Machine learning aims at generalization, not memorization, and overfitting occurs when a model performs well on training data but poorly on unseen examples. Standard defenses use held‑out validation and test sets, but repeatedly checking and adjusting based on validation data can cause it to become part of the training process, leading to overfitting. In research, benchmark datasets are reused over years, creating a hill‑climbing loop that should theoretically overfit, yet progress remains genuine. The mystery is explained by a new study on LLM‑based research agents, showing that compression—preferring simpler models that fit fewer tokens—prevents overfitting. By resetting the agent’s memory and controlling information flow, the researchers demonstrate that agents maintain generalization despite repeated benchmark evaluation, offering a precise, Occam‑razor‑based account of why real‑world ML research does not collapse into overfit models.

The paper reveals that agents that compress their knowledge, i.e., produce concise explanations that fit within limited token budgets, are less prone to overfit validation benchmarks. This compression acts as an implicit regularizer, aligning model complexity with the underlying data distribution. Consequently, even after many iterative improvements on the same benchmarks, the agents continue to generalize to fresh test sets, mirroring the observed success of human‑driven ML research.

Key findings suggest that future research agents and model designs should incorporate stronger compression constraints to enhance robustness and ensure that performance gains translate to real‑world scenarios rather than merely benchmark memorization.