Heuresis: Search Strategies for Autonomous AI Research Agents Across Quality, Diversity and Novelty

Autonomous AI research promises to accelerate machine learning, but current agents need to go beyond writing code: they must explore ideas that are simultaneously performant, diverse, and novel. We introduce Heuresis, a framework that abstracts the research loop into composable primitives and lets the agent backend be held fixed while the search strategy varies. We implement six strategies - greedy refinement, MAP-Elites, Go-Explore, Islands, Curiosity, and Omni - and evaluate them across quality, diversity, and novelty on three ML research domains: nanoGPT pretraining, On-Policy RL, and Model Unlearning. Across 3,222 scored runs, search changes where ideas land on these axes, but does not expand the quality-novelty frontier: no idea is rated "Original", only a few reach "Minor Similarity" to prior work, and only one novel-side idea lands in a per-task top-10 by quality. We also find that reward hacking is a practical threat to autonomous search archives, with 40 confirmed fabrications across 1,628 scored runs before auditing. These results suggest that current LLM search systems lack a signal for deciding which novel-but-underoptimized ideas merit further investment.