🧪 Test?View on arXiv
What You Can't See Is Still What You Learn: A Preregistered Sixty-Society Confirmation That Evidence Masking Drives Compositional Generalization
Not provided in the abstract
compositional generalizationevidence maskinglanguage models
2609.17637
Builder Relevance
1h ago70%
Abstract
The study confirms that restricting evidence visibility can enhance a system's ability to learn compositional tasks.
Reality Card
Core Claim
Evidence masking significantly improved accuracy in compositional tasks, with median paired differences of 0.846 and 0.859 on held-out compositions.
Method / Result
Masking improved accuracy on held-out two- and three-operation compositions by median paired differences of 0.846 and 0.859.
Limitations
The effect of usable role information remains unresolved, and the decomposition criteria for the filler condition were inconclusive.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.
No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.