Microsoft's ActiveSaddler Adds Curriculum Learning to Agent Harness Optimization

Microsoft's new paper proposes ActiveSaddler, which applies curriculum learning to agent harness optimization by actively selecting training tasks based on unfixed failures, improving Pass@1 by up to 7.5 points and GAIA2 by 4.4.

2026-10-04 ~ 2026-10-05 · 3 related posts

• Microsoft's ActiveSaddler adapts agent harness training scenarios, boosting Pass@1 by up to 7.5 points — dair_ai · 2026-10-04
• Microsoft's ActiveSaddler picks training tasks by open failures: +7.5 on Terminal-Bench 2.0 — rohanpaul_ai · 2026-10-05
• Microsoft's ActiveSaddler: automated curriculum learning for agent harness optimization, +4-7pt Pass@1 — rohanpaul_ai · 2026-10-05