Tel Aviv University speeds AI training

Tel Aviv University researchers said on July 9 they had released a preprint and code for a mathematical method they say can speed up some kinds of AI training, molecular simulations and complex search problems. The approach, called “adaptive resetting,” changes when a process is restarted based on its state and elapsed time rather than using a fixed restart rule. The paper says that lets researchers test many restart strategies without brute-force reruns in each case. The work appears as a preprint on arXiv and builds on earlier peer-reviewed research from the same group. ### What exactly are the researchers changing? The arXiv preprint defines adaptive resetting as a state- and time-dependent restart rule for stochastic processes. In plain terms, the reset decision is allowed to depend on what the process has already done and how long it has been running, instead of restarting at a constant rate regardless of progress. The paper says that framework can recover key observables — including first-passage-time distributions, propagators and steady states — from a single set of trajectories generated without resetting, using a reweighting procedure. (arxiv.org) The authors say that avoids repeated brute-force sampling across many candidate protocols. ### Why does a restart rule matter for AI or search? Nature Communications published a related paper from the Tel Aviv University team in August 2025 showing that adaptive resetting could be used for “informed search strategies.” The journal summary says the framework enables faster informed searches and the design of complex non-equilibrium steady states. (arxiv.org) Ynet reported on July 7 that the same line of research could have applications in algorithms, machine learning, molecular simulations and complex biological systems. (arxiv.org) A Springer Nature “Behind the Paper” post by the authors said the reset rules can be optimized directly with neural networks, which links the method to machine-learning workflows rather than only to abstract statistical physics. (nature.com) ### How does this connect to molecular simulations? The Hirshberg Lab at Tel Aviv University lists a 2025 paper, “Accelerating Molecular Dynamics through Informed Resetting,” among its publications. That work reported that an informed resetting protocol can accelerate molecular dynamics and metadynamics simulations by resetting only when a condition is met, such as being too far from a target along a reaction coordinate. (ynetnews.com) The newer adaptive-resetting paper says the broader framework should help predict and design restart strategies for these systems without exhaustively simulating every option. The authors wrote that they “use it to discover efficient protocols for accelerating molecular dynamics simulations,” extending the method from search settings into chemistry and materials problems. ### What is actually new in the July 9 release? (hirshblab.sites.tau.ac.il) ArXiv lists the adaptive-resetting manuscript as a preprint rather than a peer-reviewed journal article, and arXiv says papers on its platform are not peer reviewed. The July 9 discussion around the work centers on the public release of results and code, not on a new journal acceptance that day. The underlying research itself is not entirely new. A version of the adaptive-resetting paper was already on arXiv in 2024, and a peer-reviewed version appeared in Nature Communications in 2025. (arxiv.org) What appears new this week is the renewed public push around code, benchmarks and broader AI-facing framing. ### What should readers treat cautiously? ArXiv says its papers are open-access preprints and are not peer reviewed by the platform. (arxiv.org) The researchers’ strongest published support so far is the 2025 Nature Communications paper on adaptive resetting and the 2025 Journal of Chemical Theory and Computation paper on molecular dynamics acceleration listed by the Hirshberg Lab. The team said further validation is underway, according to the context provided for this story. (arxiv.org) Readers can currently find the method in the arXiv preprint, the peer-reviewed Nature Communications paper, and the Tel Aviv University lab publication list, which names Barak Hirshberg and Shlomi Reuveni among the researchers involved. (arxiv.org 1) (arxiv.org 2)