Weighted Adjoint Matching finetunes flow matching models with time-dependent control cost, using high early weights for balance and low late weights for detail.
Researchers found that removing whitespace from five model instances reversed a positive interaction between GSM8K and Wikitext, proving target-token asymmetry masquerades as depth specialization.
Authors provide an exact three-point counterexample to Feldman's memorization lower bound using independent fair labels and a permutation-equivariant learner.
Researchers identified settings where training on a smaller random subset outperforms a larger unique set across Maxar, SATLAS, Waymo, nuPlan, Ego4D, ImageNet, and YFCC15M datasets.
SQUELCH, a router-free rank-one expert method, improves LLaMA-3.1-8B performance by averaging 8.6 active components per token.
Researchers tested eight evaluation aspects on 26 monolingual 2B models and three larger LLMs across 80 tasks in four languages.
Researchers introduced Echo, a 4-bit all-reduce method that achieves final loss less than 0.9% above 16-bit training.
Researchers propose SRCoD, a structure-response co-calibrated diagnosis framework that outperforms state-of-the-art methods for multimodal benchmark diagnosis.
