A new research framework wants AI agents to think before they run the experiment.
Researchers describe JANUS, a system that pairs an AI research agent with a co-evolving world model in a Predict-Verify-Evolve loop. Instead of blindly trying an idea and waiting on results, the world model scores each candidate hypothesis by forecasting how much it will improve on the current best answer, how much new information it will reveal, and how worth exploring it is. The agent then decides whether to run, refine, or scrap a branch before spending compute on execution, and once real results come back, JANUS turns that feedback into reusable skills for both the agent and the world model. In tests on MLS-Bench-Lite and on real research tasks, the researchers report JANUS beat existing baselines on both how fast it improved and how far it could ultimately get.
Autonomous research agents today mostly run on what the paper calls a Generate-Execute-Feedback loop - try something, wait, adjust. That's slow and expensive when every "something" is a training run or simulation, and the search space grows fast enough to swamp trial-and-error. JANUS bets that spending a little compute on prediction up front saves more on execution later - the same logic that makes code review cheaper than shipping broken code and finding out.
Call it a planning layer bolted onto the agent, less a leap forward than an acknowledgment that letting an AI flail through every possible experiment was never going to scale.