---
title: MOST Across Research Domains
url: https://www.emergentmind.com/topics/most
type: topic
---

# MOST Across Research Domains

Searching arXiv for the cited MOST-related papers to ground the article in the corresponding records.
Search query: 1402.1570 OR 2601.10272 OR 2005.12753 OR 2403.04099 OR 2507.06590
In arXiv literature, **MOST** is not a single standardized term but a reused label spanning several unrelated technical objects. It denotes a topological extremal problem on arcs and curves, a density-based set operator for countable families, an optimal-transport framework for many-objective optimization, a speech–text Mixture-of-Experts architecture, and a text-to-motion diffusion system for rare prompts; the string “most” also appears non-acronymically in astrophysical titles concerning ultra-compact dwarf galaxies [1402.1570] [2005.12753] [2403.04099] [2601.10272] [2507.06590] [2003.02383]. The term therefore requires field-specific disambiguation.

## 1. Disambiguation across research areas

The collected usages of MOST, MoST, and MosT can be organized by domain as follows.

| Form | Domain | Meaning in the cited literature |
|---|---|---|
| MOST | Topology | Maximum cArdinality of families of Arcs or curves with bounded pairwise intersection |
| most-intersection | Logic / set theory | Set operator defined via natural density over countable families |
| MosT | Many-objective optimization | Many-objective Multi-solution Transport |
| MoST | Multimodal LLMs | Mixture of Speech and Text with Modality-Aware Mixture of Experts |
| MOST | Text-to-motion generation | Motion diffusion model via Temporal Clip Banzhaf Interaction |
| most | Astrophysics | Descriptive adjective in “most luminous” and “most massive” UCDs |

A common source of confusion is the visual similarity of these labels. The cited papers show that capitalization does not standardize meaning: the same four-letter string is reused independently in topology, logic, optimization, multimodal sequence modeling, and generative motion synthesis. This suggests that the correct interpretation of MOST is determined almost entirely by disciplinary context.

## 2. MOST in topology: maximal arc and curve families

In topology, MOST refers to the extremal problem studied by Przytycki for punctured oriented surfaces with Euler characteristic $\chi<0$ [1402.1570]. A punctured oriented surface of finite type has
$$
\chi(S)=2-2g-n,
$$
where $g$ is the genus and $n$ the number of punctures. An arc is a proper map $\alpha:(-\infty,+\infty)\to S$ limiting at $\pm\infty$ to punctures; it is **simple** if embedded and **essential** if not homotopic, relative to endpoints, into a cusp neighborhood. Two arcs are compared by geometric intersection number $i(\alpha,\beta)$.

The central theorem gives the exact maximum size of a family $\mathcal A$ of essential, simple, pairwise non-homotopic arcs with pairwise intersection number at most one:
$$
M(\chi)=2|\chi|(|\chi|+1).
$$
The upper bound uses a complete hyperbolic metric of area $2\pi|\chi|$, together with abstract “nibs,” each an ideal hyperbolic triangle of area $\pi$. There are exactly $2|\mathcal A|$ nibs, and a slit-and-overlap argument shows that no point of the surface lies in more than $2(|\chi|+1)$ nibs. The lower bound is sharp: after cutting along $|\chi|+1$ disjoint arcs, one obtains an ideal polygon with $2|\chi|+2$ sides, and the set of all diagonals together with the cutting arcs realizes exactly $2|\chi|(|\chi|+1)$ arcs.

The paper extends the $k=1$ analysis to closed orientable surfaces. If $S$ is closed of genus $g$ and $\mathcal C$ is a family of non-peripheral essential simple closed curves that are pairwise non-homotopic and satisfy $i(\gamma,\gamma')\le 1$, then
$$
|\mathcal C|\le g\,[2M(\chi)+1]+|\chi|-1,
$$
with $M(\chi)=2|\chi|(|\chi|+1)$, hence $|\mathcal C|=O(|\chi|^3)$. For uniformly bounded intersection number $k\ge 1$, the asymptotic bounds become
$$
|\mathcal A|=O_k(|\chi|^{k+1})
\quad\text{and}\quad
|\mathcal C|=O_k(|\chi|^{k^2+k+1}).
$$

A distinguished special case is the punctured sphere. If two punctures $p,p'$ are fixed, not necessarily distinct, then the maximum size of a family of essential simple arcs from $p$ to $p'$ with pairwise intersection at most one is
$$
\tfrac12 |\chi|(|\chi|+1).
$$
Small-$|\chi|$ examples illustrate sharpness: for $|\chi|=1$, one obtains $M(-1)=4$; for $|\chi|=2$, one obtains $M(-2)=12$; and on the torus $g=1$, one recovers the classical fact that at most $3$ non-homotopic simple closed geodesics can pairwise meet at most once.

## 3. MOST in logic and set theory: the most-intersection operator

In the paper by Çevik and Topal, MOST denotes the logical quantifier “most” implemented as a set operator on countable families via natural density [2005.12753]. For a subset $A\subseteq\mathbb N$, lower and upper asymptotic densities are
$$
\underline d(A)=\liminf_{n\to\infty}\frac{|A\cap\{1,2,\dots,n\}|}{n},
\qquad
\overline d(A)=\limsup_{n\to\infty}\frac{|A\cap\{1,2,\dots,n\}|}{n},
$$
and when these coincide, the natural density is
$$
d(A)=\lim_{n\to\infty}\frac{|A\cap\{1,2,\dots,n\}|}{n}.
$$
The construction is restricted to sets for which this limit exists.

For a sequence $\mathcal F=\{A_1,A_2,\dots\}$ of subsets of a universe $U$, the **most-intersection** operator is
$$
\MostInt(\mathcal F)=\bigl\{x\in U:\ d(\{\,i\in\mathbb N\mid x\in A_i\})>\tfrac12\bigr\}.
$$
In the finite case, this reduces to ordinary majority membership:
$$
\MostInt(\mathcal F)=\{x:\ |\{i:x\in A_i\}|>n/2\}
\quad\text{when }|\mathcal F|=n.
$$

The operator preserves several intersection-like laws. For two sets it is commutative and associative; it is idempotent in the sense that $\MostInt(\{A,A\})=A$; and it satisfies the unit laws $\MostInt(\{A,U\})=A$ and $\MostInt(\{A,\varnothing\})=\varnothing$. It is monotone under enlargement of the family:
$$
\mathcal F\subseteq\mathcal G \implies \MostInt(\mathcal G)\subseteq \MostInt(\mathcal F).
$$
It also contains the ordinary intersection:
$$
\bigcap_{A\in\mathcal F}A\subseteq \MostInt(\mathcal F).
$$
At the same time, the paper explicitly notes a limitation: complement does not distribute over most-intersection in general,
$$
\MostInt(\{A^c:A\in\mathcal F\})\ne (\MostInt(\mathcal F))^c.
$$

The examples clarify the departure from ordinary intersection. For $\mathcal F=\{A_1,A_2,A_3\}$ with
$$
A_1=\{1,2,3\},\quad A_2=\{2,3,5\},\quad A_3=\{3,4\},
$$
one computes
$$
\MostInt(\mathcal F)=\{2,3\}.
$$
For the infinite family where $A_i$ is the set of the first $i$ prime numbers, every prime belongs to most of the $A_i$, so
$$
\MostInt(\{A_i:i\in\mathbb N\})=\{\text{all primes}\},
$$
whereas the ordinary intersection is empty.

The operator is then applied in formal language theory and hypergraphs. For a countable family of deterministic finite automata with regular languages $L_i$, the density language
$$
L_{\mathrm{dens}}=\MostInt(\{L_i:i\in\mathbb N\})
$$
need not be regular even if every $L_i$ is regular. In hypergraphs $H=(V,E)$, the “average hyperedge” is defined by
$$
e_{\mathrm{avg}}=\MostInt(E).
$$
The paper also introduces a similarity relation $\MostSim$ on subsets of $\mathbb N$ and leaves the induced equivalence-class lattice, logical-programming implementations, and measure-theoretic generalizations as open directions.

## 4. MosT in many-objective optimization: objective–solution transport

In machine learning optimization, MosT is the **Many-objective Multi-solution Transport** framework for the regime of many objectives and comparatively few solutions, typically $n\gg m$ [2403.04099]. The motivating setting includes federated learning, multi-task learning, and mixture-of-prompt learning for LLMs. Rather than seeking a single model, the framework seeks $m$ Pareto-stationary solutions that collectively cover the $n$ objectives.

The formulation is bi-level. For each solution $s=1,\dots,m$, the upper level solves
$$
x_s\longleftarrow \arg\min_{x_s}\sum_{i=1}^n w_{i,s}f_i(x_s),
$$
where the weights are entries of an optimal transport plan $\Gamma\in\mathbb R_+^{n\times m}$. The lower-level problem is
$$
\Gamma\longleftarrow
\arg\min_{\Gamma\in\Pi(\alpha,\beta)}
\sum_{i=1}^n\sum_{s=1}^m \Gamma_{i,s}C_{i,s}
+\tau R(\Gamma),
$$
with $C_{i,s}=f_i(\theta_s)$, marginals
$$
\Pi(\alpha,\beta)=\{\Gamma\ge 0:\Gamma\mathbf 1_m=\alpha,\ \Gamma^\top\mathbf 1_n=\beta\},
$$
and an optional diversity regularizer
$$
R(\Gamma)=-\sum_{i=1}^n \max_{s\in[m]}\Gamma_{i,s}.
$$
Uniform $\alpha$ and $\beta$ are used to ensure coverage and balance.

Algorithmically, MosT alternates between an optimal-transport update and a Multi-Gradient Descent Algorithm update. The OT subproblem can be solved by IPOT or Sinkhorn. For each solution $s$, the MGDA step solves
$$
\min_{d\in\mathbb R^d}\max_{i\in[n]}
\langle \Gamma^t_{i,s}\nabla f_i(\theta_s^t),d\rangle+\tfrac12\|d\|^2
\quad\longleftrightarrow\quad
\min_{\lambda\in\Delta^n}
\Bigl\|\sum_i \lambda_i\,\Gamma^t_{i,s}\nabla f_i(\theta_s^t)\Bigr\|^2,
$$
and updates $\theta_s^{t+1}\gets \theta_s^t+\eta d_s$. The analysis includes a sparsity proposition stating that any unregularized OT solution has at most $n+m-1$ nonzero entries.

The convergence results separate non-convex and strongly convex regimes. Under $\nu$-smoothness and step size $\eta=1/(2\nu)$,
$$
\frac1T\sum_{t=1}^T\sum_{s=1}^m \beta_s\|d_s^t\|^2 = O\!\bigl(\tfrac1T\bigr),
$$
which implies convergence to a Pareto stationary set of solutions. Under $\mu$-strong convexity and stabilization of the nonzero pattern of $\Gamma$ with all nonzero entries bounded below by $\epsilon>0$, each solution converges linearly:
$$
\|\theta_s^{t+1}-\theta_s^*\|^2
\le (1-\mu\eta\epsilon)\,\|\theta_s^t-\theta_s^*\|^2.
$$

Empirically, MosT is evaluated on federated learning with $n=30$ or $206$ clients and $m=5$ models, multi-task learning with $n=4$ or $6$ tasks and $m=4$ solutions, and mixture-of-prompt learning with $n=128$ instances and $m=3$ soft prompts. The reported results state that MosT outperforms all baselines by **4–10 percent points** in average accuracy in federated learning, achieves **+2–3% absolute improvement** in average task accuracy in multi-task learning, improves test accuracy by **4–7 points** in mixture-of-prompt learning, and attains the highest hypervolume coverage on ZDT benchmarks and a fairness–accuracy trade-off setting. The discussion also notes computational limitations: each outer iteration requires solving both an OT problem and an MGDA subproblem per solution, and the convex analysis assumes stability of the $\Gamma$ sparsity pattern.

## 5. MoST in multimodal language modeling: speech–text Mixture of Experts

In multimodal language modeling, MoST denotes **Mixture of Speech and Text**, a speech–text LLM built on a sparse Mixture-of-Experts decoder with a **Modality-Aware Mixture of Experts** module [2601.10272]. The model augments a pretrained sparse MoE transformer decoder with a frozen HuBERT front end followed by a linear projection for speech, a standard text embedding layer, interleaved processing of speech and text tokens in shared decoder layers, and designated transformer blocks containing a MAMoE layer.

Each MAMoE layer receives a hidden state $h_t\in\mathbb R^d$ and modality indicator $m_t\in\{0,1\}$, with expert groups $\mathcal E_{\mathrm{text}}$, $\mathcal E_{\mathrm{audio}}$, and an always-active shared expert $E_{\mathrm{shared}}$. The final output is
$$
y_t=\sum_{j\in\mathcal I_t} w_{t,j}E_j(h_t)+E_{\mathrm{shared}}(h_t),
$$
where $\mathcal I_t=\mathrm{TopK}(p_t,K)$ is obtained from modality-masked gating. To encourage balanced expert usage within each modality group, the model adds
$$
\mathcal L_{\mathrm{balance}}
=
\lambda\sum_{g\in\{\mathrm{text},\mathrm{audio}\}}
\sum_{j\in\mathcal E_g}
\Bigl(\frac1B\sum_{t=1}^B p_{t,j}-\tfrac1{|\mathcal E_g|}\Bigr)^2.
$$
During mixed instruction fine-tuning, the loss is
$$
\mathcal L=\mathcal L_{\mathrm{instruction}}
+\alpha(\mathcal L_{\mathrm{ASR}}+\mathcal L_{\mathrm{TTS}})
+\mathcal L_{\mathrm{balance}}.
$$

The training pipeline is explicitly two-stage. Stage 1 initializes from **DeepSeek-v2 Lite**, described as a **3B-parameter MoE LLM**, and performs cross-modal post-training on **LibriHeavy (50 k h ASR corpus)**, **Common Voice v9.0**, and **VoxPopuli**. The task mix is **40 % ASR, 40 % TTS, 20 % text-only LM**, trained for **500 k steps** with **batch size 512**, **AdamW**, and **cosine LR**. Stage 2 constructs approximately **200 k multimodal instructions** by interrupted-dialogue synthesis on **SmolTalk** and text-to-speech conversion of open instruction datasets, then trains for **10 k steps** with **batch size 128** using a **40 % speech-text, 40 % text, 10 % ASR, 10 % TTS replay** mixture. Preprocessing uses **HuBERT base features (25 Hz)** and a **102 400-token vocabulary**.

The reported evaluation spans ASR, TTS, audio language modeling, and spoken question answering. On audio LM, MoST obtains an average score of **71.9**, exceeding **69.0** for Phi-4 Multimodal, **68.4** for LLaMA-Omni2, and **65.2** for MinMo. In spoken QA, the paper reports that MoST outperforms all baselines in both S$\to$T and S$\to$S settings; for **WebQ S$\to$T**, it reaches **58.2 %** versus the next best **51.8 %**, and for **WebQ S$\to$S**, **44.7 %** versus **37.5 %**. The ablation studies further report that, under controlled initialization from **Llama3.2 3B**, MAMoE yields **+7.3 % on ASR** and **+21.8 % on SQA** over dense baselines, and **+5–12 %** over vanilla MoE. The full MAMoE variant, combining modality masking and shared experts, gives the best validation and task losses, while routing entropy and Gini coefficients decrease more rapidly than in vanilla MoE. The paper characterizes MoST, to its knowledge, as the **first fully open-source speech–text large language model built on a Mixture of Experts architecture**.

## 6. MOST in text-to-motion generation: rare-text retrieval and diffusion

In text-to-motion synthesis, MOST denotes a **Motion diffusion model for Rare Text via Temporal Clip Banzhaf Interaction** [2507.06590]. The method addresses rare or unseen language prompts by combining a retrieval stage and a generation stage. The retrieval stage decomposes both text and motion into short clips and scores them using **Temporal Clip Banzhaf Interaction**, a cooperative-game-theoretic measure of clip-level coherence. The generation stage conditions a diffusion model not only on the text embedding but also on the retrieved motion clips through a motion prompt module.

At the entity level, a dual-stream encoder aligns texts and motions with a symmetric similarity score and a contrastive loss $\mathcal L_{\rm NCE}$. At the clip level, the method forms $T_s$ motion clips and $T_s$ text clips using a **1-D convolution + self-attention** pipeline, then defines a cooperative game over the clip set
$$
N=\{\tilde m^1,\dots,\tilde m^{T_s}\}\cup\{\tilde t^1,\dots,\tilde t^{T_s}\}.
$$
For a two-player coalition $\{i,j\}$, the Banzhaf interaction is
$$
I(\{i,j\})
=
\sum_{S\subseteq N\setminus\{i,j\}}
\rho(S)\Bigl[
v(S\cup\{i,j\})+v(S)-v(S\cup\{i\})-v(S\cup\{j\})
\Bigr],
$$
with $\rho(S)=1/2^{|N|-2}$. The retrieval objective is
$$
\mathcal L_{\rm R}=\mathcal L_{\rm NCE}+\lambda_{\rm B}\mathcal L_{\rm B},
$$
and the paper specifies **$\lambda_{\rm B}=0.5$**. At inference, the model retrieves the top $K_c$ motion clips for each text clip, with **$T_s=5$** and **$K_c=2$** in the reported experiments.

The generation model is a denoising diffusion probabilistic model conditioned on text prompt $c$ and retrieved motion prompts $R=\{R_i\}_{i=1}^{T_s}$. Training minimizes
$$
\mathcal L_{\rm G}
=
\mathbb E_{t,x_0,\epsilon}
\bigl\|x_0-\epsilon_\theta(x_t,t,c,R)\bigr\|_2^2,
$$
and classifier-free guidance combines conditional and unconditional predictions with guidance scale **$s\approx 3\!-\!5$**. The denoiser uses a Transformer backbone with self-attention on the noisy input, cross-attention to text, and cross-attention to each motion prompt $R_i$.

The evaluation uses **HumanML3D** with **14,616 motions**, **44,970 texts**, and approximately **28.6 h** of data, and **KIT-ML** with **3,911 motions**, **6,353 texts**, and approximately **10.3 h**. Retrieval metrics are Recall@1/2/5/10 and Median Rank; generation metrics are R-TOP@1/2/3, FID, MM-Dist, Diversity, and Multi-Modality. For rare texts, the paper introduces a weighted MM-Dist measure:
$$
\mathrm{W\!-\!MM}^j
=
10\,\frac{r_p^j}{\max r}\,\mathrm{MM}^j.
$$

The reported numbers are specific. On HumanML3D under the “all” retrieval protocol, MOST achieves **R@1 = 6.61%** versus **5.68%** for TMR and **MedR = 25** versus **28**. On small-batch retrieval, it reaches **R@1 = 69.2%** versus **67.16%**. On HumanML3D generation, MOST obtains **FID = 0.092** compared with **0.103** for ReMoDiffuse and **R-TOP1 = 0.526**. For the rare-text tail **0–5%**, it reports **FID = 0.66** versus **0.87** for ReMoDiffuse and **1.26** for Fg-T2M, and **W-MM = 16.8** versus **18.7**. The retrieved prompt length is approximately **39 frames**, described as **20% of ReMoDiffuse’s 196 frames**, while yielding higher cosine similarity to ground-truth motion features. In the user study, MOST wins significantly more often than ReMoDiffuse on text matching and realism with **$p<0.001$**.

The ablations identify the functional contribution of each component. Removing $\mathcal L_{\rm B}$ lowers **R@1 by approximately 1%** and raises FID on the **0–15%** tail from **0.34** to **0.96**. The setting **$T_s=5$** is reported as optimal, and **$K_c=2$** gives the best generalization. Cross-attention to retrieved prompts outperforms simple concatenation, with **FID 0.13 vs. 0.19** on all texts and **0.96 vs. 1.23** on rare texts. These results support the paper’s claim that fine-grained clip retrieval rather than whole-motion retrieval is central to rare-text performance.

## 7. Non-acronymic “most” in astronomy: luminous and massive UCDs

A distinct, non-acronymic use of the word appears in the astrophysical study of dark matter annihilation in the **most luminous** and **most massive** ultra-compact dwarf galaxies, namely **Fornax UCD3** and **M59 UCD3** [2003.02383]. The paper is not about MOST as an acronym, but it is relevant to the lexical disambiguation of the term in arXiv titles.

The study models five dark-matter density profiles: **NFW**, **NFW + inner spike**, **mini-spike**, **spike with tidal truncation**, and the **Frank** profile. The annihilation source term is
$$
Q(E,r)=\frac{\langle\sigma v\rangle}{2\,m_{\rm DM}^2}\,\rho^2(r)\,\frac{dN}{dE}.
$$
For Fornax UCD3, the paper takes a black-hole mass of **$3.3\times 10^6\,M_\odot$**, stellar velocity dispersion **$\sigma_*\simeq 33$ km/s**, and black-hole age **$\sim 5$ Gyr**, implying
$$
R_{sp}=\frac{G\,M_{\rm BH}}{\sigma_*^2}\sim 5\mbox{--}10\ {\rm pc}.
$$
The spike steepens the inner profile from an outer **$r^{-1}$** cusp to approximately **$r^{-2.3}$** between $r_{\rm sat}$ and $R_{sp}$, enhancing $\int \rho^2\,dV$ by several orders of magnitude relative to NFW alone.

The assumed dark-matter content is obtained by averaging estimates under the **Salpeter** and **Kroupa** stellar mass functions, yielding approximately **8%–32%** of the total mass within **200 pc**. The paper studies dark-matter masses **10 GeV** for annihilation to $\mu^+\mu^-$ and **34 GeV** for annihilation to $b\bar b$. In the radio analysis, the **VLA upper limit for M59 UCD3 at 5.8 GHz** is
$$
S_\nu\le 7.8\,\mu{\rm Jy}.
$$
Using that limit with **no diffusion** and **$B=7\,\mu$G**, the required cross sections for the **Frank** profile are approximately **$3\times 10^{-26}\,{\rm cm}^3/{\rm s}$** for $b\bar b$ at **34 GeV** and **$3\times 10^{-28}\,{\rm cm}^3/{\rm s}$** for $\mu^+\mu^-$ at **10 GeV**. In spike scenarios the required values are far smaller, ranging down to **$10^{-35}$–$10^{-32}\,{\rm cm}^3/{\rm s}$**, well below the canonical thermal value **$3\times 10^{-26}\,{\rm cm}^3/{\rm s}$**. The paper therefore argues that, in the absence of a strong $\gamma$-ray signature, synchrotron emission from annihilation products can provide sensitive indirect constraints, especially when black-hole–induced spikes are present.

Taken together, these usages show that MOST functions less as a unified concept than as a recurring label attached to unrelated formal, algorithmic, and descriptive constructs. For technical reading, the decisive question is not the string itself but the surrounding disciplinary vocabulary: Euler characteristic and arcs indicate the topological MOST; natural density indicates most-intersection; OT and MGDA indicate MosT in optimization; MAMoE and HuBERT indicate MoST in speech–text modeling; and Temporal Clip Banzhaf Interaction indicates MOST in rare text-to-motion generation.

Source: https://www.emergentmind.com/topics/most