NVIDIA Paper: Predicting Coding-Agent Performance From Base Models

An NVIDIA-led arXiv paper tackles picking base models for coding agents, noting six base models scored zero-of-six on SWE-bench Verified (five scored zero), underscoring how hard agentic performance is to predict pre-training.

2026-10-09 ~ 2026-10-09 · 2 related posts

• NVIDIA-led paper predicts coding-agent post-training gains from base models — heghbalz · 2026-10-09
• NVIDIA Paper Ranks Base Models by the One Edit That Fixes the Task — dair_ai · 2026-10-09