Where RL Will Take Search — Maximilian-David Rumpf, SID.ai

Somewhere between 30 and 50 percent of an agent's tokens get spent on searching, almost all of it up front, before any of the work you asked for. Maximilian David Rumpf treats that number as the whole opportunity. Handing a task to an agent rather than a query to a search engine roughly doubles the odds of finding the right documents, but it costs a hundred to a thousand times more and takes minutes instead of milliseconds. His diagnosis of why the classical alternative cannot close that gap is the sharpest part. A traditional pipeline rewrites the query, hits a backend, reranks, and returns, which means every decision was frozen at design time and every question receives the same fixed budget of compute. The reranker can often tell that the results it is holding do not answer the question. It has no way to act on that. It returns them anyway. What accumulates instead is a long tail of failure, patched with edge cases that can never be exhaustive. Rumpf argues search is now following a path we have watched twice already, in computer vision going from hand written edge detection to narrow detectors to general models, and in chess going from a machine full of human authored rules to a system that learned its own. Search suits reinforcement learning unusually well because the reward is verifiable, you either found the correct document or you did not, and the environment is grindable at thousands of attempts per second. His results show a specialized model landing about twenty times faster than a frontier model on the same task, five seconds against two minutes, at roughly one hundredth the cost. Speaker info: - https://x.com/maxrumpf - https://linkedin.com/in/maximiliandavid - https://maxrumpf.com Timestamps: 0:00 - Agentic search is better, and far more expensive 2:11 - Where the classical pipeline breaks 3:27 - Machine design beats human design, again 5:08 - Why search is an ideal target for RL 6:46 - Twenty times faster, a hundred times cheaper 7:36 - Keeping bad results out of the main context