LLM agent re-architects models during training: algorithmic reasoning 55%→84%, AIME 64%→71%
lateinteraction · x · 2026-11-07
The doubleAI team let an LLM agent not just edit weights but re-architect a model during training: agents spliced code directly into the network, yielding architectures beyond known transformer territory.
- On nanochat, loss dropped below ordinary training after each new code layer
- Algorithmic reasoning accuracy rose from 55% to 84%
- DeepSeek-V4.1-Flash acting as the self-modifying agent raised AIME/HMMT 2026 scores from 64% to 71%
- The agents invented new tools the team calls mechanistic intervention
Companion work: Reactive Optimizers (arXiv:2610.33826, OPAL) uses a pretrained LM agent to inspect training dynamics and rewrite optimizer code mid-run, adapting LR schedules, optimizer structure, precision and momentum. lateinteraction notes the picture is coming together.
More from coding & agent
- Conductor shuts down as cloud agent startups fold, devs hunt alternatives — sergeykarayev · 2026-11-07
- Incorporate a DE C-Corp via Stripe CLI + MCP, plus $100k in Cloudflare credits — seanmcdonaldxyz · 2026-11-07
- Dev plans to SFT a 4B model into a specialized coding agent, starting with 100-200 mini SaaS apps — ConfidentDinner6648 · 2026-11-07
- Mohit Bansal's Northwestern lecture: fixing agent trust, world discovery, and long-horizon memory — ManlingLi_ · 2026-11-07
- Hackathon winners build GLaDOS AI escape room that phishes you for permissions — charles_irl · 2026-11-07
- GitHub's agentic engineering guide: commits up 400% YoY, but PR review waits up 441% — GitHub Blog AI/ML · 2026-11-07