MARCOS in Mathematics, MRI, and AI

• MARCOS is a multi-sense term that denotes distinct research contributions in algebra, MRI technology, continuous reasoning in AI, and even paraconsistent logic.
• In mathematics, it underpins studies such as maximal tori in Hochschild cohomology, orbit categories for categorical induction, and invariance in derived algebra through concrete model applications.
• In technology and AI, MARCOS encompasses open-source MRI control systems like MaRCoS/MaRGA and latent reasoning frameworks that leverage Markov chains for improved inference speed.

MARCOS is a label that appears in contemporary research in several distinct senses. In the material represented here, it denotes, first, the surname of researchers—especially E. N. Marcos and J. M. Marcos—whose work is embedded in representation theory, homological algebra, non-classical logic, and statistical physics; second, the open-source MRI console MaRCoS, expanded as MAgnetic Resonance COntrol System; and third, the large-language-model reasoning framework MARCOS, expanded as Markov Chain of Continuous Thoughts. These usages are unrelated in subject matter but each has acquired technical meaning in its own literature (Briggs et al., 2021, Negnevitsky et al., 2022, Liu et al., 29 Sep 2025, Marcos et al., 2024).

1. E. N. Marcos in representation theory of finite-dimensional algebras

A central Marcos-associated line of work concerns the relationship between diagonalizable derivations, maximal tori in Hochschild cohomology, and fundamental groups of presentations of finite-dimensional algebras. In "Maximal tori in $HH^1$ and the fundamental group" (Briggs et al., 2021), Marcos appears as one of the main predecessors through earlier work with Farkas and Green on diagonalizable derivations of finite-dimensional algebras and their realization via additive characters on $\pi_1(Q,I)$. The paper states that its main theorem extends work of Farkas–Green–Marcos, of de la Peña–Saorín, and of Le Meur, and removes earlier restrictions by working for arbitrary finite-dimensional algebras over an algebraically closed field.

The main conceptual step is the identification of maximal tori in $HH^1(A)$ with duals of fundamental groups of presentations. The paper defines diagonalizable subalgebras of $HH^1(A)$ and proves that maximal diagonalizable subalgebras are exactly maximal tori. Its fundamental theorem states that every maximal torus of $HH^1(A)$ is of the form $\operatorname{im}(\theta_\nu)$ for some presentation $\nu:kQ\to A$, equivalently that every maximal diagonalizable subalgebra of $HH^1(A)$ is isomorphic to

$\operatorname{Hom}(\pi_1(Q,I),k^+)$

for some presentation $A\cong kQ/I$ (Briggs et al., 2021).

This leads to the equality

$\pi_1(Q,I)$0

where $\pi_1(Q,I)$1 is defined by maximizing $\pi_1(Q,I)$2 over minimal presentations, and $\pi_1(Q,I)$3 is the maximal dimension of a torus in $\pi_1(Q,I)$4. The paper then deduces that $\pi_1(Q,I)$5-rank$\pi_1(Q,I)$6 is a derived invariant, and that among self-injective algebras it is also invariant under stable equivalence of Morita type (Briggs et al., 2021).

The same paper places Marcos in a second historical line through the Bardzell–Marcos theorem on monomial algebras. Bardzell and Marcos had proved that if $\pi_1(Q,I)$7 is a finite-dimensional monomial algebra and $\pi_1(Q,I)$8, then $\pi_1(Q,I)$9 is a tree. The new result strengthens this to

$HH^1(A)$0

and extends it from monomial to semimonomial algebras. It also proves that if two finite-dimensional monomial algebras are derived equivalent, then their Gabriel quivers contain the same number of arrows, and consequently there are only finitely many monomial algebras in any given derived equivalence class (Briggs et al., 2021).

2. Cibils–Marcos orbit categories and categorical Clifford theory

A second major Marcos-associated theme is the orbit category introduced by Cibils and Marcos. In "Clifford's theorem for orbit categories" (Zimmermann, 2022), Zimmermann studies an $HH^1(A)$1-linear Krull–Schmidt category $HH^1(A)$2 with a finite group $HH^1(A)$3 acting by automorphisms and treats the orbit category

$HH^1(A)$4

as the categorical setting for a Clifford-theoretic decomposition theory. The paper explicitly says that it “provides the orbit category introduced by Cibils and Marcos,” later studied by Keller in cluster theory and by Asashiba in covering theory (Zimmermann, 2022).

Its decisive reinterpretation is monadic. If

$HH^1(A)$5

then the orbit category is naturally isomorphic to the Kleisli category $HH^1(A)$6, with

$HH^1(A)$7

This identifies the Cibils–Marcos orbit category as the Kleisli category of the monad naturally attached to the group action (Zimmermann, 2022).

The paper then formulates a categorical Clifford theorem. For an indecomposable $HH^1(A)$8, it defines the inertia subgroup

$HH^1(A)$9

constructs adjoint pairs between $HH^1(A)$0 and $HH^1(A)$1, and proves that if $HH^1(A)$2 is an indecomposable direct factor of the image of $HH^1(A)$3 in the inertia-level orbit category, then its image in the full orbit category is indecomposable. In this way, the Cibils–Marcos construction becomes the categorical analogue of induction from a normal subgroup to the ambient group (Zimmermann, 2022).

This places Marcos in a precise categorical lineage: the orbit category is not treated merely as notation, but as the correct receptacle for induction/restriction, inertia subgroups, Karoubi completion, and decomposition of indecomposables.

3. Homological algebra: $HH^1(A)$4-Hochschild theory, $HH^1(A)$5-Koszulity, and homotopy liftings

Marcos also appears centrally in several homological programs. In " $HH^1(A)$6-Hochschild (co)homology, the square of the Serre bimodule, and the Coxeter automorphism of the Tamarkin--Tsygan calculus" (Armenta, 12 Jul 2026), Armenta builds on the $HH^1(A)$7-Hochschild theory introduced by Cibils, Lanzilotta, Marcos and Solotar. The defining higher translates are

$HH^1(A)$8

and the paper proves that, for a minimal $HH^1(A)$9-projective resolution $HH^1(A)$0, the Nakayama transform $HH^1(A)$1 represents $HH^1(A)$2, while

$HH^1(A)$3

This yields the short exact sequence

$HH^1(A)$4

thereby separating a derived shadow from a strictly Morita-theoretic residue (Armenta, 12 Jul 2026).

The same paper states that, in top degree $HH^1(A)$5,

$HH^1(A)$6

and for hereditary $HH^1(A)$7,

$HH^1(A)$8

For self-injective algebras, the derived part vanishes identically, so

$HH^1(A)$9

Armenta also notes that taking Euler characteristics in the Cibils–Lanzilotta–Marcos–Solotar dimension formulas recovers Happel’s trace formula (Armenta, 12 Jul 2026).

In a different homological direction, Lü’s "Quasi-Koszulity and minimal Horseshoe Lemma" (Lu, 2011) is explicitly presented as a nongraded extension of the $\operatorname{im}(\theta_\nu)$0-Koszul program initiated by Green and Marcos. The paper defines a quasi-$\operatorname{im}(\theta_\nu)$1-Koszul module by radical conditions on a minimal projective resolution: $\operatorname{im}(\theta_\nu)$2 and proves that for a short exact sequence

$\operatorname{im}(\theta_\nu)$3

in the category of quasi-$\operatorname{im}(\theta_\nu)$4-Koszul modules, the minimal Horseshoe Lemma holds if and only if

$\operatorname{im}(\theta_\nu)$5

This is the radical-theoretic criterion extracted from the Green–Marcos degree-pattern idea (Lu, 2011).

Marcos also appears as a foundational source in Oke’s "Bracket structure on Hochschild cohomology of Koszul quiver algebras using homotopy liftings" (Oke, 2021). There, E. L. Green, G. Hartman, E. N. Marcos and Ø. Solberg are cited for providing a canonical minimal projective bimodule resolution of a Koszul quiver algebra. Oke uses the comultiplicative structure of that resolution to define homotopy lifting maps and compute Gerstenhaber brackets directly on the minimal resolution, including degree-$\operatorname{im}(\theta_\nu)$6 and degree-$\operatorname{im}(\theta_\nu)$7 cocycles and applications to Hochschild $\operatorname{im}(\theta_\nu)$8-cocycles satisfying the Maurer–Cartan equation (Oke, 2021).

Taken together, these papers show Marcos-associated work spanning higher Auslander–Reiten translates, Morita invariance questions, radical criteria for minimality, and explicit minimal-resolution techniques for Gerstenhaber structures.

4. Marcos and the logic $\operatorname{im}(\theta_\nu)$9

In non-classical logic, Marcos is associated with the identification and naming of $\nu:kQ\to A$0, a Sette-like paraconsistent logic. "Mortensen Logics" (Estrada-González et al., 2022) states that Mortensen proposed a logic $\nu:kQ\to A$1, and that Marcos later modified it to obtain a variant of Sette’s logic, identified and called $\nu:kQ\to A$2 by Marcos. The paper specifies the modification: replace the interpretation $\nu:kQ\to A$3 by the interpretation $\nu:kQ\to A$4, make the value $\nu:kQ\to A$5 designated along with $\nu:kQ\to A$6, and put $\nu:kQ\to A$7 instead of $\nu:kQ\to A$8, thereby obtaining $\nu:kQ\to A$9 (Estrada-González et al., 2022).

The same paper studies the connexive enrichment $HH^1(A)$0, obtained by replacing the $HH^1(A)$1-conditional with Mortensen’s $HH^1(A)$2-conditional. It remarks that M3V and $HH^1(A)$3 coincide in the $HH^1(A)$4-fragment, but differ because $HH^1(A)$5 retains the Sette-like $HH^1(A)$6 and $HH^1(A)$7 (Estrada-González et al., 2022).

This difference matters for inconsistency and connexivity. The paper’s headline comparison is that the inconsistency of M3V is attenuated in the connexive variant of the Sette-like $HH^1(A)$8, whereas it is exacerbated in the connexive variant of closed set logic. Concretely, the paper states that Aristotle’s Second Thesis and Abelard’s Principle,

$HH^1(A)$9

are valid in M3V but not in $\operatorname{Hom}(\pi_1(Q,I),k^+)$0 (Estrada-González et al., 2022).

The paper also uses $\operatorname{Hom}(\pi_1(Q,I),k^+)$1 in a broader claim about conditional semantics. Mortensen’s $\operatorname{Hom}(\pi_1(Q,I),k^+)$2-conditional is said to be “connexively stable,” meaning that it remains connexive when combined with the main paraconsistent negations $\operatorname{Hom}(\pi_1(Q,I),k^+)$3 and $\operatorname{Hom}(\pi_1(Q,I),k^+)$4. In that comparison, the Marcos-associated $\operatorname{Hom}(\pi_1(Q,I),k^+)$5 family supplies the $\operatorname{Hom}(\pi_1(Q,I),k^+)$6-based side of the argument (Estrada-González et al., 2022).

5. MaRCoS and MaRGA in low-field MRI

Outside mathematics and logic, the most prominent acronymic use is MaRCoS, the MAgnetic Resonance COntrol System. "MaRCoS, an open-source electronic control system for low-field MRI" (Negnevitsky et al., 2022) describes it as an open-source MRI console that combines hardware, firmware, and software for the integral control of MRI scanners. Its hardware foundation is the Red Pitaya SDRLab 122-16, based on a Xilinx Zynq-7020 system-on-chip with two embedded ARM processors, dual-channel 16-bit ADC, dual-channel 14-bit DAC, a 122.88 MHz crystal oscillator, two analog inputs for RX, two analog outputs for TX, and digital I/O. The paper states that MaRCoS can handle cycle-accurate sequences without hard length limitations, rapid bursts of events, and arbitrary waveforms, and that it does not use raster clocks or impose timing constraints beyond the hardware clock (Negnevitsky et al., 2022).

Its architecture is a streamed PC–server–FPGA design. The user works from a PC; marcos_client compiles sequence descriptions into binary instructions and sends them over Ethernet to the embedded MaRCoS server on the SDRLab; the server streams instructions into FPGA buffers and drains receive buffers; and the FPGA executes a generic timed-event language through an instruction decoder and sequencer connected to 24 output/timing FIFOs. The system supports arbitrary waveforms for RF envelope amplitude and phase, gradients, digital outputs, NCO frequencies, and receiver decimation settings (Negnevitsky et al., 2022).

"Benchmarking the performance of a low-cost Magnetic Resonance Control System at multiple sites in the open MaRCoS community" (Guallart-Naval et al., 2022) demonstrates the same console across three institutions and three scanners: a 360 mT tabletop educational/research MRI at MGH, a 50 mT human Halbach scanner at LUMC, and a 70–72 mT portable human Halbach scanner at i3M. The paper emphasizes native GUI programming, direct low-level sequence specification, and a PulSeq overlay, and shows implementations of SE, RARE, GRE, STIR, and radial non-Cartesian 3D SE. Its stated conclusion is that common sequences used in the clinic can be programmed into an open-source system relatively quickly and easily, and can produce good quality images even at this early stage of development (Guallart-Naval et al., 2022).

"MaRGA: a Graphical and Application Interface for MaRCoS" (Algarín et al., 2023) presents the user-facing layer built on top of MaRCoS. MaRGA is fully coded in Python 3, designed for intuitive scanner control, compatible with clinical environments, and based on an uncomplicated set of GUI panels and a renewed API. The API centers on the parent class mriBlankSeq, which provides sequence creation, execution, and data management; pulse-sequence classes run through sequenceRun, reconstruction and analysis through sequenceAnalysis, waveform validation and transfer through floDict2Exp, and export through saveRawData into DICOM 3.0 and .mat files (Algarín et al., 2023).

The GUI is organized around a Session window, Main window, and Post-processing window. New functionality includes session-aware metadata handling, multi-sequence protocols, comparison of reconstructions from the same session, image post-processing, and one-click automatic calibration of Larmor working frequency, RF coil efficiency, and gradient shim currents. The post-processing window supports $\operatorname{Hom}(\pi_1(Q,I),k^+)$7-space and image-space operations, including zero-padding, cosine-bell filters, BM4D denoising, Gaussian filtering, FFT, Partial Fourier by zero padding, POCS, and ART, with iterative methods programmed to be compatible with GPU execution. The paper reports operation on an experimental 0.2 T scanner designed for hard-tissue imaging and on a 72 mT portable scanner installed in the radiology department of a large hospital (Algarín et al., 2023).

A plausible implication is that MaRCoS and MaRGA together define not only an open console architecture, but an open workflow stack spanning sequence design, execution, calibration, reconstruction, metadata management, and clinical-format export.

6. MARCOS as latent reasoning in LLMs

A second acronymic use is MARCOS, expanded in "MARCOS: Deep Thinking by Markov Chain of Continuous Thoughts" (Liu et al., 29 Sep 2025) as Markov Chain of Continuous Thoughts. The paper begins from the chain-of-thought paradigm and argues that token-level reasoning is slow, creates a discrete bottleneck between steps, and entangles reasoning with token generation. MARCOS replaces token-level reasoning traces with a latent sequence of continuous thoughts, modeled as a conditional hidden Markov model in which explicit reasoning steps are observable variables and internal thoughts are hidden variables (Liu et al., 29 Sep 2025).

The input question $\operatorname{Hom}(\pi_1(Q,I),k^+)$8 is encoded by an “understander,”

$\operatorname{Hom}(\pi_1(Q,I),k^+)$9

and each latent reasoning step maintains two continuous neural states, $A\cong kQ/I$0 and $A\cong kQ/I$1. A learnable predictor $A\cong kQ/I$2 parameterizes a Gaussian prior over a step-level random variable $A\cong kQ/I$3,

$A\cong kQ/I$4

after which the “thinker” updates the latent thought state and the “speaker” can decode an explicit sentence from the shallow state (Liu et al., 29 Sep 2025).

Because this latent process is incompatible with standard supervised learning, the paper proposes a two-phase variational training scheme. A randomness encoder $A\cong kQ/I$5 infers posterior Gaussian parameters from the observed reasoning step $A\cong kQ/I$6, and training optimizes a reconstruction term, a KL term between the learned prior and posterior Gaussians, and a sparsity penalty

$A\cong kQ/I$7

The full objective is

$A\cong kQ/I$8

The paper states that the sparsity term is essential: removing it causes near-total model collapse, with accuracies dropping to around $A\cong kQ/I$9 (Liu et al., 29 Sep 2025).

Empirically, MARCOS is evaluated on GSM8K, SVAMP, and MultiArith. The headline claim is that it outperforms existing continuous reasoning methods and, for the first time, achieves performance comparable to token-based CoT, even surpassing it by $\pi_1(Q,I)$00 on GSM8K with up to $\pi_1(Q,I)$01 speedup in inference. On text supervision it reaches $\pi_1(Q,I)$02 GSM8K accuracy versus $\pi_1(Q,I)$03 for text CoT-SFT (0.5B), and on equation supervision it reaches $\pi_1(Q,I)$04 GSM8K accuracy. The paper also argues that MARCOS offers step-level instead of token-level control over randomness, opening opportunities for reinforcement learning and controlled reasoning (Liu et al., 29 Sep 2025).

The framework’s limitations are also explicit: experiments are restricted to GSM8K, SVAMP, and MultiArith; models are trained from scratch; absolute accuracies remain low compared with frontier systems; and the number of reasoning steps is fixed to $\pi_1(Q,I)$05 in the reported experiments (Liu et al., 29 Sep 2025).

7. J. M. Marcos and microscopic roughening of wetting fronts

The surname Marcos also appears in statistical physics through J. M. Marcos. "Microscopic fluctuations in the spreading fronts of circular wetting liquid droplets" (Marcos et al., 2024) studies the precursor fronts of nonvolatile liquid droplets spreading on solid substrates, now in circular geometry rather than the band geometry treated by an earlier paper by J. M. Marcos et al. The model is a lattice gas with Hamiltonian

$\pi_1(Q,I)$06

with a circular reservoir, two retained fluid layers, and kinetic Monte Carlo Kawasaki dynamics (Marcos et al., 2024).

The principal spreading observable is the average front position,

$\pi_1(Q,I)$07

implemented numerically as $\pi_1(Q,I)$08. For circular droplets the paper finds $\pi_1(Q,I)$09, with a stronger dependence on temperature and substrate wettability than in band geometry. The roughness is measured through

$\pi_1(Q,I)$10

and through the height-difference correlation function

$\pi_1(Q,I)$11

The paper concludes that circular droplets display intrinsic anomalous scaling, with different roughness exponents at short and large length scales (Marcos et al., 2024).

A further result is distributional. The fluctuation statistics are reported to be close to the Tracy–Widom probability distribution function that applies in the corresponding KPZ universality subclass, now the one expected for interfaces with an overall circular symmetry. This distinguishes the circular-droplet case from the earlier band geometry, where the relevant Tracy–Widom subclass was GOE rather than GUE. The authors present this as a geometry test of the same microscopic wetting model (Marcos et al., 2024).

In this usage, “Marcos” refers not to an acronym but to authorship within a research line on precursor-front roughening, diffusive or subdiffusive advance, intrinsic anomalous scaling, and geometry-dependent fluctuation subclasses.