Square-Root Trick Beats Attribution Curse

🔥 What happened Researchers from Google and collaborators show you can estimate all leave-one-out attributions with O(√log(1/ε)/ε) model calls instead of n. The key move: apply forward differentiation to the backpropagation circuit, not the forward one. 💡 Why it matters For a 10,000-token prompt, that's a drop from 10,000 evaluations to a few hundred at the same error bound. Prompt debugging and retrieval attribution suddenly become practical instead of academic. ⚡ Our take A rare paper that's both asymptotically elegant and immediately useful. If you're still brute-forcing LOO values, you're burning compute for nothing.