MDM-Prime: Unified Diffusion & Physics Framework

• MDM-Prime is a multifaceted framework that combines discrete diffusion generative modeling, Z′ anomaly solutions in B physics, and next-generation detector instrumentation.
• It employs partial masking with sub-token granularity and binary encoding via index shuffling to achieve a tighter variational bound and superior compute efficiency.
• The framework also links radiative neutrino mass generation and minimal dark matter models with precise nuclear instrumentation, demonstrating broad practical applications.

MDM-Prime denotes several distinct frameworks across high-energy physics, nuclear instrumentation, and discrete generative modeling. The following account provides a comprehensive overview of MDM-Prime in these contexts, with a focus on masked diffusion for discrete data, Z′-mediated B anomaly models, massive radiative dark matter and neutrino mass, and advanced focal-plane detector instrumentation.

1. MDM-Prime in Discrete Diffusion Generative Modeling

MDM-Prime primarily refers to a family of partial-masking schemes augmenting Masked Diffusion Models (MDM) for discrete data. Standard MDM generates or reconstructs token sequences by learning a denoising path reversing a forward process that masks tokens according to a time-dependent kernel: $$q_\alpha(x_t^i\mid x_0^i) = (1-\alpha_t)\,\delta_m(x_t^i) + \alpha_t\,\delta_{x_0^i}(x_t^i)$$ with $\alpha_t$ a strictly decreasing schedule, $m$ the mask, and $x_0$ the data sequence. The standard scheme processes fully unmasked or masked tokens, which leads to frequent "idle steps"—sampling transitions that make no change to the sequence—causing computational inefficiency and a loose evidence lower bound (ELBO) for training (Chao et al., 24 May 2025).

MDM-Prime introduces partial masking by mapping each token to $\ell$ invertible sub-tokens, each individually masked according to the same schedule. This substantiates a family of intermediate states between fully masked and fully unmasked, leading to a denser, finer-grained diffusion process, reduced redundant computation, and a tighter variational bound on the negative log-likelihood (Chao et al., 24 May 2025, Chao et al., 17 Mar 2026). Formally, for a token alphabet of cardinality $V$, an invertible subtokenizer $f_\ell$ generates $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$ for $b \approx V^{1/\ell}$, and the model operates at the sub-token level.

MDM-Prime's ELBO is given by: $$\mathcal L_{\rm vb}^{(\ell)} = \int_0^1 \frac{\alpha'_t}{1-\alpha_t} \mathbb E_{q_\alpha(y_0, y_t)} [\log p_\theta(y_0|y_t)] dt$$ where increasing $\alpha_t$0 (granularity) monotonically tightens the bound, provided subtoken entropy is sufficient (Chao et al., 24 May 2025, Chao et al., 17 Mar 2026). The architecture requires only minimal modifications to standard Transformer-based diffusion: input embedding concatenates sub-token representations, and the output head scores only valid (compatible with observed partial mask) sub-token-tuple candidates.

2. Binary Encoding and Index Shuffling: MDM-Prime-v2

The evolution to MDM-Prime-v2 resolves two central issues. First, prior instantiations lacked principled guidance for selecting $\alpha_t$1, making the trade-off between expressivity and computational load nontrivial. Second, common tokenizers, such as Byte-Pair Encoding (BPE), cluster frequent tokens to low indices, resulting in non-uniform, low-entropy sub-token bits under base-$\alpha_t$2 representations, which degrade model training and likelihood estimation (Chao et al., 17 Mar 2026).

MDM-Prime-v2 adopts (1) maximal granularity ($\alpha_t$3, $\alpha_t$4) so every token is a binary vector, and (2) a one-time random permutation (index shuffling) $\alpha_t$5 of token indices before binary encoding to maximize bit entropy. This combination yields the tightest theoretically possible variational bound among all invertible subtokenizers, and ensures bit marginals close to Bernoulli$\alpha_t$6. The implementation is precomputed: for each token, shuffle index then binary encode, with inverse mapping for evaluation.

Empirical results on OpenWebText demonstrate a perplexity of 7.77 under compute-optimal scaling, outperforming autoregressive (ARM) baselines at 12.99, earlier MDM at 18.94, and non-binary MDM-Prime at 13.41. Zero-shot commonsense reasoning accuracy with 1.1B parameters achieves 49.42%, surpassing peer models (Chao et al., 17 Mar 2026).

The scaling law fit follows a Chinchilla-style allocation, with optimal parameter count $\alpha_t$7 and dataset size $\alpha_t$8 under compute $\alpha_t$9: $m$0 With these methods, MDM-Prime-v2 attains $m$121.8$m$2 higher compute-efficiency than ARM and inherits order-agnostic sampling, robustness to model shape and embedding scheme, and architectural invariance (Chao et al., 17 Mar 2026).

3. Theoretical and Algorithmic Foundations

MDM-Prime's diffusion framework generalizes discrete denoising by introducing independently masked intermediate sub-token states, enabling an absorbing Markov chain over a combinatorial space of partially observed tokens (Chao et al., 24 May 2025). The forward transition is strictly absorbing—once a sub-token is masked ($m$3), it remains masked. The reverse transition samples from the model's posterior over possible original sub-tokenings for each masked position, consistent with the observed partial mask. For any sampled state, the model's output logits are zeroed for candidate tuples incompatible with already observed sub-tokens.

Crucially, the empirical tightness of the training objective is governed by sub-token entropy, which is maximized by index shuffling and binary encoding. The “carry-over” trick ensures efficient loss computation: once a sub-token is unmasked, its value is fixed and incurs no further cross-entropy loss.

This framework delivers two outcome improvements: reduction of idle steps (from 36.8% to 0.25% for $m$4), and improved sample likelihoods without recourse to autoregressive factorization; also, superior sample quality (e.g., FID for CIFAR-10 improves from 4.66 to 3.26 at $m$5 reverse steps) (Chao et al., 24 May 2025).

4. MDM-Prime in High-Energy B Physics: Z′ Models and Anomalies

MDM-Prime, or the Mixed-Down-Muon Z′ model, is a simplified framework describing new vector boson ($m$6) contributions to neutral current $m$7 anomalies. The relevant Lagrangian posits flavor off-diagonal couplings in the left-handed quark sector and a flavor-diagonal $m$8-only coupling among leptons (Allanach et al., 2019): $m$9 with $x_0$0, $x_0$1. Tree-level $x_0$2 transitions arise via $x_0$3. The dominant production at LHC proceeds via $x_0$4-quarks, with the $x_0$5 decaying to $x_0$6 or $x_0$7.

Collider bounds (ATLAS 139 fb$x_0$8, di-muon final states) and $x_0$9 mixing yield constraints on $\ell$0 for a given $\ell$1, summarized as: $\ell$2 No absolute lower bound exists for $\ell$3, as couplings can be tuned. The correlated constraints expose a restricted "allowed" region in the $\ell$4 plane, with perturbative unitarity imposing $\ell$5 at $\ell$6 TeV. The overall region is determined by the strictest among di-lepton resonance searches and $\ell$7-mixing constraints (Allanach et al., 2019).

5. Radiative Neutrino Mass, Minimal Dark Matter, and Lepton Flavor Violation

The MDM-Prime (R$\ell$8MDM) construct provides a unified framework linking radiative seesaw neutrino mass and Minimal Dark Matter. The field content includes a $\ell$9 Majorana fermion $V$0 and a $V$1 scalar $V$2, both stabilized by an accidental $V$3 symmetry. The neutral component $V$4 is cosmologically stable and constitutes a dark matter candidate with mass $V$5 TeV, determined by thermal relic abundance via SU(2) co-annihilation and Sommerfeld enhancement (Cai et al., 2011).

Radiative neutrino masses are induced at one loop through scalar and Yukawa couplings, with the mass matrix: $V$6 where $V$7 is a loop function and $V$8 the Yukawa matrix. Compatibility with light neutrino oscillation data requires at least two generations of $V$9.

Direct detection proceeds via electroweak loops, yielding a cross section $f_\ell$0 cm$f_\ell$1, below current limits but within reach of future ton-scale detectors.

Lepton flavor violation is predicted via $f_\ell$2 and $f_\ell$3–$f_\ell$4 conversion, with branching ratios potentially within reach of MEG II ($f_\ell$5 sensitivity $f_\ell$6) and future $f_\ell$7–$f_\ell$8 conversion experiments (PRISM $f_\ell$9 on Al). The dominant operator contributions are well-characterized and form a direct test of the Yukawa sector in R$$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$0MDM (Cai et al., 2011).

6. Instrumental Realizations: MDM-Prime Focal Plane Detector

MDM-Prime also pertains to the next-generation focal-plane ionization detector employing MICROMEGAS amplification at the Texas A&M University MDM spectrometer (Spiridon et al., 2019). The mechanical layout utilizes a 120 mm drift gap and a 256 μm amplification gap with a fine stainless-steel mesh. The anode plane consists of 4 rows × 7 columns of gold-plated copper pads, allowing for modular tiling to cover the full focal plane.

The effective gain is modeled as $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$1 with first Townsend coefficient $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$2 following the Diethorn model, yielding operational gas gains up to $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$3. Energy resolution is governed by the combination of Fano factor, gain variance, and electronic noise: $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$4 Beam tests demonstrated a two-fold improvement in energy-loss resolution for ions with $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$5 at $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$6–$$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$7 MeV: $$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$8–$$y_0 = f_\ell(x_0) \in \{ 0, ..., b-1 \}^{L \times \ell}$$9 compared to the original $b \approx V^{1/\ell}$0–$b \approx V^{1/\ell}$1, confirmed across multiple beams and gas pressures. Multi-row pad summing further improves statistical precision, and the system is robust against gain non-uniformity and electronic drifts via periodic calibration. The design recommendation for MDM-Prime is a modular, multi-row MICROMEGAS assembly for the full focal plane, with scalable readout architecture and optimized gas mixtures to extend isotopic resolution capability for heavy ions up to $b \approx V^{1/\ell}$2 (Spiridon et al., 2019).

---

MDM-Prime thus encompasses state-of-the-art advances in discrete diffusion model design for generative modeling, Z′ phenomenology for $b \approx V^{1/\ell}$3-physics anomalies, radiative seesaw dark matter mechanisms, and high-resolution nuclear instrumentation. In each paradigm, the adoption of partial masking, modularity, maximal entropy intermediate states, or high granularity drives technical performance and theoretical tractability.