List your company

Commonsense Reasoning on Winogrande first 1000 examples (standard)

48.9Accuracyk=64 + sink
48.79648.82348.8548.877Apr 20, 2026

Evaluation results

MethodLinks
k=64 + sinkModel size=300M, Tokens=5B, Precision=FP16, Decoding=greedy, Mode=zero-shot, Examples=1000, Depth Registers=k=64, Attention-sink loss=included2026.0448.9
k=0 (baseline)Model size=300M, Tokens=5B, Precision=FP16, Decoding=greedy, Mode=zero-shot, Examples=10002026.0448.8
k=64 (DR only)Model size=300M, Tokens=5B, Precision=FP16, Decoding=greedy, Mode=zero-shot, Examples=1000, Depth Registers=k=642026.0448.8