A new paper proposes letting Bayesian neural networks decide for themselves how many random samples they need before committing to a prediction.
Bayesian neural networks estimate uncertainty by running many Monte Carlo samples per input and averaging the results, but most implementations just pick a fixed sample count ahead of time, whether the input is a toss-up or obvious. The researchers use "confidence sequences" - a statistical tool that updates its error guarantee as each new sample arrives - to keep sampling only until the network can commit to an answer with a mathematically bounded error. That applies whether the task is picking the most likely class, estimating the full probability distribution, or clearing a probability threshold. Once the guarantee is met, sampling stops, and reported experiments show the method spending more samples on ambiguous inputs and fewer on easy ones, trimming overall inference latency versus a fixed budget.
Bayesian neural nets are prized for knowing when they don't know, but that honesty has always carried a flat computational tax: the same sample count for a clear case as for a coin-flip case. A method that scales effort to actual difficulty matters for anything where some inputs deserve a second look and others don't, like flagging uncertain medical scans or deciding when an autonomous system should hand control back to a human.
It's a preprint, not a peer-reviewed benchmark war - the efficiency gains are reported by the authors alone, and adaptive-compute tricks have a long history of looking great in papers and shrugging in production.