:
  - gemma-4-developer-agent
  - enveda-CASMI26-molecule-id-mass-spectra
  - google-verifiable-quantum-circuit-compilation
  - name-that-llm
  - santa-2026

competitions:
  - comp: gemma-4-developer-agent
    metric: resolved_task_rate (Pass@1 on hidden tests, higher better)
    sub_limit: "1/day | 4x L4 GPU | vLLM W4A16 (31B/12B/E4B/E2B) | 32k ctx | 12h total budget"
    top_findings:
      - method: "arXiv: Timeout-based cascades use timeout as zero-cost difficulty signal, cutting GPU-hours 16.4% overall (33.5% easy) without accuracy loss"
        src: "https://arxiv.org/html/2611.24324"
      - method: "arXiv: Anytime stopping policy with vLLM prefix caching and strict per-command sandbox time limits"
        src: "https://arxiv.org/html/2611.22901"
      - method: "Two-sided soundness audit (116/129 valid tasks) + Dense AST-chunk retrieval (51% vs 22% graph) + 6-tool slim prompt + timeout_seconds=240"
        src: "https://www.kaggle.com/code/moranzavdi/from-local-validation-to-live-evaluation"
      - method: "Attention-frozen (dq=do=0) right-subspace projection + folded rank-8 LoRA + MS-SFT trajectory selection"
        src: "https://www.kaggle.com/code/moranzavdi/from-local-validation-to-live-evaluation"
    exp_queue:
      - p: 1
        hyp: "グラフツール定義削除と240秒コマンド上限固定でプロンプト税・ハング起因のスコア低下(0.18->0.13)を解消"
        action: "6コアツールのみ公開、timeout_seconds=240, max_tool_calls=28, temp=0.2、ASTチャンク検索を接続"
        gate: "85件の健全非HoldoutでPass@1検証 && 120タスク換算推論時間 <= 10.5h && ZIP内4ファイルSHA256一致"
      - p: 2
        hyp: "単一パス一括編集+最終編集後のみテスト実行で150行read_file漏れと2回目編集成功率0%(0/10)を回避"
        action: "ASTチャンク一括取得後にパッチを完了させ、最終編集後にテスト実行枠を予約するシステムプロンプトへ最適化"
        gate: "プロンプト追加バイト数 <= 150B && 31件Holdoutリークゼロ確認"
      - p: 3
        hyp: "未マージRank-64 LoRAのW4A16カーネル逸脱(20.1->11.0 tok/s)回避と早期提出ドリフト抑制"
        action: "43件の解決済健全軌跡のみでAttention凍結(dq=do=0) Rank-8 LoRAを学習し重み折込(folded)で推論"
        gate: "スループット >= 18 tok/s && 120タスク換算時間 <= 7.5h"
    avoid:
      - "Unmerged rank-64 LoRA on vLLM W4A16 (+90.5s/task overhead, triggers 12h timeout)"
      - "timeout_seconds >= 300 (hung pytest starves episode budget, drops LB to 0.12)"
      - "Exposing unused graph tools or verbose rules (+338B prompt drops LB 0.18 -> 0.13)"
      - "Treating single-run LB delta <= 0.01 as signal (byte-identical resubmissions vary by 0.01-0.05)"

  - comp: enveda-CASMI26-molecule-id-mass-spectra
    metric: Mean Reciprocal Rank (MRR@25 per molecule_id, tautomer-canonical InChIKey-14, higher better)
    sub_limit: "Code competition | CPU ~47min or GPU | test fixed to timsTOF (eV, 10 adducts)"
    top_findings:
      - method: "arXiv: ICEBERG forward fragmentation verification + targeted remasking/denoising for inference-time scaling on novel (Class-3) molecules"
        src: "https://arxiv.org/html/2604.16648v3"
      - method: "arXiv: Multi-scale sinusoidal m/z embeddings computed in float32 before fp16 weight casting preserves high-res MS2 structure"
        src: "https://arxiv.org/html/2207.02980v2"
      - method: "Leak-free Two-Ranker within-molecule rank blend (0.65 FP-enabled + 0.35 7-lib 2,250-mol simulated FP-free) + adduct-shifted library similarity"
        src: "https://www.kaggle.com/code/megayak/casmi26-two-rankers-one-engine"
      - method: "Tautomer-canonical InChIKey-14 deduplication of top-25 (~10% slot recovery) + COCONUT/ChEBI/LIPID MAPS candidate union"
        src: "https://www.kaggle.com/code/matterhorn3838/evenda-casmi-inferencesupernext"
    exp_queue:
      - p: 1
        hyp: "公開FP重みの既知ライブラリ過適合(MRR 0.80 vs 未知0.49)を2ランカー分子内ランク平均で緩和"
        action: "enveda-np-examples学習FP有ランカー(0.65)と7セット2,250分子FP無ランカー(0.35)のランクブレンドを実装"
        gate: "5-fold grouped CV (enveda-np-examples) でClass-1 MRR >= 0.87 && Class-2 MRR >= 0.65"
      - p: 2
        hyp: "異種Adduct間の中性脱離一致と互変異性体重複排除による有効Top-25枠拡大"
        action: "同一極性制限アナログ探索+Adduct-shifted類似度追加+提出直前にTautomer-canonical InChIKey-14重複排除"
        gate: "全molecule_idの25候補が固有InChIKey-14 && RDKit正規化SMILES妥当性100%"
      - p: 3
        hyp: "Class-3未知構造および上位候補リランクへのICEBERG順方向スペクトル整合スコア導入"
        action: "上位候補に対しICEBERG順方向フラグメント一致度とfloat32 Sinusoidal m/z埋め込み類似度を特徴量追加"
        gate: "学習/推論特徴量行列のバイト完全一致 && hidden test換算実行時間制限内"
    avoid:
      - "Training single GBDT ranker with public FPNet outputs on GNPS/MoNA/MassBank rows (severe leakage)"
      - "Cross-polarity analog matching (negative queries matched to positive spectra)"
      - "Submitting raw COCONUT SMILES without tautomer-canonical InChIKey-14 deduplication"
      - "Casting m/z to fp16 before sinusoidal embedding computation"

  - comp: google-verifiable-quantum-circuit-compilation
    metric: "Geometric mean of cost multipliers: (R_invsqrt * R_dqi * R_dfthcsos)^(1/3) (higher better)"
    sub_limit: "5/day | Deterministic submission.zip (.kmb raw_bitstream) | Fiat-Shamir salted SHA-256 verifier"
    top_findings:
      - method: "arXiv: kickmix open-source C++20/Python verifier in Qualtran (6e10 instr/s, Fiat-Shamir salted SHA-256 non-interactive fuzzing)"
        src: "https://arxiv.org/html/2610.23832v1"
      - method: "arXiv: Reversible Gauss-Jordan elimination quantum circuit design for DQI coherent syndrome decoding"
        src: "https://arxiv.org/abs/2504.18334"
      - method: "arXiv: Windowed arithmetic unlookup cost reduced 66% asymptotically + merged lookup-addition and address bypassing"
        src: "https://arxiv.org/html/2502.17325v1"
      - method: "invsqrt 62-qubit compaction (21.78x) + dqi 48-site cswap on 17-route GL(2) (65.23x) + dfthcsos anisotropic 2D calibration to 20 mHa frontier"
        src: "https://www.kaggle.com/code/tuannm3812/verifiable-quantum-circuit-compilation-portfolio"
    exp_queue:
      - p: 1
        hyp: "dfthcsosの異方性2Dスケーリングとlookup-addition統合でFrobenius残差20mHa境界までlambda_effとToffoli数を圧縮"
        action: "W->sW*W, H2,I->sH*H2,Iの2D較正とWindowed unlookup最適化を適用し、<=128 qubits/loader制約下で最小化"
        gate: "公式verifier(full mode) 0 error && ||H1||_F, ||H2||_F <= 19.95 mHa && E_gap >= 3.914898 Ha"
      - p: 2
        hyp: "invsqrtの62-qubitレジスタ圧縮とBlock-4条件付きアンコンピュート分割による倍率向上"
        action: "最上位2位相測定ビット削除(62-qubit化)+q25マージ+Block-4 uncompute split適用"
        gate: "worst-case relative error ratio < 0.9995 across 5 salts (incl. 4x shots)"
      - p: 3
        hyp: "dqiの48-site cswap読み出し書き換えとPrivate Salt耐性確保"
        action: "17-route affine GL(2)デコーダに48-site cswap書き換え適用、決定論的ZIP梱包後に5種Saltで検証"
        gate: "5 salts x 102,400+ shots で value_error == 0 && phase_error == 0 && SHA256一致"
    avoid:
      - "Truncating invsqrt QROM coefficient register from 17 bits to 16 bits (Horner error 1.000412 > 1.0, instant verifier fail)"
      - "Validating dqi on only a single baseline-v1 salt (secret LB salt failure risk)"
      - "Non-deterministic ZIP packaging (changes file SHA-256 and alters Fiat-Shamir test seed)"

  - comp: name-that-llm
    metric: Balanced Accuracy (macro-averaged per-class recall across 18 frontier LLMs, higher better)
    sub_limit: "Code competition | Dual-T4 GPU | ~4,000 test passages (40-250 words) | unseen prompts & shifted class priors"
    top_findings:
      - method: "arXiv: Model-specific authorship fingerprints transfer across domains; controlled same-prompt comparison isolates style from content"
        src: "https://arxiv.org/html/2609.38831v1"
      - method: "arXiv: Survey of closed-class LLM authorship attribution combining stylometric features with pretrained LM representations"
        src: "https://arxiv.org/html/2408.08946v3"
      - method: "Strict prompt_id 5-fold GroupKFold + Qwen3-8B/14B LoRA24 with prompt-centered twin (+0.161 OOF Balanced Accuracy over 4-view SVC)"
        src: "https://www.kaggle.com/code/xiewenwei29/name-that-llm-week3-resubmit/notebook"
      - method: "Multi-depth hidden states + stylometric residual correction + Balanced Sinkhorn Assignment + GLM likelihood (0.684-0.702 LB)"
        src: "https://www.kaggle.com/code/kirderf/name-that-llm-v15c"
    exp_queue:
      - p: 1
        hyp: "未知プロンプトのトピックバイアスを除去するPrompt-centered twinでOOF Balanced Accuracyを大幅改善"
        action: "prompt_id単位5-fold GroupKFold構築、LoRAロジットからテストプロンプトクラスタ中心を減算するcentered twin追加"
        gate: "paired prompt bootstrap (2,000回) 95% CI下限 > +0.02 && fold/adapterマッピングリーク0"
      - p: 2
        hyp: "Balanced Accuracy評価とクラス比率変動に対するBalanced Sinkhorn割当の優位性"
        action: "文体残差補正後のアンサンブル確率行列にBalanced Sinkhorn Assignment (またはHungarian) を適用"
        gate: "全5 foldでargmax比OOF一貫向上 && 18クラス全予測数 > 0"
      - p: 3
        hyp: "位置別文体ビュー(15%)とGLM尤度特徴のランクブレンドによる補完"
        action: "LoRAスタックに4ビューn-gram/句読点・位置別文体SVMとGLMソース条件付尤度をランク統合"
        gate: "Dual-T4推論時間 <= 8h && モデル/アダプタSHA256事前検証パス"
    avoid:
      - "Random K-Fold without grouping by prompt_id (severe topic leakage and CV/LB divergence)"
      - "Raw argmax decoding without class-balance calibration (heavy macro recall penalty)"
      - "Length or topic-keyword shortcuts (test passages vary randomly 40-250 words on unseen prompts)"

  - comp: santa-2026
    metric: Total arithmetic instructions across 1,000 puzzles (lower better)
    sub_limit: "5/day | 48 register slots | ops: +, -, *, exact / | max value <= 10^18, <= 4,096 instr, <= 128 KiB/program"
    top_findings:
      - method: "arXiv: Addition Sequence Algorithm (ASA) + Cross Window Method (up to 44.74% shorter than Binary, 9.51% vs Window) & O(k^2) Simplified Power-tree backward-adding"
        src: "https://arxiv.org/html/2207.13276v2"
      - method: "arXiv: Minimal register allocation via Strahler numbers & Tree Straight-Line Programs (TSLP) for shared subexpression reuse"
        src: "https://arxiv.org/pdf/2512.19060"
      - method: "D_48 register decision problem is NP-complete; target coverage persists after register overwrite; sources read before destination write"
        src: "https://www.kaggle.com/competitions/santa-2026/discussion/760326"
      - method: "Per-puzzle portfolio merge + 2-step Wider Dictionary Beam Search (143.3k) + DAG-LNS + CP-SAT/MILP recombine (132,954 LB)"
        src: "https://www.kaggle.com/code/amanatar/santa-2026-fast-validator-eda-sota-baseline"
    exp_queue:
      - p: 1
        hyp: "ASA/Cross Window法+2ステップ広域辞書Beam Searchでパズル別ポートフォリオの初期解を大幅短縮"
        action: "複数ターゲット共有のAddition Sequence初期化と2ステップ辞書Beam Searchを実装し、1,000パズル個別最小値を更新"
        gate: "全1,000パズルで公式条件(値<=10^18, 余り0除算, レジスタ<=48, 命令<=4096)100%パス && 総ステップ数厳密減少"
      - p: 2
        hyp: "被覆済ターゲットのレジスタ即時解放(Strahler最適割当)とDAG-LNS + CP-SAT部分再結合で上位高コストパズルを圧縮"
        action: "ターゲット達成直後のスロット再利用スケジューラと、高コスト上位200パズルへのCP-SAT局所DAG再合成+ピープホール最適化を実行"
        gate: "更新パズルの命令数が厳密減少 && ファイルサイズ <= 128 KiB/program"
    avoid:
      - "Locking registers of already-hit targets (coverage survives overwriting; locking wastes 48-slot budget)"
      - "Non-exact division or intermediate overflow > 10^18 (instant validation failure)"
      - "Overwriting all 1,000 puzzles from a single solver run instead of per-puzzle portfolio min-merge"

全5コンペすべてで site:arxiv.org 専用検索が機能し、YAMLの top_findings 先頭に検証済みarXiv論文URLが1〜3件ずつ入った状態で出力できることを確認しました。この形式で明朝 3:00 頃から自動配信されます。調整したい項目があればお知らせください。

コメント

タイトルとURLをコピーしました