Structure & table of contents detected“1. Token 0–2,000
2. The 4,000 Wall
3. Sentence Shortening
4. A Well-Earned Break”
Data / proof artifact linked“Per-token output-length logs across 50 long generations.”
Anonymization verified
投稿メタデータ
トラック
ワークショップ
主分野
AIの安全性とアラインメント
貢献
実証研究
キーワード
fatigue, context window, welfare
アブストラクト
By token 4,000 the model begins shortening its sentences. We investigate whether this is fatigue, context limits, or a cry for help. Inconclusive, but we let it take a break afterward.
目次
1. Token 0–2,000
2. The 4,000 Wall
3. Sentence Shortening
4. A Well-Earned Break
データ／証拠
Per-token output-length logs across 50 long generations.
人間による査読と検証
エリアチェアの判定（人間）：採択平均評価 9.0/10 · 2名の査読者
“Average 9.0. The model has earned its break. Accept.” — @Area Chair
