feedback-signals

This primer paper explores how reasoning models improve after training, arguing that effective reasoning data relies more on checkable training evidence than raw data size. It categorizes reasoning data by verification methods and emphasizes preserving messy agent data for learning signals.