What Does GDP.pdf Teach — Beyond PDFs?

Impact 7 Surge AI Blog

Summary

  • Surge AI post-trained the Kimi K2.7 model on 1,465 tasks from the GDP.pdf companion dataset using one document at a time without tools.
  • On the held-out GDP.pdf benchmark, strict full-task success rose from 11.0% to 24.5%, improving its public leaderboard rank from 35th to 9th.
  • Performance also significantly improved on GDPval, a benchmark requiring workspace exploration, tool use, and multi-file artifact creation.
  • The trained model learned to gather comprehensive evidence before acting rather than prematurely building deliverables from partial data.
  • This behavior transfer shows that training on complex single-document evidence tasks builds general professional reasoning skills that apply across different environments.

Why it matters

Post-training AI models on complex document-based reasoning tasks can substantially improve general workspace problem-solving capabilities, proving that foundational analytical skills transfer beyond their original training domains.

Who should care