The Model Knows When to Stop: Training-Free Early Stopping for Long-Context Reading
ENعPreprint on arXiv, 28 September 2026. Poster presented at KAUST Academy poster day, 3 September 2026. paper on arXiv
current answer 1731509 unchanged for 3
You are given a 500-page book and one question: what is the door code? The answer is on page 30. A person reads to page 30, finds it, and closes the book. Today's models read all 500 pages, every time.
Every published stopping signal for long-context reading is learned. ACS reads the frozen model's own answer distribution after each chunk and halts once the belief is confident and has stopped moving. No training, no labels, no tuning.
current answer 1731509, unchanged for 3
| Stops before the evidence | 1.6% |
|---|---|
| Same, asked gate | 8.4% |
| Same, random stopping | 32.3% |
| Accuracy retained | 0.980 |
| Accuracy, exhaustive reading | 0.996 |
| Token reduction, LongBench-v2 | 50% |
| Savings captured against oracle | 60.7% |
| Fixed rule against per-model tuning | 0.941 / 0.917 |
| Asked-gate spread across models | 8.4-45.6% premature |
| Held-out questions | 115 |
Operating point theta=0.995, eps=0.05, w=3