fly51fly

AI

In the authors' tests, changing opening tokens mainly helped models add or restyle information, not consistently leave it out.

8001234567890123456789

EL

LR

RP

13 Sources, 7d ago, updated 7d ago

AI

Random biases to output logits reportedly matched or beat full-weight perturbations on most reasoning tasks tested.

6901234567890123456789

T2

KD

KK

12 Sources, 9d ago, updated 9d ago

AI

A 109M JEPA-Text model is claimed to align with vision models better than 72B LLMs.

8901234567890123456789

YL

PI

KD

18 Sources, 8d ago, updated 8d ago

AI

A post links the 2026 paper, listing H Deng and colleagues as authors and Harbin Institute of Technology.

1301234567890123456789

FL

1 Source, 27d ago, first seen 27d ago

AI

A post links to an arXiv page for the 2026 computer vision paper, naming ENPC and École Polytechnique.

1401234567890123456789

KK

FL

2 Sources, 32d ago, first seen 32d ago

AI

A Stanford paper on how language models pick between context and memory is shared on X.

6501234567890123456789

NI

FL

2 Sources, 69d ago, first seen 69d ago