CalibForge: Adversarial Task Synthesis for Agent Training
August 5, 2026
CalibForge is an autonomous system that synthesizes learnable terminal tasks by using verified solver behavior to calibrate task difficulty. It uses multi-solver disagreement or contrastive strong-pass/weak-fail relations to create 5,431 calibrated tasks.
HOW THIS AFFECTS YOU
●
builderYou can use this to generate more effective evaluation sets and training tasks for your agents.
●
researcherThis offers a systematic way to create better training distributions for agents.