KARMA preprint reports advertising-agent gains without model retraining

A KARMA preprint reports improved advertising-agent performance by editing decision knowledge while keeping model weights and executable code fixed. Across 18 domain-model settings, the framework raised the average hold-out score from 30.19 to 39.15, a gain of 8.96 points (+29.7%). KARMA diagnoses errors, compares successful and failed runs, and edits six knowledge modules. A fixed validation gate retains edits only when measured gains reach a threshold; otherwise, it restores the previous knowledge state. The evaluation used three LLM families and cases sampled from approximately half a million real-world operation logs across six e-commerce domains.