The terms that appear in many more new papers this week than in the eight weeks before, from several independent teams, grouped into topics and explained in plain language.
×3.3 compared with usual · 85 papers from 68 teams
New this week: often an event, a release, a data set or a deadline. The coming weeks will tell if it lasts.
Researchers are developing algorithms to find routes, identify bottlenecks, and connect selected points in networks, including networks whose links change over time. The increased activity spans related questions: how quickly answers can be updated, how much information must be stored, and how much approximation is acceptable.
×3.8 compared with usual · 19 papers from 19 teams
Already this frequent the week of 11 May 2026.
A model’s weights are the numerical settings learned during training; weight-space research studies how changing these settings changes its behavior. The examples suggest growing interest in more targeted adaptation and auditing, from improving image restoration to understanding what useful information fine-tuning adds.
×2.7 compared with usual · 30 papers from 29 teams
New this week: often an event, a release, a data set or a deadline. The coming weeks will tell if it lasts.
BF16 is a numerical format that reduces the memory needed to store and process neural-network values compared with standard 32-bit arithmetic. Its increased appearance fits ongoing efforts to make training and inference more efficient, while checking how numerical precision affects accuracy and practical abilities such as coding.
Each week, we count how many papers submitted that week (first versions) use each term of one to three words in their title and abstract; acronyms such as RAG or JWST count, and plurals with their singular. A term rises when its share of papers is at least twice its share over the eight weeks before, when so large a rise is very unlikely by chance (a log-likelihood ratio of at least 15), and when it comes from at least three teams, none signing more than half of the papers: one group posting a series is not a trend. An AI then groups the rising terms into topics and names them; it never decides what rises: the counts do.
New: the term was almost absent from these papers before.
Rising for several weeks: up three weeks in a row or more, not a one-week event.
Recurring: as frequent at this time last year: perhaps a yearly cycle, such as a conference deadline.
Comeback: as frequent some months ago, then less used.
Sudden rise: new this week: often an event, a release, a data set or a deadline.
The curve on each topic is the share of papers using its main term, per 10,000 papers, over the last six months. Counting the previous Monday-to-Sunday submission week starts on Tuesday at 06:00 UTC, after the normal Monday arXiv announcement. Delayed announcements may still be missing from the first count.
