LDLMIn construction

Meshchaninov et al. · May 2026Diffusion language model

Attention over both streams

Pinned implementation and paper LDLM @ e18c42a, not yet public · arXiv:2605.07933v1 · OpenWebText · 1024 · checkpoint, tokenizer and saved statistics not verified · How to read

Training Generation

Q[B, 12, T, 64]K[B, 12, 2T, 64]V[B, 12, 2T, 64]Query–key products Scale, softcap

50\tanh(S/50\sqrt{64})

Softmax over keys

A=\operatorname{softmax}(F)

A[B, 12, T, 2T]Weighted value sum O[B, 12, T, 64]

in: $Q$ [B batch, 12 heads, T positions, 64 head features] $K$ [B batch, 12 heads, 2T positions, 64 head features] $V$ [B batch, 12 heads, 2T positions, 64 head features]
out: $O$ [B batch, 12 heads, T positions, 64 head features]

Explanation

The fused kernel does not store the score and weight fields; the plate states its arithmetic step by step.

Sources2

Direct evidence

Both streams are normalized; queries come from one stream, keys and values from both, concatenated along positions; queries and keys are normalized per head; attention runs over all positions with softcap; the heads are merged and projected.

The implementation, pinned at revision e18c42acdfb7, not yet public

Direct evidence

The latent encoder: 12 heads of 64 features, query and key normalization, softcap 50, 6 layers, MLP multiplier 4, dropout 0.1, ε = 10⁻⁵.

The implementation, pinned at revision e18c42acdfb7, not yet public