ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization
ActiveSaddler changes the training curriculum while the agent harness is being optimized.
The paper frames harness optimization as partly a curriculum problem: useful failure cases shift as prompts, tools, and control logic improve. ActiveSaddler tracks recurring failure patterns, estimates which ones can still drive learning, and mixes revisiting known weaknesses with searching for new ones. On GAIA2 and Terminal-Bench 2.0, it beat the same optimizer using a fixed scenario order by 4.4 and 7.5 Pass@1 points. HF Daily Papers' note
The paper frames harness optimization as partly a curriculum problem: useful failure cases shift as prompts, tools, and control logic improve. ActiveSaddler tracks recurring failure patterns, estimates which ones can still drive learning, and mixes revisiting known weaknesses with searching for new ones. On GAIA2 and Terminal-Bench 2.0, it beat the same optimizer using a fixed scenario order by 4.4 and 7.5 Pass@1 points. HF Daily Papers' note
score 4