EdgeBench Pt 2: AI Agent Strategy Optimization Boosts Unit Test Score 6x

karminski3 · x · 2026-07-06

The tester continued sharing details of a 35B model playing the game DCSS within the EdgeBench framework. The initial bot logic was crude, only capable of basic attacks and simple resting. The high-scoring version patched in handling for edge cases like getting stuck in menus, confirmation prompts, and being entangled/confused, alongside more refined environmental scanning.

A key improvement was the resting strategy: rather than simply resting when "no enemy is in melee range," the high-scoring version actively scans surrounding monsters to assess environmental safety. This strategic shift alone boosted one test case score from 21 to 126. The entire process visually demonstrates how AI agents can translate failure feedback into strategic improvements in long-horizon tasks—the core focus of the EdgeBench framework.

Related event: ByteDance Releases EdgeBench to Evaluate Long-Horizon Agent Evolution(4 posts)→

Original post →

More from coding & agent

coding & agent channel →