--- title: 'MMProLong: Disambiguating Multi-Domain Methods' url: https://www.emergentmind.com/topics/mmprolong type: topic --- # MMProLong: Disambiguating Multi-Domain Methods 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-language model. 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 [1506.07496] [2208.02995] [2109.07030] [2605.13831]. ## 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 [1506.07496] [2208.02995] [2109.07030] [2605.13831]. ## 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 [1506.07496]. 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 \[ \lambda_{hk}^i(t\mid b_i) = \lambda_{hk,0}(t)\exp\{X_{hk,i}^{S\prime}\gamma_{hk}+W_{hk,i}(b_i,t)^\prime\eta_{hk}\}, \] where \(W_{hk,i}(b_i,t)\) may encode the current level \(Y_i^*(t)\), the current slope \(\partial Y_i^*(t)/\partial t\), or both. The linkage mechanism is the shared-random-effects construction. The same \(b_i\) enters the longitudinal trajectory directly and the transition hazards indirectly through functions of the latent biomarker trajectory. This permits transition-specific association parameters \(\eta_{hk}\) to quantify how biomarker dynamics drive each hazard. In the prostate-cancer application, the state space was \(\{0=\text{post-EBRT},1=\text{local recurrence},2=\text{salvage HT},3=\text{distant recurrence},4=\text{death}\}\), with clinically plausible transitions including \(0\to1\), \(0\to2\), \(0\to3\), \(0\to4\), \(1\to2\), \(1\to3\), \(1\to4\), \(2\to3\), \(2\to4\), and \(3\to4\) [1506.07496]. The observed likelihood is \[ L_i(\theta)=\int f_Y(Y_i\mid b_i;\theta)\,f_E(E_i\mid b_i;\theta)\,f_b(b_i;\theta)\,db_i, \] with the longitudinal Gaussian density, the random-effects Gaussian density, and the continuous-time Markov contribution to \(f_E\). Because neither the \(b_i\)-integral nor the time-integral in \(f_E\) has an analytic form, estimation uses an EM algorithm, quasi-Newton acceleration in case of slow convergence, Gauss-Kronrod quadrature for \(\int_0^t \lambda_{hh}(u\mid b_i)\,du\), and pseudo-adaptive Gauss-Hermite quadrature for \(\int db_i\), with at least 9 nodes recommended. Asymptotic variances are taken from the inverse Hessian at convergence [1506.07496]. 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"` [1506.07496]. Dynamic prediction is then based on the subject-specific posterior of \(b_i\) given marker history up to time \(s\), yielding \[ F_{hk}^i(s,t)=P(E_i(t)=k\mid E_i(s)=h,Y_i^{(s)},X_i^S) = \int P_{hk}^i(s,t\mid b)\,p(b\mid Y_i^{(s)},\hat\theta)\,db. \] The simulation study used 500 replicates with \(N=1500\) under a 3-state model \((0\to1,0\to2,1\to2)\), comparing 3, 9, and 15 Gauss-Hermite nodes; with 9 nodes, relative bias was below \(2\%\) for most parameters and confidence-interval coverage was approximately \(95\%\) [1506.07496]. In the pooled application to \(N=1474\) 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 \((0\to1,0\to2,0\to3)\), while later transitions showed weaker or non-significant effects [1506.07496]. ## 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 [2208.02995]. For an SPD matrix \(A\in\mathbb{R}^{n\times n}\) and prolongation \(P\in\mathbb{R}^{n\times n_c}\), the energy functional is \[ J(P)=\operatorname{tr}(P^\top A P)=\sum_{j=1}^{n_c}(Pe_j)^\top A(Pe_j)=\|P\|_A^2. \] The near-kernel requirement introduces a basis \(V\in\mathbb{R}^{n\times m}\) and enforces \[ \operatorname{range}(P)\supseteq \operatorname{span}(V), \qquad\text{equivalently}\qquad V=PX, \] or, under \(C\)–\(F\) splitting with \(P=[W;I]\) and \(V=[V_f;V_c]\), \[ W V_c = V_f. \] The resulting optimization problem seeks \[ P^*=\arg\min_{P\in\text{SparsityPattern}} \operatorname{tr}(P^\top A P) \quad\text{subject to}\quad V\subseteq \operatorname{range}(P). \] Using Lagrange multipliers yields the saddle-point system \[ \begin{bmatrix} K & B\\ B^\top & 0 \end{bmatrix} \begin{bmatrix} p\\ \lambda \end{bmatrix} = \begin{bmatrix} f\\ g \end{bmatrix}, \] where \(p\) is the vector of nonzeros of \(P\), \(K\) is block-diagonal with blocks extracted from the sparsity pattern of each column, \(f\) collects \(-A(\mathcal{I}_i,i)\), and \(B^\top p=g\) encodes \(W V_c=V_f\) [2208.02995]. The paper then uses null-space reduction. Writing \(P=p_0+\delta p\), enforcing \(B^\top \delta p=0\), and defining \[ \Pi_B = I - B(B^\top B)^{-1}B^\top, \] one obtains the singular projected system \[ \Pi_B K \Pi_B \,\delta w = -\Pi_B(f+K w_0), \] which is solved by a preconditioned restricted conjugate-gradient algorithm, denoted `EMIN_PCG` in Algorithm 3.1. The method maintains \(\delta w\in\ker(B^\top)\) and monitors the energy decrease \[ \Delta E_{k+1} = -\frac{(p_k^\top r_k)^2}{p_k^\top K p_k}, \] halting when \(\Delta E_k/\Delta E_1 \le \tau\) [2208.02995]. The implementation is explicitly parallel. Rather than storing \(K\), the method performs matrix-free \(Kp\) application through sparse matrix-matrix multiplication \(A P\) restricted to the fixed sparsity pattern of \(P\). Chronos partitions the graph of \(A\) with ParMETIS; each MPI rank owns a row-block of \(A\) and the corresponding rows of \(P\), \(W\), and \(B\). Because \(B\) is block-diagonal by row, local QR or SVD factorizations suffice to form \(\Pi_B=I-QQ^\top\) without communication. The preconditioner may be Jacobi or block symmetric Gauss-Seidel, again performed matrix-free [2208.02995]. The theoretical motivation connects low-energy prolongation to classical AMG ideals. The paper notes that the ideal prolongation satisfies \(A P_{\mathrm{id}}=0\) on fine rows, corresponding to zero-energy prolongation, and that optimal prolongation spans the first \(n_c\) eigenvectors of the generalized eigenproblem \(A x=\lambda \widetilde M x\), minimizing the two-grid convergence bound [2208.02995]. 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 \(10^{-2}\) energy drop, while Gauss-Seidel required approximately 2–4; an operating point near 2 Jacobi PCG iterations increased setup cost by at most about \(20\%\), reduced solver iterations by at least about \(40\%\), and lowered total time by about \(12\%\)–\(25\%\) [2208.02995]. 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 \(10\%\)–\(30\%\), iteration count by up to \(60\%\), and total time by \(5\%\)–\(55\%\); on the 73M and 134M cases, GAMG either failed to converge or was \(2\)–\(3\times\) slower [2208.02995]. ## 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 [2109.07030]. The observed data comprise treatments \(A_1,\dots,A_T\), baseline and time-varying covariates partitioned into common-cause proxies \(X_t\), treatment-inducing proxies \(Z_t\), and outcome-inducing proxies \(W_t\), and a final outcome \(Y\). Under consistency, \(Y=Y_{\overline A(T)}\). The target model is \[ E[Y_{\overline a}\mid V] = m(\overline a,V;\beta), \] with additive linear and saturated MSMMs given as explicit examples [2109.07030]. The framework is positioned against the sequential randomization assumption \[ Y_{\overline a}\perp A_t\mid \overline A(t-1),\overline L(t), \] 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 \(W_t\) satisfy \[ W_t \perp (A_t,Z_t)\mid U_t,\overline X(t),\overline A(t-1), \] while treatment-inducing proxies \(Z_t\) satisfy \[ (A_t,Z_t)\perp Y\mid U_t,\overline X(t),\overline A(t-1), \] together with the requirement that \(Z_t\) have no direct effect on the outcome [2109.07030]. Identification further requires completeness conditions ensuring that the proxies are sufficiently informative about the unmeasured confounders \(U_t\). Under these conditions, the paper establishes the existence of outcome-bridge functions \[ h_t(\overline w(t),\overline a(t),\overline x(t)), \] and treatment-bridge functions \[ q_t(\overline z(t),\overline a(t),\overline x(t)), \] solving sequences of Fredholm integral equations of the first kind. The outcome-bridge route leads to a proximal \(g\)-formula: if the recursive equations for \(h_T,\dots,h_0\) hold, then \[ E[Y_{\overline a(T)}\mid V] = E[h_0(W_0,\overline a(T),X_0)\mid V]. \] The treatment-bridge route yields a proximal IPW representation, with the final expression \[ E[Y_{\overline a(1)}\mid V] = E\!\left[Y\mathbf{1}\{\overline A=\overline a\}\,q_T(\overline Z,\overline a,\overline X)\mid V\right]. \] The recursive structure “peels off” one time step at a time [2109.07030]. Estimation uses low-dimensional working models for \(h_t(\cdot;b_t)\) and \(q_t(\cdot;t_t)\), fitted through moment equations. On that basis, the paper defines three estimator classes for \(\beta\): 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 \(h\)-models or the \(q\)-models are correct [2109.07030]. 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 \(n=500\) and \(n=1000\), 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 \(95\%\) confidence-interval coverage for PDR was approximately \(0.95\) under double robustness [2109.07030]. The real-data application involved \(n=1010\) rheumatoid arthritis patients, treatments \(A_0,A_1\in\{0,1\}\) representing methotrexate use at baseline and 6 months, outcome \(Y\) equal to tender joints at 12 months, and proxies divided into \(Z_t\), \(W_t\), and \(X_t\). POR, PIPW, and PDR all agreed on a significant protective effect of baseline MTX, with \(\beta_2\approx -0.5\) tender joints in the saturated MSMM, while the standard SRA-based DR estimator was weaker and non-significant; in the cumulative MSMM, PDR gave \(\hat\beta_1\approx -0.29\) with \(95\%\) CI \((-0.42,-0.16)\), compared with SRA-DR \(-0.17\) and CI \((-0.41,0.06)\) [2109.07030]. ## 5. MMProLong as a 128K long-context vision-language model 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 [2605.13831]. The initialization keeps the transformer-vision-language backbone unchanged except for positional modifications: the original 32K context window is extended to \(L_{\max}=131{,}072\), the model retains mRoPE rotary embeddings with base frequency \(\omega_0=1\times 10^6\), and Dynamic-NTK scaling sets \[ t = 128K/32K = 4,\qquad \omega=\omega_0\cdot t = 4\times 10^6. \] FlashAttention and 8-GPU sequence/FSDP parallelism are used to enable efficient attention over 128K tokens [2605.13831]. 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 \(+5\)–\(6\) points, whereas OCR-based objectives yield poor performance even after extra SFT; the final recipe therefore uses VQA only [2605.13831]. 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 \(p_p(L)\) over \(L\in[32K,128K]\), while an alternative “long-biased” distribution \(p_\ell(L)\) has \(83.9\%\) of examples above 100K tokens, compared with \(23.6\%\) for pool-native. Table 8 shows that pool-native outperforms long-biased consistently: \(+1.3\) at 64K for `extract-single`, \(+0.1\) for `extract-multi`, \(+1.7\) for `reasoning`, and \(+1.4\) overall [2605.13831]. Another ablation varies task-mixture weights. Grouping `extract-single` and `extract-multi` as information extraction, the best mixture is retrieval-heavy with \[ \alpha_{es}+\alpha_{em}+\alpha_r=1,\qquad \alpha_{es}=\alpha_{em}=0.4,\qquad \alpha_r=0.2, \] that is, an \(8{:}2\) extraction:reasoning ratio. The generic loss is standard autoregressive cross-entropy, \[ L = -\sum_{t=1}^{|y_i|}\log P(y_{i,t}\mid y_{i,