arXiv:2604. 19341v2 Announce Type: replace-cross Abstract: Scientific discovery often requires many cycles of proposing, testing, and refining candidate solutions.
Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.
arXiv:2606. 29082v1 Announce Type: cross Abstract: Would experience designing faster GPU kernels also help close in on a long-standing open mathematical conjecture?
The paper investigates Test‑Time Scaling (TTS) for large language models (LLMs) in the context of automated scientific equation discovery, an open‑ended task where models iteratively search candidate equations using observed data for feedback. It frames equation discovery as a unified iterative search that encompasses Best‑of‑N, sequential refinement, tree search, and evolutionary methods, and studies how compute allocation—particularly search width—affects performance under fixed budgets. Experiments on the LLM‑SRBench dataset show that increasing search width with more compute improves results, while other factors like population‑branching split and controller choice have smaller impacts, indicating that controlling exploration versus exploitation is key to scaling LLM‑based equation discovery.
ExplorationBench is a new benchmark designed to evaluate AI systems’ ability to conduct scientific exploration in verifiable alien worlds. It comprises two sandbox environments—AlienCode and AlienLogic—each containing discovery targets, tasks, flawed manuals, and tool‑call schemas that force systems to formulate hypotheses, design experiments, and iterate on results. Ten AI systems were tested, revealing that while the best performers can learn and apply unfamiliar rules, their progress varies across exploration trajectories and can even regress with continued exploration.
ExplorationBench is a benchmark designed to evaluate AI systems’ ability to conduct scientific exploration in verifiable alien worlds, where rules are executable and can be precisely checked. It consists of two sandboxes—AlienCode and AlienLogic—each offering discovery targets, tasks, flawed manuals, environmental feedback, and tool‑call schemas. The benchmark tests whether systems can generate new hypotheses, design experiments, and iterate on results, rather than merely recalling pre‑trained knowledge, and finds that top performers can acquire and apply unfamiliar rules, though performance varies across exploration trajectories.
MOSAIC‑SR is a new symbolic regression method that combines a pretrained Transformer with search‑based refinement. The Transformer generates multiple initial sketches, which seed searches that jointly recover equation structure and constants using scale‑aware optimization and symbolic repair. On the SRSD‑Feynman dataset and six other benchmarks, MOSAIC‑SR achieves the highest symbolic solution rate and ranks among the top two in predictive accuracy, even when irrelevant dummy variables are present.
arXiv:2609.15973v1 Announce Type: new Abstract: Foundation models have progressed from learning and reasoning over existing knowledge, to increasingly learning through action, tool use, and outcome f...