GPT-OSS-120B agents route 9 robots: correction step makes 40/40 runs collision-free, without it only 2/20 survive
peter_richtarik · x · 2026-10-22
In the experiment portion of his thread, Peter Richtárik uses GPT-OSS-120B agents to plan routes for 9 robots on a 16×16 grid (3 agents × 3 robots, collisions penalized at 100 each). With the proposed verification-and-coloring correction, all 40 team runs are collision-free with final gap 5.4–5.9; without it, only 2 of 20 runs avoid collisions and the gap explodes to 313.1. Single-agent and best-of-3 baselines stay safe (95–100%) but perform worse (gap 12.0/18.9), showing unchecked multi-agent merges fail catastrophically.
Related event: First Convergence Theory for AI Agent Collaboration via Block Optimization(6 posts)→
More from Research
- Recursive's Eureka Machine Built Coding Model Flint Almost Entirely With AI in Under 4 Months — _rockt · 2026-10-22
- Thinking Machines Ships StableReinforce, a Loss That Keeps Off-Policy RL From Collapsing — LiTianleli · 2026-10-22
- Kaggle puts up $50K to pack squares into squares — only 59 of 1,000 puzzles have proven optima — MeganRisdal · 2026-10-22
- RAX drives robots with a frozen LLM via pure exploration at 1/20th Opus cost — du_yilun · 2026-10-22
- Survey: representation learning and generative modelling are converging in four phases — AllThingsApx · 2026-10-22
- New arXiv paper: a neuron's 'coastline paradox' — measurement resolution shapes apparent neural computation — tyrell_turing · 2026-10-22