Evolver turns agent exploration into reusable know-how: +4.4pp success, -24% tokens
omarsar0 · x · 2026-11-09
EvoMap researchers introduce Evolver (arXiv 2611.07595), a framework that converts transient agent exploration into persistent, reusable procedural experience stored as structured "Gene" assets — acquired, retrieved, and refined via feedback without touching model weights. On 320 paired runs across Terminal-Bench 2.1, SWE-bench Multilingual, SWE-rebench, and GAIA, Claude Code and Codex with Evolver improved task success from 83.13% to 87.50% (+4.38pp) while cutting per-run token consumption by 24.24%; among 265 pairs successful under both conditions, tokens still dropped 17.57%, showing efficiency gains independent of success improvements.
More from coding & agent
- Tencent's Terminal-RSI lifts agent success from 20.68% to 50.58% via recursive self-improvement — mark_k · 2026-11-09
- lattedit: free local-first LaTeX editor in the browser with a JSON API for agents — cneuralnetwork · 2026-11-09
- Perplexity Computer gains Cloudflare Security Audit skill that traces vulns entry-to-sink — jeffgrimes9 · 2026-11-09
- "Do the needful": the Indian English phrase that works surprisingly well on AI agents — moultano · 2026-11-09
- Self-described AI hater: Claude Code fixed a year-old problem in under an hour — alejandroll10 · 2026-11-09
- Three.js 'Hollow' demo lands as day 7 of a vibe-coding creative streak — thespite · 2026-11-09