A frozen small model reads 43,687 events with no error in 299 tests
The model only records each change, and a counter external to the model adds the changes in sequence.
Researchers show a method to monitor a condition through a long stream of events with a frozen language model. The weights of the model do not change. The model records each change, and an external integer register adds the changes in sequence. A frozen 0.8B model with a selector read each change in 299 of 299 windows of 43,687 events.
How it works
The team adds a selector to the frozen model, to read each change. With access only through an API, the model gives the changes as text. A confidence rule can flag an uncertain update or keep some candidates.
The numbers
On a clean machine, the same package got 298 of 299 new windows correct. This was below the target of 0.99 for the lower bound. A follow-up for one error gave 472 of 473. With the register, all 7 models that had 2 tests got all 5 documents correct. By reasoning only, 3 of 10 models did this. Cost for each document fell 55% for Sonnet 5.5.
What is not known
A regular expression with an executor also got all 15 fixed-format documents correct. On 500 free-form comments, a 4B reader found the recorded value in 68.8% of the comments. A baseline got 61.0%. The paper gives no test on company systems in use.
Insight: Plain text is the link that works
In October 2026, a paper found that a frozen model reads only a cache near the cache that it makes. It cannot use the cache of a different model. This paper uses only text between a frozen model and external code: the model records changes and code adds them. This reads as a move of correct work from costly reasoning to cheap code, and the buyer pays less.
Sources
Posted