Long Horizons: Why generalization is the real benchmark
A long-horizon solver is not finished when it runs for hours. The real benchmark is generalizing coherent work, recovery and handoffs across different models.
Lynn Cole / 标签
agent evaluation
A thread through the work.
Training competence and engineering taste separately
Recursive training improved correctness while weakening engineering taste. A controlled learning-rate ablation and a failed promotion gate changed the next experiment.
Can play-training teach engineering taste?
A small Qwen training run tests whether play, critique and revision can change engineering taste while preserving correctness. The results guide the next replication run.