MADPOT: Diverse Contexts and Applications

• MADPOT is a polysemous label used across disciplines, denoting distinct frameworks from differentiable experimental potentials to debate-based on-policy training.
• In molecular augmented dynamics, MADPOT represents an experimental potential that aligns simulated observables with experimental targets via a differentiable L2 loss, yielding realistic atomistic structures.
• In medical imaging and other domains, MADPOT encapsulates diverse methodologies—from CLIP adaptation with partial optimal transport to policy optimization—addressing challenges like data noise and multi-stage metric aggregation.

MADPOT is not a single standardized technical term. In recent arXiv literature it appears explicitly as “Medical Anomaly Detection with CLIP Adaptation and Partial Optimal Transport”, while in several other works it is used only by contextual interpretation to denote a key potential, model family, or optimization framework built around a method named MAD. The resulting usages span atomistic structure generation, universal interatomic potentials, on-policy LLM distillation, preference optimization, medical image anomaly detection, and validity auditing in agent-based policy optimization (Shiri et al., 9 Jul 2025, Zarrouk et al., 23 Aug 2025, Mazitov et al., 24 Jun 2025, Wang et al., 2 May 2026, Rho, 6 Oct 2025, Zhang et al., 27 Jun 2026). This suggests that MADPOT is best treated as a polysemous label whose meaning is determined entirely by disciplinary context.

1. Nomenclature and disambiguation

The available literature associates MADPOT with multiple non-equivalent constructs. In one paper the acronym is explicit; in others it is a contextual interpretation supplied by the paper summary.

Context	Meaning of MADPOT	Source
Medical imaging	Medical Anomaly Detection with CLIP Adaptation and Partial Optimal Transport	(Shiri et al., 9 Jul 2025)
Molecular simulation	The experimental potential $\tilde{V}$ in Molecular Augmented Dynamics	(Zarrouk et al., 23 Aug 2025)
Atomistic ML	Universal interatomic potentials trained on the Massive Atomic Diversity dataset	(Mazitov et al., 24 Jun 2025)
LLM distillation	Multi-Agent Debate-based On-Policy Training/Distillation	(Wang et al., 2 May 2026)
Preference optimization	Query-level reference to Margin-Adaptive DPO	(Rho, 6 Oct 2025)
ABM+MOEA validity	Metric Aggregation Divergence in Policy Optimization Pipelines	(Zhang et al., 27 Jun 2026)

A recurring source of confusion is that only the medical-imaging usage is an explicit acronym introduced by its paper title. In the molecular, dataset, distillation, preference-optimization, and ABM+MOEA settings, the term is absent from the original paper title or formulation and is instead interpreted from surrounding terminology. This matters because the underlying objects are structurally different: a differentiable penalty potential, a family of universal PES surrogates, a multi-teacher distillation loop, an instance-weighted DPO variant, and a pipeline-level validity threat are not variants of a common algorithmic core.

2. MADPOT in molecular augmented dynamics

In "Molecular augmented dynamics: Generating experimentally consistent atomistic structures by design" (Zarrouk et al., 23 Aug 2025), the contextual meaning of MADPOT is the experimental potential $\tilde{V}$ added to the molecular-dynamics Hamiltonian. The augmented Hamiltonian is

$\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$

where $T$ is kinetic energy, $V$ is the interatomic potential energy, and $\tilde{V}$ penalizes mismatch between simulated observables and experimental targets. The corresponding augmented objective is

$$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$

and the forces are

$$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$

The paper specializes $\tilde{V}$ to an element-wise weighted $L_2$ loss over vector-valued observables,

$\tilde{V}$0

with per-point weights often chosen proportional to inverse experimental uncertainty. The differentiability requirement is central: MAD accepts any observable whose simulated counterpart can be cast as a function of differentiable atomic descriptors. In the implementation, SOAP-type local environments, analytic kernels, and finite cutoffs ensure that gradients are well-defined and computationally efficient.

The demonstrated observables are X-ray diffraction, neutron diffraction, pair distribution function, and X-ray photoelectron spectroscopy. For diffraction, the simulated intensity is built from the Debye scattering equation and compared to experiment with a weighted $\tilde{V}$1 loss on $\tilde{V}$2. For PDF, the method uses a smoothed pair histogram

$\tilde{V}$3

with Gaussian broadening to regularize gradients. For XPS, the simulated spectrum is a kernel density estimate over ML-predicted per-atom core-electron binding energies,

$\tilde{V}$4

with $\tilde{V}$5 eV.

The workflow consists of melt-quench initialization, parameter selection for $\tilde{V}$6, $\tilde{V}$7, and weights, MAD annealing in NVT or NPT with augmented forces at every step, and a final relaxation under the MLP alone. The paper uses periodic systems of about $\tilde{V}$8 atoms for pure carbon examples and about $\tilde{V}$9 atoms for a-C:D, with a Bussi thermostat for NVT, $\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$0 Å in most cases and $\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$1 Å for nanoporous carbon. The underlying interatomic model is a GAP trained to ab initio data, and the implementation is in TurboGAP.

Empirically, MAD identifies experimentally consistent metastable structures for glassy carbon, nanoporous carbon, ta-C, a-C:D, and a-CO$\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$2 using the same initial structure family and the same underlying MLP. Reported outcomes include a glassy-carbon density of approximately $\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$3 g/cm$\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$4 versus approximately $\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$5 g/cm$\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$6 for the control, a ta-C $\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$7 fraction of approximately $\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$8 versus $\mathcal{H}(R,P) = T(P) + V(R) + \tilde{V}(R),$9 in the control and close to an experimental value of about $T$0, suppression of CO$T$1 formation in oxygen-rich amorphous carbon, and near-perfect reproduction of the low-$T$2 neutron-diffraction region for a-C:D. The method does not guarantee a global optimum, remains sensitive to initialization and weighting, and can overfit noisy spectra or induce unrealistic density changes if the experimental forces are overemphasized.

3. MADPOT as universal interatomic potentials trained on Massive Atomic Diversity

In "Massive Atomic Diversity: a compact universal dataset for atomistic machine learning" (Mazitov et al., 24 Jun 2025), MADPOT naturally denotes universal interatomic potentials trained on the Massive Atomic Diversity (MAD) dataset. Their intended scope is universal, charge-unaware prediction for arbitrary atomic geometries and compositions across organic and inorganic systems, including bulk crystals, 2D materials, surfaces, clusters, and molecules, while remaining robust under aggressive distortions and out-of-equilibrium configurations.

MAD contains 95,595 structures spanning 85 elements from $T$3 to $T$4, excluding At. The structural subsets are MC3D, MC3D-rattled, MC3D-random, MC3D-surface, MC3D-cluster, MC2D, SHIFTML-molcrys, and SHIFTML-molfrags. Its design philosophy differs from conventional discovery-oriented datasets: it maximizes atomic diversity via aggressive distortions with near-complete disregard for stability, prioritizes a single consistent electronic-structure protocol over per-material accuracy, covers both organic and inorganic systems, and remains compact enough for converged uniform DFT and accessible retraining.

The reference calculations use Quantum ESPRESSO v7.2 compiled with SIRIUS and managed via AiiDA, with PBEsol, SSSP v1.2 efficiency pseudopotentials, $T$5 Ry wavefunction cutoff, $T$6 Ry charge-density cutoff, Marzari-Vanderbilt-DeVita-Payne cold smearing with spread $T$7 Ry, and $T$8-centered $T$9-point grids at resolution $V$0 Å$V$1 along periodic dimensions. Spin polarization, dispersion corrections, and explicit charge labels are omitted for consistency. Convergence exceeds $V$2 for most subsets, while MC3D-random converges at about $V$3.

The paper does not provide complete model-training details for the resulting potentials, but it states that MAD has already enabled universal interatomic potentials competitive with models trained on traditional datasets containing two to three orders of magnitude more structures. A concrete example is PET-MAD, whose final-layer features define the paper’s latent cartography. The per-structure descriptor is

$V$4

where $V$5 is a 512-dimensional environment feature. Sketch-map is then used to project these descriptors into a low-dimensional latent space using the transformed loss

$V$6

Comparative mapping shows MAD covering a larger 3D latent spread than MPtrj and Alexandria and long descriptor-distance tails, especially for MC3D-random. The paper recommends an $V$7 train:validation:test split and provides benchmark packs, Chemiscope visualizations, and a PET-MAD featurizer. Its limitations are explicit: no spin, no dispersion, no strong-correlation treatment, under-representation of noble gases and lanthanides, and aggressive distortions that populate high-energy and potentially unphysical regions. A plausible implication is that MADPOT in this sense is optimized for broad structural robustness rather than for domains requiring subtle relative energetics without task-specific fine-tuning.

4. MADPOT in on-policy distillation and agentic training

In "MAD-OPD: Breaking the Ceiling in On-Policy Distillation via Multi-Agent Debate" (Wang et al., 2 May 2026), MADPOT is naturally interpreted as Multi-Agent Debate-based On-Policy Training/Distillation. The method addresses two linked problems: the single-teacher capability ceiling in OPD and the instability induced by compounding errors in multi-step agentic trajectories. MAD-OPD replaces a single teacher with a collective of $V$8 teachers that debate for $V$9 rounds over the student’s on-policy state, producing privileged debate context $\tilde{V}$0 that is visible to teachers but hidden from the student.

The OPD objective with privileged information is

$\tilde{V}$1

After the final debate round, each teacher outputs a confidence score $\tilde{V}$2, converted into weights

$\tilde{V}$3

The resulting MAD-OPD loss is

$\tilde{V}$4

OPAD extends this to agentic settings with step-level supervision on states $\tilde{V}$5. The paper’s main theoretical contribution is a task-adaptive divergence principle. For agentic tasks it selects JSD with $\tilde{V}$6, exploiting the bound

$\tilde{V}$7

which ensures stable gradients under privileged teacher-student gaps. For code generation it selects reverse KL,

$\tilde{V}$8

because its mode-seeking geometry favors a single coherent implementation path.

Training uses AdamW with $\tilde{V}$9, $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$0, learning rate $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$1, cosine decay, $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$2 warmup, weight decay $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$3, gradient clipping at norm $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$4, effective batch size $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$5, BF16 precision, context length $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$6 for agentic tasks and $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$7 for code, and full-vocabulary distillation logits. The debate temperature is $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$8, and agentic simulation caps at $$U_{\mathrm{aug}}(R) = E_{\mathrm{MLP}}(R) + \sum_k \lambda_k\,L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big),$$9 environment steps.

Across six teacher-student configurations involving Qwen3 and Qwen3.5 models and five benchmarks, MAD-OPD ranks first across all six configurations. In the $$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$0B$$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$1B$$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$2B setting, it improves the agentic average by $$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$3 points over single-teacher OPD ($$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$4 versus $$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$5) and the code average by $$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$6 points ($$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$7 versus $$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$8). In one code example, a $$F_i = -\nabla_{R_i}U_{\mathrm{aug}}(R) = F_i^{\mathrm{MLP}} - \sum_k \lambda_k \nabla_{R_i}L_k\!\big(O_k(R),O_k^{\mathrm{exp}}\big).$$9B student trained under $\tilde{V}$0B$\tilde{V}$1B debate surpasses the $\tilde{V}$2B teacher on LiveCodeBench v6, with pass@1 of $\tilde{V}$3 versus $\tilde{V}$4 and BoN@16 of $\tilde{V}$5 versus $\tilde{V}$6. The stated limitations are compute overhead scaling with $\tilde{V}$7, the requirement for teacher token-level distributions and a shared vocabulary, and sensitivity to teacher diversity and confidence calibration.

5. MADPOT in preference optimization

In "Margin Adaptive DPO: Leveraging Reward Model for Granular Control in Preference Optimization" (Rho, 6 Oct 2025), the method is named MADPO, but the supplied query treats it as a MADPOT-related usage. MADPO is an instance-level extension of DPO that learns a reward model and then reweights the DPO loss according to estimated preference margins. The underlying DPO pairwise margin is

$\tilde{V}$8

and the standard DPO loss is

$\tilde{V}$9

MADPO first fits a BTL reward model

$L_2$0

with loss

$L_2$1

It then defines a continuous margin-adaptive coefficient

$L_2$2

and a piecewise weight

$L_2$3

The final objective is

$L_2$4

The stated effect is to amplify low-margin, informative pairs and dampen high-margin, easy pairs, without filtering data or permitting negative $L_2$5. The analysis proves bounded gradient and Hessian,

$L_2$6

and a stability bound with respect to reward-estimation error under identifiability, smoothness, and bounded-policy-term assumptions.

The experiments use a controlled sentiment-generation task based on IMDB prompts, with google/gemma-3-270M fine-tuned by LoRA, an oracle reward from cardiffnlp/twitter-roberta-base-sentiment-latest mapped by $L_2$7, $L_2$8 preference pairs per quality tier, $L_2$9 training pairs, and $\tilde{V}$00 held-out examples. All methods use $\tilde{V}$01. Best mean oracle rewards are: DPO $\tilde{V}$02, IPO $\tilde{V}$03, $\tilde{V}$04-DPO $\tilde{V}$05, and MADPO $\tilde{V}$06 on High/Medium/Low Quality data. Relative to the next-best baseline, MADPO improves by $\tilde{V}$07, $\tilde{V}$08, and $\tilde{V}$09, respectively. Ablations show amplification-only is highest or near-highest, regularization-only improves over DPO, and full MADPO remains competitive with amplification-only while providing explicit regularization safeguards. The limitations are dependence on reward-model quality, potential bias amplification under aggressive weighting, domain-transfer risk, and the fact that the experiments were conducted on a $\tilde{V}$10M-parameter model with simulated preferences.

6. MADPOT in medical anomaly detection

The explicit acronym usage appears in "MADPOT: Medical Anomaly Detection with CLIP Adaptation and Partial Optimal Transport" (Shiri et al., 9 Jul 2025). Here MADPOT is a CLIP-based framework for anomaly classification and anomaly segmentation in medical imaging under few-shot, zero-shot, and cross-dataset conditions. The method is designed for modalities and organs that vary widely, with anomalies that are subtle, localized, and weakly labeled.

The vision backbone is CLIP ViT-L/14, frozen across $\tilde{V}$11 layers. A visual adapter is inserted at layer $\tilde{V}$12 and a projector at layer $\tilde{V}$13, each a small learnable linear stack, with residual blending

$\tilde{V}$14

A shared mapping and task-specific heads generate detection and segmentation features:

$\tilde{V}$15

$\tilde{V}$16

The text branch uses two classes, normal and abnormal, with $\tilde{V}$17 learnable prompts per class by default. For class $\tilde{V}$18, the $\tilde{V}$19-th prompt is

$\tilde{V}$20

and the fused class prototype is

$\tilde{V}$21

MADPOT aligns local image patches to prompts through entropic Partial Optimal Transport:

$\tilde{V}$22

subject to

$\tilde{V}$23

The cost matrices use cosine distance, with $\tilde{V}$24, maximum $\tilde{V}$25 iterations, early stopping at $\tilde{V}$26, and best transport ratio $\tilde{V}$27.

Detection combines a POT branch and a contrastive-similarity branch,

$\tilde{V}$28

$\tilde{V}$29

$\tilde{V}$30

Segmentation combines POT logits and fused-prompt similarities, interpolates them to input resolution by bicubic interpolation, and normalizes them by softmax. The training loss is

$\tilde{V}$31

with $\tilde{V}$32.

The reported experiments use the BMAD benchmark across brain MRI, liver CT, chest X-ray, retinal OCT, and histopathology. In few-shot evaluation with $\tilde{V}$33 normal and $\tilde{V}$34 abnormal samples per class, MADPOT averages $\tilde{V}$35 AUC for anomaly classification and $\tilde{V}$36 for anomaly segmentation. Dataset highlights include HIS AC $\tilde{V}$37, Chest AC $\tilde{V}$38, OCT17 AC $\tilde{V}$39, Brain AC/AS $\tilde{V}$40, Liver AC/AS $\tilde{V}$41, and RESC AC/AS $\tilde{V}$42. In zero-shot evaluation, AC averages $\tilde{V}$43, surpassing MVFA by $\tilde{V}$44, while AS averages $\tilde{V}$45 and trails the best average by $\tilde{V}$46, largely because of RESC. In cross-dataset transfer, MADPOT improves over MVFA by $\tilde{V}$47 AUC on average.

Ablation results indicate that CL+POT is the strongest and most balanced prompt-learning strategy, OT or POT alone underperform severely, and the adapter-projector combination outperforms either component alone by $\tilde{V}$48 AC and $\tilde{V}$49 AS over adapter-only. The stated limitations are residual CLIP domain gap, sensitivity to transport ratio and temperature, weaker zero-shot AS on some datasets such as RESC, and the need to validate the number of prompts and context length.

7. MADPOT as metric aggregation divergence in policy optimization

In "Metric Aggregation Divergence: A Hidden Validity Threat in Agent-Based Policy Optimization and a Contractual Remedy" (Zhang et al., 27 Jun 2026), MADPOT denotes Metric Aggregation Divergence in Policy Optimization Pipelines. The object of study is not a model but a structural validity threat in ABM+MOEA workflows when the same substantive metric is independently re-implemented across pipeline stages.

Let $\tilde{V}$50 be a simulation trajectory, with replications $\tilde{V}$51. A metric path is

$\tilde{V}$52

where $\tilde{V}$53 extracts a trajectory-level construct and $\tilde{V}$54, $\tilde{V}$55 perform time and replication aggregation. The paper formalizes binary or norm-based cross-stage divergence,

$\tilde{V}$56

value divergence

$\tilde{V}$57

and rank divergence $\tilde{V}$58, where $\tilde{V}$59 is Spearman correlation over a policy set $\tilde{V}$60.

The motivating pipeline has three stages: optimizer, tournament/evaluator, and inference. In the EpidemiOptim architecture, Stage 1 aggregates stochastic rollouts by means per objective for NSGA-II fitness; Stage 2 re-evaluates candidates with deterministic runs under eval=True; Stage 3 selects a champion using an equal-weight normalized cost sum

$\tilde{V}$61

with champion $\tilde{V}$62. Because these stages implement distinct metric paths, champion disagreement can occur even when each stage is internally coherent.

The paper’s central empirical finding is that in a faithful replication of the EpidemiOptim architecture, the Stage-1 versus Stage-3 champion disagreement rate is $\tilde{V}$63 across $\tilde{V}$64 independent NSGA-II runs, with $\tilde{V}$65 CI $\tilde{V}$66. In a $\tilde{V}$67-seed policy-flip experiment, divergent aggregation causes the optimizer to recommend the wrong champion in $\tilde{V}$68 of replications, with mean welfare gap $\tilde{V}$69 units and Gini inequality gap $\tilde{V}$70 units. In a follow-up inference audit, $\tilde{V}$71 of $\tilde{V}$72 flipped seeds cross the significance boundary itself. A complementary enterprise follow-up yields the predicted null under near-commensurable rankings, with Spearman $\tilde{V}$73 and zero champion flips across $\tilde{V}$74 seeds. In a public Lake Problem DPS rerun, the archived published-path recommendation reaches joint-threshold success $\tilde{V}$75, while a shared contract-path rule reaches $\tilde{V}$76.

The proposed remedy is the metric contract: a single shared callable injected and enforced across all stages,

$\tilde{V}$77

with all weights and scalings supplied by the contract rather than re-implemented locally. Enforcement relies on dependency injection, coincidence checks, an immutable interface, and unit or integration tests asserting equality of stage outputs. The runtime overhead is reported as approximately $\tilde{V}$78. The paper emphasizes that the contract guarantees structural identity, not semantic correctness of the chosen metric. Its broader claim is that MADPOT functions as an architectural analogue of researcher degrees of freedom: an inconsistency introduced by modular pipeline design rather than by explicit analytical choice.

The cross-domain record therefore does not support a single canonical meaning of MADPOT. Instead, the term designates a family of context-specific constructs whose only commonality is association with a parent method or problem labeled MAD. In atomistic simulation it is a differentiable experimental potential; in universal PES learning it denotes models trained on a deliberately distorted dataset; in LLM training it names a debate-conditioned on-policy supervision loop or, by query-level extension, an instance-weighted DPO variant; in medical imaging it is a concrete CLIP-POT architecture; and in ABM+MOEA evaluation it names a structural validity threat and its contractual remedy.