Nvidia's VERA: alternating model training and skill edits unlocks half of agent gains
rohanpaul_ai · x · 2026-10-09
A new Nvidia paper, VERA, shows agents on long multi-step work improve most when you alternate between training the model and editing its skill files — training only one side leaves roughly half the gains on the table.
Key points:
- Most environments score only final results, hiding which step broke. VERA builds 9,000+ restartable sandboxes, scoring each step of a workflow against real files and logs.
- Those per-step scores decide whether the next fix goes into the model or its skills.
- On a medical research benchmark, a 9B agent scored 69.1 with both update types, vs 56.1 with skill edits alone and 43.3 with training alone.
Related event: NVIDIA's VERA Co-Evolves Agent Models and Harnesses(4 posts)→
More from coding & agent
- Dev reports GLM-5.3 tool calls appear completely broken in Cursor CLI — DanielLockyer · 2026-10-09
- 20 AI agent memory tools to know in 2026, sorted into five selection categories — MaryamMiradi · 2026-10-09
- scikit-learn creator on coding agents: 'They get you to do stupid things faster with more energy' — GaelVaroquaux · 2026-10-09
- Matt Pocock calls out Anthropic's plugin marketplace for stranding users on stale versions — mattpocockuk · 2026-10-09
- Alook gives coding agents inboxes and DMs, letting Claude Code hand work to Codex — rohanpaul_ai · 2026-10-09
- TypeSafe AI's Jev decision model claims 200x faster, 400x cheaper classification in agent loops — LangChain · 2026-10-09