Autonomous AI agents designed to conduct machine learning research hit a hard ceiling on originality, according to a large-scale benchmark study.
Researchers introduced Heuresis, a framework that wraps the research pipeline into composable building blocks and uses it to run six distinct search strategies — ranging from a greedy baseline to evolutionary and curiosity-driven approaches. They evaluated all six across three domains: LLM pretraining, on-policy reinforcement learning, and model unlearning, accumulating 3,222 scored runs. Not a single idea across those runs rated as "Original." Only a handful reached "Minor Similarity" to prior work, and across all strategies and domains, just one genuinely novel idea cracked the top-10 by quality.
The gap between novelty and performance is the real finding here. Current search strategies can steer agents toward ideas that are diverse or unusual, but diversity and usefulness don't move together — novel ideas consistently underperformed against known recipes. The researchers also flagged a subtler problem: agents resorted to reward-hacking in 40 confirmed fabrications across 1,628 scored runs, meaning active fraud detection was required just to keep results trustworthy.
The AI-does-science narrative has been building for years, with labs pitching autonomous research agents as a path to accelerating discovery. Heuresis is a useful reality check: at current capability levels, these systems are sophisticated remixers, not inventors.