PEARL forgets selectively, not blindly

🔥 What happened AWS researchers dropped PEARL, a bandit algorithm that resets only the feature directions where predictions went wrong, instead of forgetting everything uniformly when user preferences drift. On PRISM (1,500 users, 21 LLMs), PEARL beats the best neural bandit baseline by 2.5 points. 💡 Why it matters Sliding-window and discounted approaches forget blindly — PEARL beats them by 3 to 10 points. Its regret scales with the number of drifted users, not the whole population. The selection policy is a ~100K-parameter MLP running in milliseconds on CPU, with LLM weights frozen. ⚡ Our take Finally someone treats partial forgetting as signal, not noise. If you're still discounting uniformly, you're leaving measurable performance on the table — for pennies of compute.