GEAR open-source framework lifts agent benchmark pass rates by up to 13 points
Pale-Training-6944 · reddit · 2026-11-01
The team open-sources GEAR, a meta-learning-inspired "scikit-learn for RSI" (recursive self-improvement) framework, with code and a paper.
Core idea
- Frames RSI as three familiar problems: efficient sampling of candidate changes, reliable evaluation, and diversity preservation against premature convergence
- A task agent produces execution traces; a meta-agent uses those traces and eval feedback to improve prompts, tools, workflows and context management — separating search from execution and evaluation
- Ships a working GEPA variant with shared failure diagnosis and staged evaluation, aiming to make GEPA, RRSI and Harness² easy to implement and compare
Results (GPT 5.6 Luna, medium reasoning)
- AutomationBench Marketing pass rate: 27.0% → 40.0%
- AutomationBench Support: 17.0% → 25.0%
- Terminal-Bench 2.1: 52.8% → 62.9%
- At max effort on Marketing, Luna with the optimized harness hit 88.88% objective completion vs 84.08% for Astra + Codex; strict pass rates 53% vs 57%
- With the same optimized harness, Luna cost averaged $0.0568/task vs $1.1858 for Astra — 95.21% cheaper (excluding optimization/infra costs)
The team welcomes feedback on candidate evaluation and diversity preservation under limited budgets.
Related event: Open-Source GEAR Framework Automates Recursive Self-Improvement for Agents(2 posts)→
More from coding & agent
- AI-Driven Frame Rate Optimization Cuts Render Time by 9%-15% Across Six Scenes — AIandDesign · 2026-11-01
- Parcha founder: You may not need an agent after all — decision models vs workflows — blaizedsouza · 2026-11-01
- Claude Code agents burned 9.45M tokens a day reading their own notes — WWEtrump · 2026-11-01
- Open-source agent filters Reddit feed with cheap LLM, under 1 cent per 100 posts — Ill_Fisherman8352 · 2026-11-01
- Berkeley/Stanford paper: agents gain up to 39% by wiping memory and restarting from best result — alexcovo_eth · 2026-11-01
- 4516 contributions in a year: dev turns GitHub heatmap into a Halloween easter egg — Abhishekcur · 2026-11-01