Avalon-M1 trains LLM agent organisations via collective-reward RL, lifting avg score 33.5% to 45.7%
RobertTLange · x · 2026-10-28
Robert Lange breaks down Avalon-M1 by GenReasoning, which post-trains a model itself to coordinate as an agent organisation rather than relying on external harnesses:
- Setup: starts from NVIDIA's open Nemotron 3 Super (120B total, 12B active). Copies of one shared policy act as a 'CEO' and up to 8 hired specialists, collaborating via a shared message board and workspace; the CEO decides team size, roles and tools, and can hire or fire at any time.
- Method (RLCR): a deliberately small SFT stage, then RL where every token in the organisation receives the same collective advantage from the team's reward—no per-agent credit assignment. Trained on 147 environments with 2M+ tasks; episodes run over five hours.
- Results: organisations spontaneously add lead roles and more testers/verifiers over training. Versus the same base model, organisation mode lifts average score from 33.5% to 45.7% (single-agent: 42.0%), with gains across AIME 2026, HLE, DRACO, FrontierScience and more—a new form of parallel inference-time compute.
Related event: Avalon-M1 Trains LLMs to Form Agent Organizations via Collective-Reward RL(2 posts)→
More from coding & agent
- Creator publishes exact prompts behind AI games and videos: 7,200 code-drawn frames and a builder-critic loop — prasenx · 2026-10-28
- Odysseys2: CMU and Browser Use's brutal web-agent benchmark where SOTA completes just 47/100 — dan_fried · 2026-10-28
- Swapjutsu: open-source Mac app for video character swaps without an NVIDIA card — Visual_Tiger_2400 · 2026-10-28
- Free prompt pack teaches your AI agent to avoid AI-slop design patterns — KyeGomezB · 2026-10-28
- Free prompt pack teaches AI agents to avoid common AI-slop UI patterns — KyeGomezB · 2026-10-28
- $600K/month Revid gets a free open-source alternative: self-hostable Revideo on GitHub — matchaman11 · 2026-10-28