A model that learned yesterday's news overnight
On 13 October 2026, Fintan learned a real day of news into a model's own weights overnight. The next morning another model wrote questions about it, and the model answered with nothing added to the prompt.
Every setting and pass mark was written down before the news existed, and the test ran once. Every question and every answer is on this page, including the misses.
Questions about 200 articles from BBC, The Guardian, NASA, the Federal Reserve and others, crawled after the settings were fixed. Marks come from a strict automatic scorer, which agreed with three blind judges on 99% of answers.
Four of the 266
Picked to show the idea. Every question is in the list below, including the misses.
Every question
The judges later marked 35 questions as broken or unfair. They stay in the list with a label. On the other 231, Qwen alone got 6.1% and with Fintan 87.9%.
40 test people each tell the assistant 12 things about themselves on day one, and 4 of those change on day two. Fintan learns each person's own knowledge overnight, and the next morning they ask in their own words. The people and their questions were written by a different model family. All 40 are here.
Fintan adds learning to whatever model it runs on. The same sealed day of news, with nothing tuned for each family:
| Model | Alone | With Fintan |
|---|---|---|
| Qwen2.5-7B | 7.9% | 87.6% |
| Llama-3.1-8B | 7.1% | 92.1% |
| Gemma-2-9B | 11.3% | 94.0% |
| Mistral-7B | 10.2% | 90.2% |
| Qwen2.5-14B | not run | 89.8% |
| Qwen2.5-72B (80 of the questions) | 7.5% | 95.0% |
Bigger models do better with it. Mistral uses its one-hour automatic setup. The 14B row uses the first night's settings. The 86.8% on the News tab is the sealed run itself; 87.6% here is the same day with the current night.
Also measured
On the standard lifelong benchmark for adding facts to a model (ZsRE, 1,000 facts, Llama-3-8B), Fintan scores 0.958. The best published method, MEMOIR, scores 0.93; it adds facts one at a time, and Fintan adds them in nights of 100. Questions unrelated to the new facts keep 99.4% of their predicted words.