MACI: Diverse Research Frameworks

Updated 14 July 2026

MACI is a polysemous term encompassing distinct research artifacts such as network experiment frameworks, multilingual abusive comment datasets, blockchain voting protocols, fuzzy rule interpolation methods, and LLM coordination architectures.
Its network experiment framework replaces ad hoc scripts with an integrated workflow that separates configuration from environment parameters and supports scalable, parallel execution using tools like AWS EC2 and Proxmox.
MACI implementations in abusive content detection and blockchain governance demonstrate measurable improvements, including high accuracy in sentiment analysis and enhanced anti-collusion features via encrypted, privacy-preserving voting.

MACI is a polysemous research term used for several unrelated artifacts, methods, and frameworks across networking, abusive-content detection, blockchain governance, fuzzy systems, multimodal learning, conformal reliability, and LLM coordination. Across the cited literature, it denotes a network-experiment framework, a multilingual abusive-comment dataset, a privacy-preserving on-chain voting primitive, a fuzzy rule interpolation method, several causal-intervention modules, and multiple “Multi-Agent Collaborative Intelligence” architectures; in mathematical analysis, closely related search results also arise from the surname Macià, which is not an acronym (Frömmgen et al., 2018).

1. Polysemy and nomenclature

The term appears in distinct technical lineages rather than as a single canonical concept. In the literature considered here, the most explicit expansions are IIIT-D Multilingual Abusive Comment Identification, Minimal Anti-Collusion Infrastructure, Modified a-Cut based Interpolation, Model-Agnostic Causal Inference, Modality-conflict-Aware Causal Intervention, Multi-LLM Adaptive Conformal Inference, and Multi-Agent Collaborative Intelligence (Rehman et al., 2024).

Usage of MACI	Domain	Representative source
IIIT-D Multilingual Abusive Comment Identification	Multilingual abusive content detection	(Rehman et al., 2024)
Minimal Anti-Collusion Infrastructure	Ethereum voting / dispute resolution	(Gangemi et al., 2023)
Modified a-Cut based Interpolation	Fuzzy rule interpolation	(Alzubi et al., 2019)
Model-Agnostic Causal Inference	Multimodal emotion recognition	(Zhong et al., 12 Jun 2025)
Modality-conflict-Aware Causal Intervention	Multimodal LLM hallucination mitigation	(Jiang et al., 19 May 2026)
Multi-LLM Adaptive Conformal Inference	Reliable LLM response filtering	(Noh et al., 1 Feb 2026)
Multi-Agent Collaborative Intelligence	LLM planning and debate control	(Chang, 28 Jan 2025)

This distribution has a practical consequence for literature search: acronym-only retrieval is ambiguous, and domain context is essential. A plausible implication is that citations should usually expand the acronym on first use, because the same token names datasets, algorithms, cryptographic infrastructure, and coordination architectures.

2. MACI as a framework for network experiments

In "Don't Repeat Yourself: Seamless Execution and Analysis of Extensive Network Experiments," MACI is the first bespoke framework for the management, the scalable execution, and the interactive analysis of a large number of network experiments (Frömmgen et al., 2018). Its motivation is methodological: ad hoc script collections for running experiments and plotting results tend to evolve into substantial, project-specific infrastructure, and MACI was designed to replace that repeated reinvention with an integrated workflow.

The framework has three architectural pillars. The experiment management layer distinguishes an experiment study template, an experiment study, and an experiment as the atomic execution unit. A central design choice is the explicit separation between configuration parameters and environment parameters, because later analysis often asks how an implementation behaves as environment changes, or conversely how environment-independent configuration choices behave under controls. The execution layer treats experiments as independent atomic jobs and therefore as embarrassingly parallel workloads; the implementation supports worker infrastructures including AWS EC2 and Proxmox under an IaaS model. The analysis layer integrates the SciPy toolchain—notably Jupyter, NumPy, and pandas—and adopts an OLAP (hyper) cube view for filtering, aggregation, drill-down, and visualization, including box plots and automatically generated Pareto frontiers (Frömmgen et al., 2018).

The implementation stack is unusually concrete for an experience paper. The backend is a .NET Core server application exposing a REST API and a ready-to-use Java interface; experiment control uses Python scripts; the frontend is web-based; optional docker-compose can deploy the backend, Jupyter/SciPy, and a Mininet worker with a single command on major operating systems. The paper reports successful integration with Mininet, ns-3, and a custom Java-based simulator. It further claims reproducibility through automatic tracking of meta-information such as version number and commit identifiers, and gives two practical indicators of efficiency: reproducing a prior MPTCP study required only six lines of code beyond the single-experiment setup, and a full sequential DASH study would have exceeded 40 hours, making parallel execution important in practice (Frömmgen et al., 2018).

The paper is explicit about limits. It does not provide a formal scheduling algorithm, a formal data schema, controlled user studies, detailed scaling curves, or rigorous head-to-head comparison against existing experiment-management systems. Its statistical support is exploratory rather than inferential. MACI is therefore best understood not as a new simulator or analytics engine, but as workflow glue for extensive, repetitive, multidimensional communication-systems experimentation.

3. MACI as the IIIT-D Multilingual Abusive Comment Identification dataset

In "User-Aware Multilingual Abusive Content Detection in Social Media," MACI stands for the IIIT-D Multilingual Abusive Comment Identification dataset, originating from the IEEE BigMM challenge and collected from the Moj social media application (Rehman et al., 2024). It supports binary abusive content detection over short social-media comments, with labels Abusive (1) and Non-abusive (0), and is positioned as a benchmark for low-resource Indic languages with code-mixing and transliteration effects.

The dataset contains comments in 10 Indic languages: Hindi, Marathi, Tamil, Kannada, Gujarati, Telugu, Bengali, Odia, Malayalam, and Assamese; the paper explicitly notes that English and Punjabi are not present in MACI. Its total size is 665K comments, with 352,376 non-abusive comments (52.98%) and 312,656 abusive comments (47.02%), so it is described as fairly balanced. The language distribution is highly skewed, however: Hindi alone contributes 307,179 comments, whereas Assamese contributes 2,780. This imbalance is methodologically important when interpreting per-language scores (Rehman et al., 2024).

The paper’s preprocessing pipeline materially shapes MACI experiments. It removes rows with missing comments or missing feature values, extra spaces, punctuation, digits, erroneous entries, and “insignificant words” using a stopword list containing 2,250 words across Indic languages, with English stopwords from NLTK. Comments are transliterated into Roman script with IndicXlit, expressed as $t_c = IndicXlit(c)$ ; MuRIL and mBERT use transliterated comments, whereas XLM-R uses original-script comments because that performed better. The pipeline also converts emojis and emoticons to text, lowercases uppercase characters, and augments the training set with abusive spelling variants from an abusive lexicon containing more than 17K abusive words, including spelling variants. The paper states that augmentation is performed only on the training portion, but it does not provide exact train/validation/test ratios or split sizes for MACI (Rehman et al., 2024).

The proposed detection system has three modules: Text Feature Learning (TFL), Social Feature Learning (SFL), and Joint Feature Learning (JFL). On MACI, the crucial dataset-specific constraint is the absence of user IDs, so user-history features are not used. Instead, the social branch uses post-level features and especially post polarity, defined over the abusive vs non-abusive comment distribution on a post. The paper reports that on MACI the point-biserial correlation magnitude of post polarity is $-0.857$ , larger than that of contextual embeddings (0.752), and identifies this as the strongest MACI-specific signal. The full model uses an ensemble of six configurations—MuRIL, mBERT, and XLM-R, each at sequence lengths 64 and 128—with majority voting, confidence-based tie resolution, and fallback to the best individual model, which the paper states is MuRIL with sequence length 128 in general (Rehman et al., 2024).

Empirically, the proposed method reaches Accuracy 91.97, Precision 90.12, Recall 93.08, and F1 91.58 on MACI, outperforming all reported baselines. The paper states an average F1 gain of 9.52% on MACI over state-of-the-art baselines, and the best single additional feature in ablation is post polarity, which with text embeddings yields F1 91.01. The paper also reports strong per-language gains, with the proposed method exceeding transformer baselines in every language and especially large improvements in Malayalam, Marathi, Assamese, Kannada, and Gujarati (Rehman et al., 2024).

Several caveats are integral to the dataset’s interpretation. The paper does not numerically document split ratios, does not discuss split hygiene at the post level, and does not deeply analyze possible information leakage from estimating post polarity using comment distributions on the same post. It also notes that the method benefits from prior user or post history and may be less effective without it. These are not peripheral concerns: for MACI specifically, missing user IDs, post-level dependence, and language imbalance are part of the dataset’s methodological identity.

4. MACI in blockchain voting and fuzzy rule interpolation

In Ethereum-oriented cryptographic governance, MACI denotes Minimal Anti-Collusion Infrastructure. In "Towards a Privacy-Preserving Dispute Resolution Protocol on Ethereum," it is one of two core zero-knowledge layers, combined with Semaphore in the architecture Proof of Humanity → Semaphore → MACI (Gangemi et al., 2023). The paper defines MACI as a protocol that allows users to vote on-chain with a greatly increased collusion resistance and explains its function through zk-SNARKs, encrypted vote submission, message overwriting, and public key switching. Within the proposed dispute-resolution system, MACI is used primarily in Phase 0 and Phase 1: judges register their public keys in the MACI smart contract, receive one voice credit, and then privately vote for a disputing party and submit a proposal; the MACI coordinator computes the tally and commits only aggregate scores such as $V_A$ and $V_B$ , not individual ballots. Phase 2 then gives the disputing users voice credits equal to those aggregate scores for a later quadratic voting round over proposals (Gangemi et al., 2023).

The paper is careful about division of labor. Semaphore provides anonymous group membership and is described as a Sybil-protection mechanism; MACI provides encrypted voting, coordinator-based tallying, and anti-collusion or anti-bribery properties. The main caveat is equally explicit: MACI relies on a single trusted coordinator, and the paper calls this its “major flaw.” A compromised coordinator invalidates the intended privacy and anti-collusion guarantees. The design is therefore not trustless; it is a coordinator-dependent privacy layer embedded in a broader governance protocol.

In fuzzy systems, MACI denotes Modified a-Cut based Interpolation. In "Fuzzy Rule Interpolation Methods and Fri Toolbox," it is introduced as a single-step fuzzy rule interpolation method proposed by Tikk and Baranyi (Alzubi et al., 2019). The paper states that its main idea is a vector's description of the fuzzy sets for eliminating the abnormality problem in the conclusion. Fuzzy sets are described in two vectors space, representing the left and right flank of the $a$ -cut levels; the method applies a linear transform $b' = bT$ , with formulas such as $b'_0 = b_0 \cdot \sqrt{2}$ and $b'_1 = b_0 - b_1$ , computes the interpolated consequent reference point, and then maps back with $T^{-1}$ (Alzubi et al., 2019).

The paper attributes two principal strengths to this MACI: the conclusion is always a convex and normal fuzzy set, and the method can handle multi-dimensional antecedents. It also notes a possible disadvantage: in some instances MACI does not keep the piecewise linearity of the membership functions. Yet in the paper’s seven unified numerical benchmark examples, MACI is reported as suitable across the tested single-dimensional and multi-dimensional settings, and the conclusions section states that MACI, IMUL, CRF, GM and SCALE MOVE methods did not suffer from abnormality and piecewise linearity according to the unified numerical examples (Alzubi et al., 2019). This produces a nuanced characterization: generally, piecewise-linearity preservation may be a limitation; in the benchmarked cases of that paper, the limitation did not manifest.

5. MACI as causal intervention and reliability machinery in machine learning

A distinct contemporary usage is Model-Agnostic Causal Inference, introduced inside the CIDer framework for robust multimodal emotion recognition. In "Towards Robust Multimodal Emotion Recognition under Missing Modalities and Distribution Shifts," MACI addresses out-of-distribution degradation, while a separate module, MSSD, handles missing modalities (Zhong et al., 12 Jun 2025). The causal graph uses nodes $L$ , $-0.857$ 0, $-0.857$ 1, $-0.857$ 2, $-0.857$ 3, and $-0.857$ 4, and MACI targets both label bias during training and language bias during testing. Training uses an interventional formulation,

$-0.857$ 5

implemented by the Multimodal Causal Module (MCM) through class-level modality prototypes and class-specific classifiers. Testing adds fine-grained counterfactual texts, retaining only strongly class-skewed frequent words using the coefficient of variation threshold $-0.857$ 6 and a top 100 most frequent words constraint, and debiases prediction by subtracting a counterfactual term weighted by $-0.857$ 7. The paper reports that MACI is lightweight—37,405 parameters—and portable enough to improve external backbones such as DLF+MACI and MPLMM+MACI (Zhong et al., 12 Jun 2025). Its caveats include a tailored causal graph, a test-time equal-class assumption, heuristic counterfactual construction, and an unspecified numerical value for $-0.857$ 8.

In multimodal LLMs, MACI denotes Modality-conflict-Aware Causal Intervention. "Causal Evidence for Attention Head Imbalance in Modality Conflict Hallucination" identifies hallucination-driving heads and hallucination-resisting heads by head-level path patching on paired clean and conflict inputs (Jiang et al., 19 May 2026). The paper’s central mechanistic finding is an imbalance: the summed positive importance exceeds the summed absolute negative importance by approximately 1.51× on average across five MLLMs; the top-5 resisting heads account for 27.2% of resisting importance, whereas the top-5 driving heads account for only 14.0% of driving importance. MACI operationalizes this by training a conflict detector from averaged resisting-head activations, using a Lasso logistic regression probe with object-conflict AUROC 0.89–0.95, and conditionally zeroing the outputs of selected driving heads only when conflict is detected. Across Qwen2.5-VL-7B-Instruct, Qwen3-VL-8B-Instruct, InternVL3-8B, LLaVA-NeXT-7B, and LLaVA-7B, MACI achieves the largest hallucination reduction among compared inference-time baselines on MMMC object conflict, and it transfers zero-shot to SCI-SemanticConflict with mean hallucination reduction of -7.9 pp (Jiang et al., 19 May 2026). The method is explicitly easier to deploy on open-source models with access to internal head activations than on API-only systems.

A third machine-learning usage is Multi-LLM Adaptive Conformal Inference for post hoc reliability filtering of LLM outputs. In "Multi-LLM Adaptive Conformal Inference for Reliable LLM Responses," MACI decomposes a response into claims, estimates claim-level factuality with an ensemble of verifier LLMs, aggregates those scores multiplicatively, and then calibrates a filtering threshold conformally so that, with probability at least $-0.857$ 9, all retained claims are factual (Noh et al., 1 Feb 2026). The method models each example as $V_A$ 0, defines group-conditional validity through a grouping function $V_A$ 1, and calibrates separate quantiles for each group. The ensemble in the main experiments uses Llama-3.3-70B-Instruct, Qwen-2.5-72B-Instruct, and DeepSeek-V3. Empirically, at target coverage 90%, retention on MedLFQA is 0.50 for MACI versus 0.02 for BCI and 0.31 for CCI; on WikiBio, 0.25 versus 0.01 and 0.11; and on ExpertQA, the paper emphasizes that MACI is the method that reliably aligns to the requested target coverage in the aggregate table. On 500 WikiBio test samples, reported wall-clock time is 598.98 s for MACI versus 1643.91 s for CCI (Noh et al., 1 Feb 2026). The theoretical guarantee remains group-conditional rather than fully conditional, and it inherits the usual exchangeability caveat of conformal inference.

6. MACI as Multi-Agent Collaborative Intelligence

A separate lineage uses MACI for Multi-Agent Collaborative Intelligence in LLM reasoning and planning. In "MACI: Multi-Agent Collaborative Intelligence for Adaptive Reasoning and Temporal Planning," MACI is a framework with three key components: a meta-planner (MP), a collection of agents, and a run-time monitor (Chang, 28 Jan 2025). The meta-planner is defined as

$V_A$ 2

mapping objectives and explicit constraints into a workflow graph $V_A$ 3. The framework separates planning from validation, keeps agent context minimal, and augments explicit constraints with common-sense and derived constraints through

$V_A$ 4

Its two empirical case studies are the Traveling Salesperson Problem and Thanksgiving Dinner Planning. For $V_A$ 5 TSP, the paper reports that MP + Claude, MP + GPT4o, and MP + DeepSeek all reached the stated 60 minutes goalpost. In Thanksgiving planning, all three LLMs paired with the MACI workflow produced feasible sequential plans, while under a 3-hour delay to James’s flight, DeepSeek and Claude found valid reactive plans and GPT4o failed (Chang, 28 Jan 2025). The paper is explicit that the framework remains partially manual: “the feedback loop for refining $V_A$ 6 must be performed manually.”

In "Multi-Agent Collaborative Intelligence: Dual-Dial Control for Reliable LLM Reasoning," the same acronym names a more explicit debate-control architecture (Chang et al., 6 Oct 2025). This MACI introduces two independent dials: an information dial that gates evidence by quality and a behavior dial that schedules contentiousness from exploration to consolidation. A moderator tracks disagreement $V_A$ 7, overlap $V_A$ 8, evidence quality $V_A$ 9, and argument quality from a cross-family judge, CRIT; stopping is plateau-based rather than fixed-round. The paper gives theory-lite guarantees for nonincreasing dispersion and provable termination, and a budget-feasible scheduler with regret bound $V_B$ 0. On clinical diagnosis, full MACI reaches Acc@1 77.8, MRR 0.856, ECE 0.081, Brier 0.158, with 18.4 thousand tokens and 2.7 rounds on average, outperforming the best single model and a fixed-contentiousness debate baseline. On news-bias detection, MACI reaches MAD 0.12, ECE 0.093, Brier 0.167, with 8.2 thousand tokens and 2.4 rounds, outperforming unscheduled debate and simple partisan averaging (Chang et al., 6 Oct 2025). The paper also validates CRIT for order invariance and judge-swap stability, while cautioning that stability depends on using high-capability judges.

A later conceptual extension, "The Missing Layer of AGI: From Pattern Alchemy to Coordination Physics," again uses MACI for Multi-Agent Collaborative Intelligence, but here as an explicit coordination stack layered atop pretrained LLMs (Chang, 5 Dec 2025). The paper links MACI to UCCT through the anchoring score

$V_B$ 1

and frames reasoning as a phase transition whose engagement probability is

$V_B$ 2

Its architectural triad is baiting (behavior-modulated debate), filtering (CRIT-based Socratic judging), and persistence (transactional memory). Unlike the two 2025 papers, however, this work does not present a dedicated MACI experimental section with benchmark tables or ablations; it is primarily a conceptual architecture paper (Chang, 5 Dec 2025). A plausible synthesis across these three MACI papers is that “Multi-Agent Collaborative Intelligence” evolved from workflow-centric planning, to measurable dual-dial debate control, to a broader coordination thesis for LLM-based reasoning.

7. Macià as surname rather than acronym

Searches for “MACI” in mathematical analysis can also retrieve work by Fabricio Macià, where the term is not an acronym. "Uniform estimates for the solutions of the Schrödinger equation on the torus and regularity of semiclassical measures" is a paper by Tayeb Aïssiou, Dmitry Jakobson, and Fabricio Macià that explicitly builds on earlier work of Nalini Anantharaman and Fabricio Macià (Aïssiou et al., 2011). It studies the Schrödinger flow on the arithmetic flat torus and proves the uniform estimate

$V_B$ 3

together with improved regularity of weak-* limits of $V_B$ 4 under oscillation assumptions, obtaining

$V_B$ 5

and, under stronger nonresonance conditions, $V_B$ 6 regularity (Aïssiou et al., 2011).

This literature is unrelated to the acronymic uses above, but it matters bibliographically. For researchers working across arXiv-scale corpora, “MACI” therefore names both a cluster of acronym-based systems and a distinct surname-based line in semiclassical analysis. The ambiguity is lexical rather than conceptual, yet it materially affects indexing, retrieval, and citation disambiguation.