Adversarial Context

In the adversarial-context task, RASP-Tuner incurred higher cumulative regret than CMA-ES and GP-UCB. The paper treats the adversarial-context result as central evidence that RASP-Tuner is not universally better than context-blind methods. The adversarial-context task used iid Gaussian contexts independent of a fixed quadratic objective. RASP-Tuner failed under the adversarial-context setting because retrieval attached prompts and best-parameter entries to random contexts. Offline screening and monitoring are recommended because retrieval can harm optimization when features have no signal.