MazeBench adds novelty-based loop stopping to keep long runs from wasting compute
xeophon · x · 2026-07-28
MazeBench’s authors show how they stop agents from wasting compute when they loop: they let runs continue for roughly O(M) tokens and terminate only when a novelty score says the agent is spinning in circles.
The chart in the post shows novelty steadily falling over the course of a run, illustrating why the loop detector is needed in long-horizon maze tasks.
Related event: MazeBench Open-Sourced: 3D Open-World Benchmark for Long-Horizon Planning(14 posts)→
More from coding & agent
- Coding agent runs are getting long — one dev suggests breathing, not parallel sessions — teropa · 2026-10-16
- Dev builds MCP layer so Claude Code can work on servers without ever holding SSH keys — NETMESAFE · 2026-10-16
- As coding agents run longer autonomously, one dev's tip: don't parallel-session, just breathe — teropa · 2026-10-16
- 'Refusing to code with AI in 2026 is like refusing Google in 2005' — prasenx · 2026-10-16
- Claude Code adds idle compaction to save 1h prompt cache on idle sessions — EricBuess · 2026-10-16
- Handling agents that crash mid-task: the hard part is the unknown outcome across the network — Responsible-Dare5139 · 2026-10-16