Meridian Notes
The papers that mattered this week, and why.
A short weekly newsletter. Every item has the same four parts, so you can read it in five minutes and know exactly where to go deeper.
Four parts, every item
- New: what the paper does that was not done before, in one sentence
- Evidence: the result and the comparison that supports it
- Catch: the limit the authors state, or the one we see
- Who cares: who should read it, and which section
Each issue carries five to eight items, a link to that week’s Conversation and Brief, and a calendar of upcoming venues and deadlines.
Meridian Notes · one itemIllustrative format, fictional paper
A cheaper way to make small models reason over long problems
Example Paper Title · Venue 2027 · first author A. Researcher
NEW
A training method that lets a small model hold its accuracy on problems three times longer than it was trained on.
EVIDENCE
Better than the strongest baseline on four of five task families; an ablation shows most of the gain comes from the method, not the data.
CATCH
Tested up to mid-sized models only, and on English tasks.
WHO CARES
Teams serving small models on long documents. Read §4 and Table 4.
What Notes is not
- Not a press-release digest: we read the paper, not the announcement
- Not a hype cycle: an item with no catch is not an item
- Not sponsored editorially: any sponsor is a labelled line at the end