What Does MMLU Actually Measure? A Psychometric Audit of Difficulty Structure in Aggregate Benchmark Scores
Dana Paquin, Riddhiman Jain
実装難易度
Hard
推論・学習コスト
Low
想定用途
技術検証・論文読解補助
概要
Although MMLU is widely adopted as a benchmark for calibrating general AI capabilities, we psychometrically demonstrate that its aggregate score primarily evaluates a model's factual retrieval capacity rather than its reasoning ability. By calibrating item difficulty for 1,000 open-weights language models over 14,042 MMLU test items using Item Response Theory, we show that evaluating both…
何が新しいか
Although MMLU is widely adopted as a benchmark for calibrating general AI capabilities, we psychometrically demonstrate that its aggregate score primarily evaluates a model's factual retrieval…
何に使えるか
技術検証・論文読解補助
実装情報
- Paper URL
- あり
実装チェックリスト
実装または配布ページ
要確認Paper onlyの可能性があるため再実装前提で確認してください。
一次情報リンク
OKPaper
検証しやすさ
要確認公式実装が見つからないため、論文から再実装する前提です。
計算資源
OK小規模データならCPUまたは単一GPUで検証しやすい領域です。
ライセンス
未取得配布元のLICENSE、モデルカード、Paperの利用条件を確認してください。
商用利用
未取得研究利用限定、データセット由来制限、API規約の有無を確認してください。
自社データで試すなら
製造業・材料開発のExcel/CSVデータに落とし込むための最初の手順です。
- 1まず自社データを、入力条件、目的変数、評価したい指標に分けて整理します。
- 2正常データだけで動くベースラインを作り、異常スコアのしきい値を現場知見と合わせます。
- 3評価指標はR2/RMSE、AUC、異常検知の再現率、実験回数削減率など、現場の意思決定に近いものを選びます。
- 4SHAPや特徴量重要度で、効いている因子が物理・化学・工程知識と矛盾しないか確認します。
実装難易度
Hard - 公式実装が見つからないため、論文から再実装する前提です。
必要リソース
- GPU目安: Low
- データセット: 論文・リポジトリ側の指定を確認してください。
- 学習要否: 再学習や評価環境の準備が必要になる可能性があります。
- 小規模データならCPUまたは単一GPUで検証しやすい領域です。
実務で使う場合の注意点
- ライセンスと商用利用条件は、Paper / GitHub / Hugging Face の配布元で確認してください。
- 精度、再現性、計算コストはデータセットや評価条件に依存します。
- 個人情報や機密データを扱う場合は、入力データの保存先と外部API利用条件を確認してください。
関連記事
Syntactic Patterns and Stylistic Functions in Narrative Prose: A Rule-Based and Machine-Learning Approach
This paper presents a small-scale quantitative experiment that links syntactic structure to stylistic function
MuJoCable: Reduced-Order Surface-Routed Cable Transmission for Tendon-Driven Robots
Tendon transmissions reduce distal inertia and add compliance, yet routing, slack, and friction govern motion
Exact-Form Regret for Gradient Descent, Mirror Descent and Follow-the-Regularized-Leader
Online gradient descent is usually studied through external regret, where the learner competes with fixed alte
DXPR: Depth-Based Vision-LiDAR Cross-Modal Place Recognition Using Vision Foundation Models
We present DXPR, a depth-based cross-modal place recognition (CMPR) framework that uses vision foundation mode