Verified, then logged.
Sun 22 Nov 2026 · 00:17 KST / 15:17 UTC한국어 | ENSearch /

Breaking01 / AI

Weight transfers improve maths reasoning in language models

The researchers reported larger gains when the transferred layers overlapped more with the region linked to reasoning.

An unoccupied media research workstation displays one fictional Chinese period-drama scene on one monitor and two alternative scenes side by side on another. AI-generated image

Researchers reported on 19 Nov 2026 that transferring weights from selected layers improved mathematical reasoning on average across three families of language models.

According to the paper’s abstract, the researchers used a method for analysing the organisation of the brain’s cortex to compare base models with counterparts trained for reasoning. The overall patterns were broadly similar, with changes concentrated in a continuous span of layers.

The researchers transferred weights, the internal values set during training, into models trained to follow instructions. They reported greater gains when the transferred layers overlapped more with the region linked to reasoning.

The researchers found that this region formed gradually as reasoning performance began to improve during training.

The researchers also found that models still solved more of the problems they had previously answered correctly after receiving weights from the reasoning-related region than after receiving weights from other layers.

The abstract did not disclose the size of the average improvement in mathematical reasoning.