Kimi K2.7 reportedly improved on tasks without PDFs after GDP.pdf companion-task training

Surge AI reports that post-training Kimi K2.7 on 1,465 GDP.pdf companion tasks improved its results beyond PDF work. Its GDP.pdf score rose from 11.0% to 24.5%; on GDPval, full-task success more than doubled and the 90% rubric pass rate rose 13.2 percentage points, including on tasks without PDFs. Based on an analysis of the model’s steps, Surge AI hypothesizes that the training taught it to gather more evidence before acting.