A new technique lets a single AI model shed excess baggage on the fly, tailoring itself to whatever job it is handling at that exact moment.
Mixture-of-Experts models save computation by activating only a few expert subnetworks per token, but every expert still has to sit in memory, which caps how big a model can be deployed. Existing pruning methods fix that by calibrating one mask offline and applying it to every request afterward, regardless of the task. Researchers found this static approach breaks down in multi-agent systems, where a single backbone model juggles many different roles and tasks that each lean on different experts. Their fix, called Dynamic Expert Pruning, trains a lightweight predictor once on workflow transcripts, then reads an agent's own system and task prompts to generate a custom expert mask in a single forward pass, with no per-task calibration required.
That prompt-reading trick matters because multi-agent pipelines, where one model plays planner, coder, and critic in turn, are becoming a default way companies deploy LLMs in production. A pruning method that adapts per request, rather than locking in one subset of experts for everything, could let teams run noticeably leaner deployments without sacrificing accuracy, particularly when they want to keep only a small number of experts active.
The results come from the researchers' own benchmarks across several model sizes and architectures, not a live production system. Whether the gains survive messier real-world prompts and actual latency budgets is the test that still needs running.