# Jiaxuan Zou argues pretraining must inject multi-turn tool-use patterns for agents

> Original title: Notes on MAI, Beam, ML4 & Nemotron: pretraining(PT) can't just watch NLL, and CPT isn't just lr annealing. Inject multi-turn tool pattern...

- Source: X search: AI launch posts (introducing, just launched)
- Published: 2026-11-11T18:17:20.000Z
- Bad Influence: https://badinfluence.news/items/sc6xnrgoq28uax7pfgb7jc5sv
- Original: https://x.com/SmartPig_Joe/status/2120581656558121192

## Summary

Jiaxuan Zou argues that pretraining and continued pretraining for agents should not rely on static NLL alone. He says multi-turn tool-use patterns must exist in the training data so the model starts reinforcement learning from a higher score, and cites reports from MAI, Reflection AI, Mistral, and NVIDIA to support the claim.

## Post

Notes on MAI, Beam, ML4 & Nemotron: pretraining(PT) can't just watch NLL, and CPT isn't just lr annealing. Inject multi-turn tool patterns in CPT to raise the pre-RL start. Part of my recent big-picture thoughts after 3 months at pretrain data team. https://x.com/i/article/2120572005624033280