What Fits (Into Few Tokens) Doesn't Overfit: Compression and Generalization in ML Research Agents

The paper presents findings on the resilience of machine learning strategies against overfitting in benchmark-driven settings, specifically for LLM-driven research agents. It introduces two information bottlenecks—output compression and input compression—and demonstrates that high-performance models can be effectively reproduced using minimal prompts and one-bit feedback across various datasets, indicating that successful strategies are highly compressible. This has implications for practitioners, suggesting that efficient model exploration and validation can be achieved with reduced complexity, potentially improving the robustness and efficiency of model training processes.