Topic
Curriculum Learning
2 episodes
-
Harness S: Learning from Sibling Harnesses for Generalizable Agent Training
Asteria and Draco unpack Harness-S as a practical answer to a stubborn agent-training problem: harness improvements usually help only inside the improved harness. Their take is that the paper's real contribution is not 'many harnesses are better than one,' but a training recipe for borrowing specialized execution support only on failures, then distilling that behavior back into a single policy that still works under the original base harness.
-
Axion 1.0: Open Agent Foundation Models, Simplified
ByteDance's Axion 1.0 demonstrates that a three-stage training pipeline—two lightweight SFT passes followed by RL—can produce agent foundation models competitive with frontier commercial systems, while releasing the complete reproducible recipe including data construction, verification, and training code.