MUNI: Cross-Domain Applications

• MUNI is a cross-disciplinary term denoting municipal service systems, municipal bonds, market microstructure models, urban transit agencies, and multimodal latent diffusion frameworks.
• In municipal operations, MUNI employs multi-view spatiotemporal encoders and intra/inter-type interactions to significantly improve 311 service-time predictions with reduced error metrics.
• MUNI also encompasses financial valuation using supervised similarity, ratio-modeling in high-frequency trading, urban mobility optimization through game-theoretic frameworks, and innovative machine learning approaches in diffusion modeling.

“MUNI” is not a single technical object but a domain-dependent label used in several research literatures. In the supplied corpus it refers, among other usages, to municipal non-emergency service systems and municipal transit agencies, the municipal bond market, Muni Toke’s ratio-modeling line in market microstructure, and the multimodal generative framework “MUNI: Multimodal Unified Latent Diffusion for Coherent Any-to-Any Generation” (Asif et al., 22 Aug 2025, Saha et al., 2024, Toke et al., 2020, Yeo et al., 15 Jun 2026, Guo et al., 2024, Zambrano et al., 6 Feb 2025, Chomei, 2023). The term therefore functions as a cross-domain homonym whose meaning must be fixed by context, methodology, and notation.

1. Disambiguation across research domains

The principal usages represented in recent arXiv literature can be organized as follows.

Usage of “MUNI”	Research referent
Municipal operations	City 311 service-time estimation
Finance	Municipal bond relative valuation
Market microstructure	Muni Toke ratio models for LOB events
Urban mobility	Municipal transit design and regulation
Machine learning	Multimodal Unified Latent Diffusion

In municipal operations, the relevant object is the city 311 system, where residents submit non-emergency requests such as missed garbage collection, potholes, and noise complaints, and the modeling problem is service-time estimation under spatial, temporal, and type heterogeneity (Asif et al., 22 Aug 2025). In finance, “muni” denotes municipal bonds, with the central problem being relative valuation in a market where only around 2% of over a million outstanding securities trade daily (Saha et al., 2024). In high-frequency finance, “MUNI” refers to Muni Toke’s Cox-type and marked ratio models for limit order books, where intensity ratios remove a common baseline intensity from the estimation problem (Toke et al., 2020, Chomei, 2023). In transportation, “MUNI” denotes a municipal transit agency context, used to motivate optimization and game-theoretic frameworks for multimodal mobility systems with transit, AMoD, ride-hailing, micromobility, and walking (Guo et al., 2024, Zambrano et al., 6 Feb 2025). In multimodal generative modeling, “MUNI” is the name of an end-to-end latent diffusion architecture for any-to-any generation (Yeo et al., 15 Jun 2026).

2. Municipal service systems: spatial–temporal–type learning for 311 requests

Within municipal service operations, the central task is to estimate the service time of non-emergency requests. The formulation in “MuST2-Learn” partitions a city into $M$ regions with index set $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$ and defines $N$ request types with index set $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$. At day $t$, the aggregated intra-type features for region $i$ and type $l$ are request volume $r^l_{i,t}\in\mathbb{R}$ and average service time $d^l_{i,t}\in\mathbb{R}$, assembled into the temporal sequence

$$e_i^l=[e_{i,t}^l]_{t=t'-T}^{t'-1}\in\mathbb{R}^{2T},\qquad e_{i,t}^l=[r^l_{i,t},d^l_{i,t}].$$

For a newly submitted request $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$0, the target is realized service time $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$1 and prediction $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$2, with training by mean squared error and evaluation by MAE, RMSE, and MAPE (Asif et al., 22 Aug 2025).

The architecture is explicitly multi-view. An intra-type spatiotemporal encoder first applies a Transformer to each regional sequence,

$\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$3

then stacks hidden states across regions and applies a 1D-CNN,

$\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$4

to capture cross-region influence under irregular district layouts. An inter-type interaction encoder then attends across the set of type embeddings $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$5 to model resource-coupled dependence across heterogeneous request types. This design is motivated by empirical Pearson correlations greater than $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$6 among several Solid Waste request types. An intra-type variation encoder addresses long tails and within-type dispersion using Gaussian Process Regression with an RBF kernel, producing posterior mean $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$7 and variance $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$8, and augments these with an LLM-derived workload score $\mathcal{S}\equiv\mathcal{M}=\{1,\dots,M\}$9 extracted from request text. The request-level embedding is

$N$0

Fusion is performed by concatenating $N$1 and passing it through an MLP with ReLU activations.

The reported Chattanooga study uses nearly 170,000 311 requests from 2022–2024, with response time capped at 80 days for analysis; Newark contributes approximately 7,500 requests for generalization experiments. Hyperparameters include learning rate $N$2, an 80/20 train/test split, look-back window $N$3 days, Transformer with 32 hidden units and 4 heads, inter-type attention with 64 hidden units and 4 heads, and an MLP with 64 hidden units. Inference latency is reported as less than 1 ms per request on the stated hardware. Representative Chattanooga results include MAE $N$4 for New Garbage Container, $N$5 for Bulk Trash, $N$6 for Missed Recycle, $N$7 for Missed Garbage, $N$8 for Brush Collection, and $N$9 for Bagged Yard Waste. Overall, MuST2-Learn reduces mean absolute error by at least 32.5% across request types versus all baselines, lowers MAPE by up to 84.6%, and lowers MAE by up to 97.1% relative to DeepSTA. On Newark, MAPE is reported as 8–12% across six types, with Prophet outperforming the method on Animal Complaint. Ablation results show that the intra-type variation encoder is the most impactful component; for one type, removing it increased MAPE from 24.90% to 94.72%. The paper also identifies a mid-2024 Chattanooga brush pickup policy change that degraded RMSE/MSE for that type, emphasizing drift sensitivity and the need for model refresh or change-point detection in production (Asif et al., 22 Aug 2025).

3. Municipal bonds: supervised similarity for relative valuation

In fixed-income research, “muni” denotes the municipal bond market, characterized by extreme heterogeneity and sparse trading. The working universe in the cited study is approximately 225K securities after filters on credit quality, small deal sizes, and coupons. Structural heterogeneity arises from tax status, sector-specific covenants, heterogeneous credit ratings, insurance, varied call structures, and deal sizes; sparse trading is reflected in less than 10% of securities trading more than 25 days per year and 50% trading no more than 5 days per year (Saha et al., 2024).

The proposed solution is a supervised similarity framework built on a multi-output CatBoost model trained to predict yield and option-adjusted spread. CatBoost is used because it handles categorical variables via ordered boosting and target encoding without leakage, forms symmetric trees, and captures nonlinear interactions. The learned proximity is tree-importance-weighted. For Random Forests, proximity is the fraction of trees in which two instances land in the same leaf, whereas for boosted trees the paper defines

$\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$0

with tree importance

$\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$1

where $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$2 is the MultiRMSE after adding tree $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$3. Relative value for bond $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$4 is then measured by

$\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$5

The feature set comprises 22 variables, including 11 categorical and 11 numerical attributes such as State, Rating, Tax Status, Sector Code, Put-Call, Funding, Use of proceeds, Payment Frequency, Days-to-maturity, Age, Coupon, Bonds by obligor, Amount Issued, Time-to-call, and Deal Amount. Yield and OAS are winsorized, and six months of MSRB trade data preceding November 1 are used to weight training samples linearly by recency. Generic groups are first defined by issuing state and maturity bands, and cohorts are then selected either by duration-times-spread or by CatBoost similarity.

Regression benchmarks on the November 1, 2023 test fold report, for OAS, CatBoost $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$6, MAE $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$7, MSE $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$8, and MAPE $\mathcal{K}\equiv\mathcal{L}=\{1,\dots,N\}$9; for yield, CatBoost reports $t$0, MAE $t$1, MSE $t$2, and MAPE $t$3. Across six-month backtests spanning 2019–2024, the similarity-based ranking generally outperforms yield-only and rule-based DxS approaches, with higher combined backtest metrics and lower variability across market regimes. Average SHAP analysis indicates that ratings, days-to-maturity, and obligor-related features are the main OAS drivers, while days-to-maturity, put-call optionality, and obligor-related features dominate yield prediction. The method is therefore a supervised cohort construction procedure rather than a purely heuristic nearest-neighbor rule (Saha et al., 2024).

4. Muni Toke’s ratio-modeling framework in market microstructure

In high-frequency finance, “MUNI” refers to Muni Toke’s ratio-modeling program for limit order book events. The marked extension defines process indices $t$4 and mark indices $t$5, with intensities

$t$6

where $t$7 is a common baseline intensity, $t$8 is a process-level state-dependent factor, and $t$9 is a normalized mark-level conditional distribution. In the exponential/logit specification,

$i$0

and $i$1 is a multinomial logit over marks. Baseline invariance follows because the unknown $i$2 cancels in ratios, yielding the process softmax

$i$3

and the conditional mark softmax

$i$4

Estimation proceeds via quasi-log-likelihoods on the ratios, with consistency and asymptotic normality under stationarity, strong mixing, and identifiability conditions. For the pooled estimator,

$i$5

The marked model was applied to 36 Euronext Paris stocks in 2015, using imbalance, last trade sign, signed spread, and Hawkes covariates, and the best marked ratio specification achieved out-of-sample average accuracies of 0.877 for side prediction, 0.774 for aggressiveness, and 0.667 globally, outperforming Hawkes and non-marked ratio benchmarks (Toke et al., 2020).

The empirical extension to 222 Tokyo Stock Exchange stocks retains the Cox-type intensity ratio structure for market orders and studies richer depth-wise and lagged imbalance covariates. With ask-side and bid-side market orders as the two marks, the relative intensities reduce to logistic forms for $i$6 and $i$7, estimated by maximizing the partial log-likelihood

$i$8

The study reports that best-level imbalance, near-best imbalance, last trade sign, the spread-weighted last sign, and one-lag imbalance are the most predictive covariates; predictive accuracy reaches approximately 77–78% for the next market-order side, and calibrating every 1–2 weeks improves performance relative to daily calibration, while much longer windows degrade it because of parameter drift. QAIC, QCAIC, and QBIC favor parsimonious specifications such as “imb 1_la 1” and “imb 1_e_es_la 1,” which also attain high predictive accuracy (Chomei, 2023).

5. Municipal transit agencies: optimization and game-theoretic regulation

In urban mobility research, “MUNI” denotes a municipal transit agency setting in which public transit is coordinated with AMoD, ride-hailing, micromobility, and walking. One line of work formulates Transit-Centric Multimodal Urban Mobility with AMoD as a bilevel design–choice problem. The operator chooses PT frequencies $i$9, AMoD fleet allocations $l$0, and a pricing parameter $l$1 to minimize passenger disutility, while passengers choose routes according to generalized costs and discrete choice. The objective aggregates expected waiting, excess waiting, and walking. For transit legs, expected wait is $l$2; for AMoD legs, expected wait is

$l$3

The paper linearizes the route-choice map by first-order approximation and solves a sequence of large LPs. In the Chicago case study, the network contains 40 bus routes, the Red Line inbound rail, 48 five-minute intervals from 06:00–10:00, 12,400 commuters, and 2,276 distinct OD pairs. Under capital cost equivalence, replacing 20% of buses with 162 AMoD vehicles reduces average disutility from approximately 8.88 minutes to approximately 7.48 minutes, whereas under passenger car equivalence replacing 20% of buses with 82 AMoD worsens average disutility to approximately 13.02 minutes. The reported policy conclusion is transit-centric: AMoD is effective for dispersed local access and first/last mile, but trunk bus and rail capacity must be preserved (Guo et al., 2024).

A second line frames the same municipal-agency context as a hierarchical game among the municipality, service providers, and travelers. The municipality selects taxes $l$4, subsidies $l$5, public transit fares $l$6, and infrastructure decisions $l$7; providers choose prices and fleet allocations; travelers choose mode shares under generalized costs. The lower-level traveler equilibrium is computed by a convex program of the form

$l$8

subject to feasibility and capacity constraints, while provider equilibrium is obtained through best responses or KKT conditions, and the upper-level municipal problem is an MPEC. Optional modules add BPR congestion,

$l$9

and M/M/s-type waiting-time approximations for fleet-based services. The framework includes a graphical user interface for scenario analysis and has demonstrations in Lugano, Boston/Cambridge, and Kyiv. The transportation papers therefore use “MUNI” not as a modeling primitive but as the municipal operator whose policy levers, service levels, and equity constraints structure the optimization or game (Zambrano et al., 6 Feb 2025).

6. MUNI as multimodal unified latent diffusion

In machine learning, “MUNI” names an end-to-end multimodal latent diffusion framework for any-to-any generation. The problem is to model both subset-conditioned generation $r^l_{i,t}\in\mathbb{R}$0 for all subsets $r^l_{i,t}\in\mathbb{R}$1 and unconditional joint sampling $r^l_{i,t}\in\mathbb{R}$2. MUNI introduces modality-specific encoders $r^l_{i,t}\in\mathbb{R}$3, modality-specific expressive decoders $r^l_{i,t}\in\mathbb{R}$4, and a single shared flow-based prior $r^l_{i,t}\in\mathbb{R}$5, with factorized decoder likelihood

$r^l_{i,t}\in\mathbb{R}$6

Subset posteriors are built by aggregating unimodal experts, using product aggregation or Hellinger aggregation. Prior training uses conditional flow matching on the linear path $r^l_{i,t}\in\mathbb{R}$7, with loss

$r^l_{i,t}\in\mathbb{R}$8

and encoder-side ELBO-correct weighting $r^l_{i,t}\in\mathbb{R}$9.

The paper’s main claim is that standard multimodal variational aggregation is insufficient once a learned prior and expressive decoders are coupled. MUNI therefore introduces a routed objective with three structural choices: non-mixture aggregation, target-detached self-reconstruction via stop-gradient on the target modality encoder, and prior learning only on full and leave-one-out routes. The routed objective is

$d^l_{i,t}\in\mathbb{R}$0

The latent is explicitly aligned with coherence sufficiency, predictive sufficiency, and minimality, expressed by conditions such as $d^l_{i,t}\in\mathbb{R}$1 and $d^l_{i,t}\in\mathbb{R}$2 for missing modalities.

Empirically, on PolyMNIST-Quadrant-Labels, MUNI reports verifier accuracies of 0.9131 for single-label-to-image digit conditioning, 0.9999 for quadrant conditioning, 0.9346 for multi-label-to-image generation, and 0.4841 for unconditional coherence. On a large-scale image–text–audio benchmark trained with pairwise data only, MUNI reports many-to-one alignment of 93.42 AIS for $d^l_{i,t}\in\mathbb{R}$3 and 87.29 AIS for $d^l_{i,t}\in\mathbb{R}$4, outperforming FlowBind and OmniFlow in those reported settings. The framework does not require fully paired triplets in principle and does not rely on text-aligned embeddings or deterministic matched-dimensionality mappings. Its stated limitations include computational cost, modality imbalance, possible loss of fine-grained detail under strong latent compression, and the fact that the general objective does not exploit known asymmetric cross-modal structure (Yeo et al., 15 Jun 2026).