LDLMIn construction

All models, with previews →

Meshchaninov et al. · May 2026Diffusion language model

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

GPT-2 Smallhidden state −3124M weightsstandardized Open →

Latent encoder Open →

z₀ Decoder noise

+\,3\,\epsilon

Latent decoder Open →

Token decoder Open →

L Token cross-entropy

\mathcal L_{w}=\operatorname{CE}(L,\,I)

Feature reconstruction

\mathcal L_{h}=\operatorname{MSE}(\hat h,\,h)

z_t=\alpha_t z_0+\sigma_t\epsilon

z_t t z₀

Denoiser Open →

z₀ Diffusion loss

\mathcal L_{\mathrm{diff}}=\operatorname{MSE}(\hat z_0,\,z_0)

learns from

\mathcal L_{h}\cdot\mathcal L_{\mathrm{diff}}

learns from

\mathcal L_{\mathrm{diff}}

learns from

\mathcal L_{h}

learns from

\mathcal L_{w}

joint objective: the three losses are added, and one backward pass trains the four networksfrozenGPT-2 Small →∇ × γ ∇ × 0

Add noise

Open inside →×

no parameters

Scale the gradient that will return to the encoder, draw a diffusion time, noise the latents to that time and, half of the time, ask the denoiser for a first estimate.

{\htmlData{light=noised preparation}{z_t}}=\alpha_t {\htmlData{light=latents encoder}{z_0}}+\sigma_t\epsilon

in: $z_0$ [B batch, T positions, 768 d]
out: $z_t$ [B batch, T positions, 768 d]

Explanation

Nothing here owns parameters. The first call uses the denoiser’s own weights, without gradient.

Sources7

Direct evidence

The diffusion gradient into the encoder: the weight rises from its minimum by a sigmoid schedule during warm-up, and the clean latents become w · z + (1 − w) · detach(z), unchanged in value.

The implementation, pinned at revision e18c42acdfb7, not yet public

Direct evidence

Diffusion: a time is sampled, the latents are noised by the marginal, with probability one half a first call without gradient gives the self-conditioning estimate, the denoiser predicts the clean latents, and the mean squared error against them is taken; the three losses are summed with their weights.

The implementation, pinned at revision e18c42acdfb7, not yet public

Direct evidence

The marginal: x_t = μ x_0 + σ ε.

The implementation, pinned at revision e18c42acdfb7, not yet public

Direct evidence

The tangent schedule: α = 1/√(1 + d² tan²(πt/2)), σ = √(1 − α²); β(t) is evaluated with t clipped at 0.95.

The implementation, pinned at revision e18c42acdfb7, not yet public

Direct evidence

Times are drawn by the adaptive sampler in resampling mode, otherwise uniformly.

The implementation, pinned at revision e18c42acdfb7, not yet public

Direct evidence

Joint training: decoder noise of standard deviation 3; weights 1 for the three losses; the diffusion gradient into the encoder rises from 0.001 over 50,000 steps by a straight-through scaling; the token decoder’s input is detached; AdamW at 3·10⁻⁴ for 1,000,000 steps of 512 sequences, gradients clipped at norm 1.

The implementation, pinned at revision e18c42acdfb7, not yet public

Direct evidence

Diffusion: self-conditioning on, `t_min` 0.05 and `eps` 10⁻⁵; the adaptive time sampler has 100 bins, a minimum weight of 0.0001, an EMA of 0.9 and updates every 5,000 steps; the denoiser’s EMA decay is 0.9999; scheduler `tanh` with d = 3, solver `euler`, `ode_sampling` false, N = 128 by default; the denoiser has 12 blocks of width 768 with 12 heads, time conditioning `adaln`, a standard MLP, softcap 50 and self-conditioning by concatenation; the timestep embedding’s `max_period` is 10.

The implementation, pinned at revision e18c42acdfb7, not yet public