Single-occurrence fact
Factual information that appears only once in the training corpus. Such facts are considered to force models to perform reasoning-based integration and transfer, and they are a key data attribute affecting the scientific discovery capabilities of large language models.
Timeline (last 90 days)
The study defines factual information that appears only once in the training corpus as a "monofact".
The study proposes from the perspective of "missing probability mass" that monofacts are the key to triggering scientific discovery.
Experimental results show that the more monofacts there are, the more scientific discoveries the models make.
The positive correlation between monofacts and model discovery capability holds across different models, different sampling methods, and different inference budgets.
Through further experiments, the study confirms that the positive correlation between monofacts and discovery capability cannot be explained by training set diversity alone.
The cross-domain transfer of the discovery capability improvement brought by monofacts holds conditionally and requires pairing with data in the target domain.
All Facts (6)
The study defines factual information that appears only once in the training corpus as a "monofact".
50%UnverifiedThe study proposes from the perspective of "missing probability mass" that monofacts are the key to triggering scientific discovery.
50%UnverifiedExperimental results show that the more monofacts there are, the more scientific discoveries the models make.
50%UnverifiedThe positive correlation between monofacts and model discovery capability holds across different models, different sampling methods, and different inference budgets.
50%UnverifiedThrough further experiments, the study confirms that the positive correlation between monofacts and discovery capability cannot be explained by training set diversity alone.
50%UnverifiedThe cross-domain transfer of the discovery capability improvement brought by monofacts holds conditionally and requires pairing with data in the target domain.
50%