MaM Library: Multi-Domain Reusable Resources

• MaM Library is defined as a family of reusable research resources that span distinct forms—from traditional software libraries to empirical corpora.
• In cosmology, MiMeS exemplifies a header-only C++ library with a Python interface that numerically solves axion misalignment mechanisms using adaptive Runge–Kutta methods.
• In medical AI, systems like MAM-CLIP and MAM-AI bundle code, pretrained weights, and datasets to enhance tasks such as mammography analysis and clinical diagnosis.

Searching arXiv for papers mentioning “MaM” or “library” to ground the article in current literature. The current literature suggests that “MaM Library” is not a single standardized artifact but a family resemblance across several reusable research resources. In the most literal sense, it denotes MiMeS, a header-only C++ library with a thin Python interface for solving the axion or axion-like particle misalignment mechanism in arbitrary cosmological histories. In medical AI, the same expression is used more loosely for reusable codebases, pretrained weights, datasets, prompt suites, curated corpora, and evaluation harnesses associated with systems named MAM or MaM, including mammography models, on-device retrieval-augmented generation, and modular multi-agent diagnosis frameworks (Karamitros, 2021, Gulluk et al., 19 May 2026, Ren, 28 Jun 2026, Zhou et al., 24 Jun 2025).

1. Terminological scope and disciplinary spread

Across current research usage, “library” appears in at least three distinct senses: a software library in the conventional API-centered sense, a reusable model-and-code package, and a curated empirical or textual corpus distributed for downstream scientific use. The following artifacts illustrate that spread.

Artifact	Domain	Reusable form
MiMeS	Axion/ALP cosmology	Header-only C++ library with Python interface
MAM-CLIP	Mammography AI	Code, pretrained weights, extraction scripts, released TEKNOFEST PNGs
MAM-AI	On-device medical QA	Android app, guideline corpus, benchmarks, evaluation harness
MAM-E	Mammographic synthesis	Source code, pretrained weights, graphical user interfaces
MAM	Multimodal medical diagnosis	Code for a modular multi-agent framework

A nearby but distinct nomenclatural case is MaStar, the MaNGA Stellar Library. MaStar is a large empirical stellar library with more than 8000 stars, wavelength coverage $3622$–$10{,}354$ Å, spectral resolution $R \approx 2000$, and relative flux calibration better than $3\%$, built to match MaNGA spectroscopy for spectral modeling and stellar population analysis (Yan et al., 2017). It is not usually presented as a “MaM Library,” but it illustrates how “library” in astronomy denotes a calibrated empirical reference set rather than an executable software framework.

A plausible source of confusion is therefore nomenclatural overlap rather than methodological continuity. The common element is reusability; the technical content, evaluation criteria, and intended users vary sharply by field.

2. MiMeS as the literal misalignment-mechanism library

MiMeS is the clearest instance in which “MaM Library” denotes a conventional scientific software library. It is a header-only C++ library, with a thin Python interface, designed to solve the axion or axion-like particle misalignment mechanism in arbitrary cosmological histories. The library numerically integrates the full nonlinear axion equation of motion, identifies the onset of oscillations, tracks the adiabatic regime, and computes the relic abundance including anharmonic effects and entropy injection (Karamitros, 2021).

The underlying dynamics are expressed in terms of the axion angle $\theta$, with

$$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$

and the potential

$$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$

MiMeS recasts the system in the dimensionless variable

$u \equiv \log\frac{a(t)}{a_{\rm ini}},$

using a first-order system for $y=(\theta,\zeta)$ with $\zeta \equiv d\theta/du$. This formulation is explicitly cosmology-agnostic because the code consumes user-provided interpolants of $10{,}354$0 and $10{,}354$1, rather than assuming radiation domination or any other built-in history. The input file tabulates $10{,}354$2, $10{,}354$3 in GeV, and $10{,}354$4, which allows radiation domination, early matter domination, kination, decaying-fluid cosmologies, late reheating, and other nonstandard backgrounds to be supplied directly.

Mass modeling is similarly modular. MiMeS supports a data-driven QCD axion mass via the lattice topological susceptibility $10{,}354$5, with

$10{,}354$6

or a user-supplied function $10{,}354$7 with arbitrary functional form. The library ships Standard Model thermodynamics through the Cosmo class and supports callbacks for extrapolations outside the tabulated mass range. Its numerical core uses NaBBODES with adaptive embedded Runge–Kutta methods, including the Rosenbrock–Wanner solvers ROS34PW2 and RODASPR2 for stiff systems and Dormand–Prince as an explicit embedded method. Recommended tolerances are at or below $10{,}354$8; for very high $10{,}354$9 and large $R \approx 2000$0, long double precision and a Rosenbrock method are recommended.

MiMeS exposes scientifically interpretable outputs rather than only final abundances. These include $R \approx 2000$1, $R \approx 2000$2, $R \approx 2000$3, $R \approx 2000$4, the entropy injection factor $R \approx 2000$5, the full time series $R \approx 2000$6, and per-peak diagnostics including the adiabatic invariant $R \approx 2000$7. The repository is released under the MIT license, version $R \approx 2000$8, with stable releases at HepForge and a documented C++ and Python interface. In this domain, “MaM Library” is therefore an actual computational library in the ordinary software-engineering sense.

3. Mammography-centered reusable resources

In mammography, “library” more often denotes a reusable model-and-resource package than a standalone numerical library. MAM-CLIP is a vision-LLM and an accompanying reusable resource for mammography AI, designed to improve BI-RADS classification by pretraining on mammography atlases and then fine-tuning for prediction. It curates $R \approx 2000$9 image-caption pairs from the Atlas of Mammography and the ACR BI-RADS Atlas, uses a ConvNeXt visual encoder with a pretrained PubMedBERT language encoder, and trains with a CLIP-style contrastive objective plus a masked language modeling loss that uses image embeddings to help predict masked words (Gulluk et al., 19 May 2026).

The project is explicitly “library-like” in two senses: it provides code, pretrained weights, and data extraction and preprocessing scripts, and it exposes a mammography-specific vision-language backbone that can be reused for downstream tasks such as BI-RADS prediction or density classification. Pretraining is performed with batch size $3\%$0, AdamW, learning rate $3\%$1, $3\%$2 epochs, and a single NVIDIA A100. Fine-tuning on TEKNOFEST and EMBED uses 4-fold cross-validation and macro-averaged F1. Reported gains range from $3\%$3 macro F1 with approximately $3\%$4K training samples to $3\%$5 with $3\%$6K samples. An especially notable result is that, with more than $3\%$7K labeled samples, pretraining on $3\%$8 atlas image-text pairs yields about $3\%$9 average macro F1 over adding $\theta$0 more labeled images to the ImageNet baseline. This positions the released artifact not merely as a model checkpoint but as a reusable mammography pretraining resource.

MAM-E occupies a different part of the mammography pipeline: synthetic image generation and lesion inpainting. It uses Stable Diffusion with DreamBooth-style fine-tuning from stable-diffusion-v1-5, a frozen latent VAE, a UNet with cross-attention conditioned by a fine-tuned CLIP text encoder, and a mask-aware latent concatenation scheme for inpainting. Images are resized to $\theta$1, encoded into $\theta$2 latents, and lesion inpainting stacks $\theta$3 along channels. The prompts encode view, breast density or breast area size, and vendor, while two graphical user interfaces support healthy mammogram synthesis and lesion drawing or inpainting (Montoya-del-Angel et al., 2023).

Its empirical basis is the combined use of OMI-H and a Siemens subset of VinDr-Mammo, totaling $\theta$4 images, of which $\theta$5 are healthy and $\theta$6 contain lesions. Hyperparameters explored include batch sizes $\theta$7–$\theta$8, training steps $\theta$9–$$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$0, and learning rates $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$1, $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$2, and $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$3, with MSE on $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$4-prediction. The radiologist reader study reports AUROC $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$5 for distinguishing real from synthetic images on $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$6 images with balanced real and synthetic distribution, while XAI heatmaps from a CAD classifier highlight synthetic lesions similarly to real lesions. In this setting, the “library” concept is embodied by released code, weights, and GUIs rather than by a formal API abstraction.

4. Clinical AI system releases: from on-device RAG to modular diagnosis

MAM-AI presents another meaning of “MaM Library”: a released, fully offline clinical AI system together with its knowledge base, benchmarks, and evaluation tools. The system runs entirely on a commodity Android device. A query is embedded with EmbeddingGemma-300M, matched against a curated corpus of $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$7 guideline documents and $$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$8 heading-aligned passages in a local SQLite vector store, and answered with citations by Gemma 4 E4B int4. Retrieval returns top-$$\ddot{\theta}(t)+3H(t)\dot{\theta}(t)+m_a^2(T)\sin\theta(t)=0,$$9 passages by cosine similarity; the bundle is checksum-pinned, every source document has a SHA-256 checksum in the manifest, and every passage carries a truncated SHA-256 content hash for auditable citations (Ren, 28 Jun 2026).

The release is unusually broad: the Android application, the medical guideline build pipeline, the benchmark datasets mamabench and mamaretrieval, and an evaluation harness for retriever-only, oracle-context, end-to-end, and latency tests. On the reported device, a OnePlus tablet with Snapdragon 8 Elite, $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$0 GB RAM, and Android 15, latency with Gemma 4 E4B is about $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$1 s median total on GPU and about $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$2 s median total on CPU at $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$3. The evaluation identifies a central architectural finding: on-device retrieval is already strong, while the small generator remains the limiting component. EmbeddingGemma-300M ranks third of seven retrievers, lenient $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$4, weighted precision $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$5, and strict hit rate $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$6. By contrast, adding retrieved context does not improve the small generator’s answers on average. The deployed G1 prompt nevertheless reduces deflection from approximately $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$7 to approximately $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$8, while Gemma 4 is chosen over a more helpful same-size alternative because it is about twice as faithful to its sources and produces fewer dangerous clinical errors. Here the “library” is best understood as a bundled research release for offline point-of-care RAG.

A related but architecturally distinct medical usage is MAM, the Modular Multi-Agent Framework for Multi-Modal Medical Diagnosis. This system decomposes diagnosis into five LLM-based agents: a General Practitioner, Specialist Team, Radiologist, Medical Assistant, and Director. The workflow is explicitly formalized: $$V(\theta,T)=m_a^2(T)f_a^2\left[1-\cos\theta\right].$$9

$u \equiv \log\frac{a(t)}{a_{\rm ini}},$0

$u \equiv \log\frac{a(t)}{a_{\rm ini}},$1

$u \equiv \log\frac{a(t)}{a_{\rm ini}},$2

followed by unanimity voting with $u \equiv \log\frac{a(t)}{a_{\rm ini}},$3 and consensus when $u \equiv \log\frac{a(t)}{a_{\rm ini}},$4, after which $u \equiv \log\frac{a(t)}{a_{\rm ini}},$5 (Zhou et al., 24 Jun 2025).

The framework supports text, image, audio, and video modalities and uses off-the-shelf components such as Medichat-Llama3-8B, HuatuoGPT-Vision-7B, Qwen-Audio-Chat, and VideoLLaMA2-7B. No new loss function or additional fine-tuning under the framework is reported. Across MedQA, PubMedQA, Brain Tumor MRI, DeepLesion, NIH Chest X-rays, PathVQA, PMC-VQA, Heartbeat, SoundDr, and MedVidQA, reported improvements relative to baselines range from $u \equiv \log\frac{a(t)}{a_{\rm ini}},$6 to $u \equiv \log\frac{a(t)}{a_{\rm ini}},$7. Supporting studies report $u \equiv \log\frac{a(t)}{a_{\rm ini}},$8–$u \equiv \log\frac{a(t)}{a_{\rm ini}},$9 gains from role assignment over direct prompting and $y=(\theta,\zeta)$0–$y=(\theta,\zeta)$1 gains in diagnostic discernment over random-choice baselines. Although presented as a framework rather than a “library” in the strict sense, its code release, prompt templates, swappable modules, and explicit message-passing interfaces make it library-like in practical deployment.

5. Other technical meanings of MAM that are not libraries

Not every prominent “MAM” in the literature denotes a library. In speech translation, MAM means Masked Acoustic Modeling, a self-supervised objective for end-to-end speech-to-text translation. The method masks $y=(\theta,\zeta)$2 of acoustic frames from 80-dimensional log-Mel filterbanks plus 3-dimensional pitch features and trains the encoder to reconstruct the original acoustic signal through a deconvolutional head. The combined objective is

$y=(\theta,\zeta)$3

with reconstruction

$y=(\theta,\zeta)$4

On MuST-C over eight English-to-target directions, joint training with span masking improves by $y=(\theta,\zeta)$5 BLEU on average without transcriptions, English-speech pretraining on Libri-Light improves by $y=(\theta,\zeta)$6 BLEU, and arbitrary acoustic pretraining on AudioSet improves by $y=(\theta,\zeta)$7 BLEU (Chen et al., 2020). This MAM is a training objective, not a software library, even though it is amenable to library implementation.

In the theory of abstract machines, MAM denotes the Maximal MAM, a reasonable implementation of the maximal strategy of the lambda-calculus. Its state is

$y=(\theta,\zeta)$8

with frame $y=(\theta,\zeta)$9, current code $\zeta \equiv d\theta/du$0, argument stack $\zeta \equiv d\theta/du$1, global environment $\zeta \equiv d\theta/du$2, and phase $\zeta \equiv d\theta/du$3. The machine implements maximal $\zeta \equiv d\theta/du$4-reduction through commutative, exponential, and multiplicative transitions, and the paper proves a weak bisimulation between machine execution and maximal $\zeta \equiv d\theta/du$5-reduction. Its overhead bound is

$\zeta \equiv d\theta/du$6

with the same asymptotic RAM-step cost (Accattoli, 2017). This is a formal abstract machine, not a reusable software package, but it contributes to the ambiguity of the acronym.

These cases matter because they show that “MAM” alone is not a reliable indicator of software packaging, release practice, or disciplinary lineage. The same acronym names an optimization objective, an abstract machine, a medical multi-agent framework, and a set of clinical AI resource bundles.

6. Common themes, limitations, and interpretive cautions

A plausible unifying theme across these artifacts is modular reuse rather than a shared technical core. MiMeS separates cosmology, mass modeling, and solver choice; MAM-CLIP separates vision-language pretraining from downstream BI-RADS fine-tuning; MAM-AI separates embedder, corpus, generator, and evaluation harness; the multi-agent medical MAM separates triage, retrieval, diagnosis, synthesis, and voting. In each case, the reusable unit is not merely data or code in isolation but a composable research workflow.

The limitations are equally domain-specific. MiMeS assumes the standard misalignment initial condition $\zeta \equiv d\theta/du$7 and does not support kinetic misalignment in version $\zeta \equiv d\theta/du$8. MAM-CLIP pretraining uses only $\zeta \equiv d\theta/du$9 atlas pairs, and domain shift between educational atlases and clinical screening remains a consideration. MAM-E currently operates at $10{,}354$00 and 8-bit depth, which the authors identify as a limitation for tasks requiring high-resolution contrast. MAM-AI identifies the small generator, rather than retrieval, as the binding constraint for safe and helpful on-device RAG, and the system is explicitly a research prototype rather than a fielded product. The medical multi-agent MAM has no real-world clinical validation yet and remains constrained by the biases and error modes of its base LLMs and LVLMs.

A frequent misconception would be to treat “MaM Library” as the name of a single codebase. The literature instead supports a narrower interpretation in some fields and a broader one in others. In particle cosmology, it can mean a true software library; in medical AI, it often means a released ecosystem of models, corpora, prompts, and evaluation assets; in other areas, “MAM” may not denote a library at all. This suggests that the term is best read contextually, with the surrounding discipline determining whether “library” refers to an executable API, an empirical reference corpus, or a reproducible research package.