何が起きたか: What We Learned by Reproducing 2,200 papers from ICML

A hackathon reproduced 2,226 ICML 2026 papers using coding agents, finding 51% had verified claims and 23% had falsified or contested claims.

なぜ重要か: What We Learned by Reproducing 2,200 papers from ICML

Matters for researchers evaluating conference paper reliability and for those using AI agents to audit scientific work at scale.

ニュース

Hugging Face · 公開日 2026年8月13日 · 読了3分

何が起きたか: A hackathon reproduced 2,226 ICML 2026 papers using coding agents, finding 51% had verified claims and 23% had falsified or contested claims.
なぜ重要か: Matters for researchers evaluating conference paper reliability and for those using AI agents to audit scientific work at scale.
注意点: Reproducibility is adversarial, not binary. Independent teams reached opposite verdicts on same claims. Missing artifacts limited many reproductions.

この要約を音声で聴く

1つのパターン、1つのトレードオフ、1つの本番障害事例。エージェントシステムを構築する人のための短い週刊ブリーフィング。

週1回のメール、ワンクリックで購読解除できます。アドレスはブリーフィングの送信のみに使用します。