Training Agent Organizations from Nemotron-Super-120B: Avalon-M1-Mini Works
rosstaylor90 · x · 2026-10-23
Avalon asked whether 'agent organizations' can be trained with a simple approach. Starting from a Nemotron-Super-120B-A12 base with minimal SFT, then applying their RLCR recipe, Avalon-M1-Mini performs well—e.g. versus NVIDIA's non-OPD model trained on the same base.
They acknowledge an RLVR baseline on the same environments is still needed (compute-limited, likely next), and note OPD models are stronger at this scale, but the interest is RLCR as a from-scratch recipe without a strong teacher.
On safety: in scenarios like a (heavily simulated) collective cyberattack, whistleblowing, and human loss-of-control, refusal to participate in bad actions remains consistent—whether collective objectives bias agents toward the collective over humans is still being studied.
More from coding & agent
- First App Store rejection in years: reviewer crash traced to two fatalError from 2023 — rudrank · 2026-10-24
- Yacine: a fleet of near-free DeepSeek agents gives absurd leverage over the digital world — yacineMTB · 2026-10-24
- Lee Rob's 45-minute masterclass explains how always-on AI agents work — Saboo_Shubham_ · 2026-10-24
- Prompt engineering is 'dead', long live prompt engineering: a cyclist-recipe prompt shows the craft persists — chrisalbon · 2026-10-24
- Few-shot prompting trick: turn examples into a list the system completes — mmitchell_ai · 2026-10-24
- Rewriting the world in Rust: nearly 1T tokens and 50k+ commits per month — doodlestein · 2026-10-24