ICML 2026
D3
KaTeX
α‑entmax → AdaSplash
Interactive explainer
α‑entmax → AdaSplash
A touch-friendly tour from adaptive sparse attention to fast differentiable sparse attention (AdaSplash‑2). Drag α, watch tokens snap to zero, follow τ from bottleneck to 1–2 iterations.
sparse attn
iPad‑ready