MMProLong: Disambiguating Multi-Domain Methods

• MMProLong is a multi-domain research label that denotes four distinct technical constructions across biostatistics, numerical linear algebra, causal inference, and vision-language modeling.
• In biostatistics, it refers to a joint longitudinal–multi-state model linking biomarker data and transition hazards via shared random effects, while in AMG it minimizes energy in prolongation operators.
• Each MMProLong usage employs tailored mathematical frameworks, highlighting the need for precise domain context and explicit arXiv identifiers to avoid ambiguity.

Searching arXiv for papers associated with “MMProLong” to ground the article in the current literature. Search query: MMProLong MMProLong is an overloaded research label rather than a single canonical method. In the arXiv record supplied here, it denotes four unrelated technical constructions: a shared-random-effects joint model for longitudinal and multi-state clinical progression; a constrained energy-minimization prolongation procedure for algebraic multigrid (AMG); a proximal causal-inference framework for longitudinal marginal structural mean models (MSMMs); and a long-context continued pre-training recipe for a 128K-context 7B vision-LLM. The shared acronym masks substantial differences in mathematical object, inferential target, computational substrate, and application domain, ranging from biostatistics and semiparametric causal inference to sparse linear solvers and multimodal large-model training (Ferrer et al., 2015, Janna et al., 2022, Ying et al., 2021, Wang et al., 13 May 2026).

1. Nomenclature and domain-specific meanings

In the available literature, “MMProLong” functions as a local shorthand attached to domain-specific methods. It does not identify a unified family of algorithms.

Usage of “MMProLong”	Domain	Core technical object
Joint multi-state/longitudinal framework	Biostatistics, survival analysis	Linear mixed model + multi-state proportional hazards linked by shared random effects
Parallel energy-minimization prolongation	Numerical linear algebra, AMG	Prolongation operator $P$ minimizing $\operatorname{tr}(P^\top A P)$ under near-kernel constraints
Proximal inference for longitudinal MSMMs	Causal inference	Outcome-bridge and treatment-bridge identification with POR, PIPW, and PDR estimators
Long-context continued pre-training recipe	Vision-language modeling	128K-context Qwen2.5-VL-7B variant trained with long-document VQA mixtures

This polysemy has practical consequences for citation and retrieval. A reference to “MMProLong” is not self-disambiguating; the relevant arXiv identifier or surrounding domain vocabulary is necessary to determine whether the subject is a joint likelihood over biomarker trajectories and transition intensities, a matrix-free PCG scheme for AMG prolongation, a proximal bridge-function estimator under unmeasured confounding, or a long-context multimodal training recipe (Ferrer et al., 2015, Janna et al., 2022, Ying et al., 2021, Wang et al., 13 May 2026).

2. MMProLong as a joint longitudinal–multi-state model

In Ferrer et al., MMProLong denotes a joint model for a longitudinal process and a multi-state process, divided into a linear mixed sub-model for longitudinal data and a multi-state sub-model with proportional hazards for transition times, both linked by shared random effects (Ferrer et al., 2015). The longitudinal component observes marker measurements $Y_{ij}$ at times $t_{ij}$ and assumes the Gaussian linear mixed-effects specification

$$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$

with $b_i \sim N(0,D)$ and $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$, independent of $b_i$. The multi-state component treats $E_i(t)\in S=\{0,1,\dots,M\}$ as a non-homogeneous Markov process and models each transition intensity $h\to k$ as

$\operatorname{tr}(P^\top A P)$0

where $\operatorname{tr}(P^\top A P)$1 may encode the current level $\operatorname{tr}(P^\top A P)$2, the current slope $\operatorname{tr}(P^\top A P)$3, or both.

The linkage mechanism is the shared-random-effects construction. The same $\operatorname{tr}(P^\top A P)$4 enters the longitudinal trajectory directly and the transition hazards indirectly through functions of the latent biomarker trajectory. This permits transition-specific association parameters $\operatorname{tr}(P^\top A P)$5 to quantify how biomarker dynamics drive each hazard. In the prostate-cancer application, the state space was $\operatorname{tr}(P^\top A P)$6, with clinically plausible transitions including $\operatorname{tr}(P^\top A P)$7, $\operatorname{tr}(P^\top A P)$8, $\operatorname{tr}(P^\top A P)$9, $Y_{ij}$0, $Y_{ij}$1, $Y_{ij}$2, $Y_{ij}$3, $Y_{ij}$4, $Y_{ij}$5, and $Y_{ij}$6 (Ferrer et al., 2015).

The observed likelihood is

$Y_{ij}$7

with the longitudinal Gaussian density, the random-effects Gaussian density, and the continuous-time Markov contribution to $Y_{ij}$8. Because neither the $Y_{ij}$9-integral nor the time-integral in $t_{ij}$0 has an analytic form, estimation uses an EM algorithm, quasi-Newton acceleration in case of slow convergence, Gauss-Kronrod quadrature for $t_{ij}$1, and pseudo-adaptive Gauss-Hermite quadrature for $t_{ij}$2, with at least 9 nodes recommended. Asymptotic variances are taken from the inverse Hessian at convergence (Ferrer et al., 2015).

The implementation is provided in R through JMstateModel(), combining and extending mstate and JM. The workflow uses nlme::lme for the longitudinal fit, mstate::msprep and expand.covs for multi-state data preparation, survival::coxph stratified on transition for the survival object, and then JMstateModel() with options such as parameterization="both" and method="spline-PH-aGH" (Ferrer et al., 2015). Dynamic prediction is then based on the subject-specific posterior of $t_{ij}$3 given marker history up to time $t_{ij}$4, yielding

$t_{ij}$5

The simulation study used 500 replicates with $t_{ij}$6 under a 3-state model $t_{ij}$7, comparing 3, 9, and 15 Gauss-Hermite nodes; with 9 nodes, relative bias was below $t_{ij}$8 for most parameters and confidence-interval coverage was approximately $t_{ij}$9 (Ferrer et al., 2015). In the pooled application to $$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$0 prostate-cancer patients from the RTOG 9406 and BCCA cohorts, the final joint model included both current level and slope in every transition hazard. Baseline iPSA, T-stage, and Gleason significantly shaped short-term drop and long-term rise of log PSA; current PSA level and particularly slope strongly increased the hazard of every first transition $$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$1, while later transitions showed weaker or non-significant effects (Ferrer et al., 2015).

3. MMProLong as parallel energy-minimization prolongation for AMG

In AMG, MMProLong denotes a constrained minimization procedure for constructing prolongation operators with low energy while preserving near-kernel components in the interpolation space (Janna et al., 2022). For an SPD matrix $$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$2 and prolongation $$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$3, the energy functional is

$$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$4

The near-kernel requirement introduces a basis $$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$5 and enforces

$$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$6

or, under $$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$7–$$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$8 splitting with $$Y_{ij} = Y_i^*(t_{ij}) + \varepsilon_{ij} = X_i^L(t_{ij})^\prime \beta + Z_i(t_{ij})^\prime b_i + \varepsilon_{ij},$$9 and $b_i \sim N(0,D)$0,

$b_i \sim N(0,D)$1

The resulting optimization problem seeks

$b_i \sim N(0,D)$2

Using Lagrange multipliers yields the saddle-point system

$b_i \sim N(0,D)$3

where $b_i \sim N(0,D)$4 is the vector of nonzeros of $b_i \sim N(0,D)$5, $b_i \sim N(0,D)$6 is block-diagonal with blocks extracted from the sparsity pattern of each column, $b_i \sim N(0,D)$7 collects $b_i \sim N(0,D)$8, and $b_i \sim N(0,D)$9 encodes $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$0 (Janna et al., 2022).

The paper then uses null-space reduction. Writing $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$1, enforcing $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$2, and defining

$\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$3

one obtains the singular projected system

$\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$4

which is solved by a preconditioned restricted conjugate-gradient algorithm, denoted EMIN_PCG in Algorithm 3.1. The method maintains $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$5 and monitors the energy decrease

$\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$6

halting when $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$7 (Janna et al., 2022).

The implementation is explicitly parallel. Rather than storing $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$8, the method performs matrix-free $\varepsilon_i \sim N(0,\sigma^2 I_{n_i})$9 application through sparse matrix-matrix multiplication $b_i$0 restricted to the fixed sparsity pattern of $b_i$1. Chronos partitions the graph of $b_i$2 with ParMETIS; each MPI rank owns a row-block of $b_i$3 and the corresponding rows of $b_i$4, $b_i$5, and $b_i$6. Because $b_i$7 is block-diagonal by row, local QR or SVD factorizations suffice to form $b_i$8 without communication. The preconditioner may be Jacobi or block symmetric Gauss-Seidel, again performed matrix-free (Janna et al., 2022).

The theoretical motivation connects low-energy prolongation to classical AMG ideals. The paper notes that the ideal prolongation satisfies $b_i$9 on fine rows, corresponding to zero-energy prolongation, and that optimal prolongation spans the first $E_i(t)\in S=\{0,1,\dots,M\}$0 eigenvectors of the generalized eigenproblem $E_i(t)\in S=\{0,1,\dots,M\}$1, minimizing the two-grid convergence bound (Janna et al., 2022). Energy minimization with exact interpolation constraints therefore balances approximation of low-energy modes and operator complexity.

The reported experiments cover medium-size tests, weak scaling, and seven large real-world problems up to 134M degrees of freedom. On medium-size problems, Jacobi required approximately 4–8 PCG iterations for a $E_i(t)\in S=\{0,1,\dots,M\}$2 energy drop, while Gauss-Seidel required approximately 2–4; an operating point near 2 Jacobi PCG iterations increased setup cost by at most about $E_i(t)\in S=\{0,1,\dots,M\}$3, reduced solver iterations by at least about $E_i(t)\in S=\{0,1,\dots,M\}$4, and lowered total time by about $E_i(t)\in S=\{0,1,\dots,M\}$5–$E_i(t)\in S=\{0,1,\dots,M\}$6 (Janna et al., 2022). In weak scaling on tetrahedral elasticity from 222k to 124M dofs over 1 to 512 nodes, the method maintained bounded complexities and favorable time-to-solution. Against PETSc/GAMG and classical smooth-prolongation AMG on large application problems, it reduced operator complexity by $E_i(t)\in S=\{0,1,\dots,M\}$7–$E_i(t)\in S=\{0,1,\dots,M\}$8, iteration count by up to $E_i(t)\in S=\{0,1,\dots,M\}$9, and total time by $h\to k$0–$h\to k$1; on the 73M and 134M cases, GAMG either failed to converge or was $h\to k$2–$h\to k$3 slower (Janna et al., 2022).

4. MMProLong as proximal causal inference for complex longitudinal studies

In semiparametric causal inference, MMProLong denotes a framework for longitudinal marginal structural mean models when sequential randomization fails because measured covariates are only imperfect proxies of latent confounding (Ying et al., 2021). The observed data comprise treatments $h\to k$4, baseline and time-varying covariates partitioned into common-cause proxies $h\to k$5, treatment-inducing proxies $h\to k$6, and outcome-inducing proxies $h\to k$7, and a final outcome $h\to k$8. Under consistency, $h\to k$9. The target model is

$\operatorname{tr}(P^\top A P)$00

with additive linear and saturated MSMMs given as explicit examples (Ying et al., 2021).

The framework is positioned against the sequential randomization assumption

$\operatorname{tr}(P^\top A P)$01

which requires complete measurement of time-varying confounders. Instead, it introduces proximal independence assumptions for the proxy variables. For the two-period case, outcome-inducing proxies $\operatorname{tr}(P^\top A P)$02 satisfy

$\operatorname{tr}(P^\top A P)$03

while treatment-inducing proxies $\operatorname{tr}(P^\top A P)$04 satisfy

$\operatorname{tr}(P^\top A P)$05

together with the requirement that $\operatorname{tr}(P^\top A P)$06 have no direct effect on the outcome (Ying et al., 2021). Identification further requires completeness conditions ensuring that the proxies are sufficiently informative about the unmeasured confounders $\operatorname{tr}(P^\top A P)$07.

Under these conditions, the paper establishes the existence of outcome-bridge functions

$\operatorname{tr}(P^\top A P)$08

and treatment-bridge functions

$\operatorname{tr}(P^\top A P)$09

solving sequences of Fredholm integral equations of the first kind. The outcome-bridge route leads to a proximal $\operatorname{tr}(P^\top A P)$10-formula: if the recursive equations for $\operatorname{tr}(P^\top A P)$11 hold, then

$\operatorname{tr}(P^\top A P)$12

The treatment-bridge route yields a proximal IPW representation, with the final expression

$\operatorname{tr}(P^\top A P)$13

The recursive structure “peels off” one time step at a time (Ying et al., 2021).

Estimation uses low-dimensional working models for $\operatorname{tr}(P^\top A P)$14 and $\operatorname{tr}(P^\top A P)$15, fitted through moment equations. On that basis, the paper defines three estimator classes for $\operatorname{tr}(P^\top A P)$16: proximal outcome-regression (POR), proximal IPW (PIPW), and proximal doubly robust (PDR). The PDR estimating function combines bridge estimates and an MSMM restriction, and Theorem 5 states that the PDR estimator is consistent if and only if either the $\operatorname{tr}(P^\top A P)$17-models or the $\operatorname{tr}(P^\top A P)$18-models are correct (Ying et al., 2021). The paper also characterizes regular and asymptotically linear estimators and gives the semiparametric efficiency bound, with the efficient influence function stated in the supplement.

The simulation study, reported for $\operatorname{tr}(P^\top A P)$19 and $\operatorname{tr}(P^\top A P)$20, found that PDR bias was near zero when either nuisance model was correct; POR and PIPW were biased under misspecification of their respective nuisance models; empirical standard error was approximately equal to the theoretical efficiency bound when both bridge models were correct; and $\operatorname{tr}(P^\top A P)$21 confidence-interval coverage for PDR was approximately $\operatorname{tr}(P^\top A P)$22 under double robustness (Ying et al., 2021). The real-data application involved $\operatorname{tr}(P^\top A P)$23 rheumatoid arthritis patients, treatments $\operatorname{tr}(P^\top A P)$24 representing methotrexate use at baseline and 6 months, outcome $\operatorname{tr}(P^\top A P)$25 equal to tender joints at 12 months, and proxies divided into $\operatorname{tr}(P^\top A P)$26, $\operatorname{tr}(P^\top A P)$27, and $\operatorname{tr}(P^\top A P)$28. POR, PIPW, and PDR all agreed on a significant protective effect of baseline MTX, with $\operatorname{tr}(P^\top A P)$29 tender joints in the saturated MSMM, while the standard SRA-based DR estimator was weaker and non-significant; in the cumulative MSMM, PDR gave $\operatorname{tr}(P^\top A P)$30 with $\operatorname{tr}(P^\top A P)$31 CI $\operatorname{tr}(P^\top A P)$32, compared with SRA-DR $\operatorname{tr}(P^\top A P)$33 and CI $\operatorname{tr}(P^\top A P)$34 (Ying et al., 2021).

5. MMProLong as a 128K long-context vision-LLM

In long-context vision-language modeling, MMProLong is a 128K-context-enabled 7B multimodal model obtained by long-context continued pre-training of Qwen2.5-VL-7B-Instruct (Wang et al., 13 May 2026). The initialization keeps the transformer-vision-language backbone unchanged except for positional modifications: the original 32K context window is extended to $\operatorname{tr}(P^\top A P)$35, the model retains mRoPE rotary embeddings with base frequency $\operatorname{tr}(P^\top A P)$36, and Dynamic-NTK scaling sets

$\operatorname{tr}(P^\top A P)$37

FlashAttention and 8-GPU sequence/FSDP parallelism are used to enable efficient attention over 128K tokens (Wang et al., 13 May 2026).

The continued pre-training budget is 5B tokens on multimodal long-document tasks. The primary data mixture consists of long-document VQA with three task types: extract-single, extract-multi, and reasoning. OCR transcription tasks, OCR-full and OCR-needle, are included only for comparison. Under the controlled 5B-token budget, Table 1 reports that the three VQA tasks each improve downstream long-document VQA by $\operatorname{tr}(P^\top A P)$38–$\operatorname{tr}(P^\top A P)$39 points, whereas OCR-based objectives yield poor performance even after extra SFT; the final recipe therefore uses VQA only (Wang et al., 13 May 2026).

A central ablation concerns sequence-length distribution. Document spans are sampled from a PDF pool of 1.5M documents with 32–50 pages, producing a “pool-native” distribution $\operatorname{tr}(P^\top A P)$40 over $\operatorname{tr}(P^\top A P)$41, while an alternative “long-biased” distribution $\operatorname{tr}(P^\top A P)$42 has $\operatorname{tr}(P^\top A P)$43 of examples above 100K tokens, compared with $\operatorname{tr}(P^\top A P)$44 for pool-native. Table 8 shows that pool-native outperforms long-biased consistently: $\operatorname{tr}(P^\top A P)$45 at 64K for extract-single, $\operatorname{tr}(P^\top A P)$46 for extract-multi, $\operatorname{tr}(P^\top A P)$47 for reasoning, and $\operatorname{tr}(P^\top A P)$48 overall (Wang et al., 13 May 2026). Another ablation varies task-mixture weights. Grouping extract-single and extract-multi as information extraction, the best mixture is retrieval-heavy with

$\operatorname{tr}(P^\top A P)$49

that is, an $\operatorname{tr}(P^\top A P)$50 extraction:reasoning ratio. The generic loss is standard autoregressive cross-entropy,

$\operatorname{tr}(P^\top A P)$51

The optimization recipe in Table A.1 is explicit: AdamW, learning rate $\operatorname{tr}(P^\top A P)$52 with cosine decay, $\operatorname{tr}(P^\top A P)$53 warmup, global batch size of 4M tokens across 32 sequences, and stopping when $\operatorname{tr}(P^\top A P)$54, corresponding to approximately 1,250 updates (Wang et al., 13 May 2026). Short-context mixing is not required: Table 4 and Figure 1 show that pure long-context training with $\operatorname{tr}(P^\top A P)$55 short data attains the best long-document VQA, scoring 59.56 at 64K and 55.84 at 128K, while short-context average declines only slightly from 66.47 to 65.48 (Wang et al., 13 May 2026).

Performance gains are reported against Qwen2.5-VL-7B. On long-document VQA, the average rises from 52.24 to 59.56 at 64K and from 48.94 to 55.84 at 128K, for an average gain of $\operatorname{tr}(P^\top A P)$56 points (Wang et al., 13 May 2026). The model also extrapolates beyond its training window without additional training: at 256K it reaches 55.09 versus a baseline 38.12, and at 512K it reaches 52.52 versus 19.49. Transfer to downstream long-context tasks is likewise reported. On webpage needle retrieval (MM-NIAH) at 128K, the average is 42.28 versus 12.17 for Qwen2.5-VL-7B, and the retrieval subtask rises from 11.33 to 57.83. On VTCBench, overall performance improves from 48.23 to 52.73, with reasoning 22.88 versus 15.63 and memory 40.50 versus 33.83. On long-video understanding, Video-MME, MLVU, and LongVideoBench rise to 67.78, 73.55, and 62.08, compared with 65.1, 70.2, and 60.43 respectively (Wang et al., 13 May 2026).

6. Comparative interpretation and disambiguation

The four usages of MMProLong are methodologically unrelated, and treating them as variants of a single framework would be incorrect. In the joint-modeling literature, MMProLong is a likelihood-based framework with latent random effects mediating dependence between biomarker evolution and multi-state event intensities (Ferrer et al., 2015). In AMG, it is an optimization problem over sparse prolongation operators, solved by projected PCG and motivated by low-energy interpolation of near-kernel modes (Janna et al., 2022). In proximal causal inference, it is a semiparametric identification and estimation strategy based on proxy variables, completeness, and bridge functions under failure of sequential randomization (Ying et al., 2021). In multimodal foundation models, it is a continued pre-training recipe and trained checkpoint for long-context VLM behavior at 128K and beyond (Wang et al., 13 May 2026).

A useful disambiguation principle is therefore to identify the mathematical primitive attached to the term. If the surrounding notation involves $\operatorname{tr}(P^\top A P)$57, $\operatorname{tr}(P^\top A P)$58, $\operatorname{tr}(P^\top A P)$59, and transition intensities $\operatorname{tr}(P^\top A P)$60, the reference is to joint longitudinal–multi-state clinical modeling (Ferrer et al., 2015). If the central objects are $\operatorname{tr}(P^\top A P)$61, $\operatorname{tr}(P^\top A P)$62, $\operatorname{tr}(P^\top A P)$63, $\operatorname{tr}(P^\top A P)$64, and $\operatorname{tr}(P^\top A P)$65, it is the AMG prolongation method (Janna et al., 2022). If the notation centers on $\operatorname{tr}(P^\top A P)$66, $\operatorname{tr}(P^\top A P)$67, $\operatorname{tr}(P^\top A P)$68, $\operatorname{tr}(P^\top A P)$69, $\operatorname{tr}(P^\top A P)$70, $\operatorname{tr}(P^\top A P)$71, and $\operatorname{tr}(P^\top A P)$72, it is the proximal MSMM framework (Ying et al., 2021). If the discussion concerns Qwen2.5-VL-7B, mRoPE, 128K context, long-document VQA, and a 5B-token budget, it is the long-context LVLM model (Wang et al., 13 May 2026).

This suggests that “MMProLong” is best understood as a collision of local acronyms rather than as a stable transdisciplinary concept. For technical communication, the arXiv identifier is therefore not merely bibliographic metadata but an essential part of the concept’s specification.