CRACI: A Polysemous Research Label

• CRACI is a polysemous research label that serves as a domain-dependent index used for various research programs, highlighting diverse methodologies across fields.
• In clinical speech assessment, CRACI refers to a cross-lingual retrieval-augmented approach that improves dysarthria severity classification through an align-retrieve-fuse pipeline.
• In industrial systems, CRACI defines a cloud-native reference architecture leveraging microservices, semantic enrichment, and distributed orchestration to optimize the compute continuum.

Searching arXiv for papers using the term “CRACI” and closely related canonical forms. CRACI does not denote a single standardized construct across recent arXiv literature. Instead, it functions as a domain-dependent label that is either used canonically, as in the industrial-systems paper titled "CRACI: A Cloud-Native Reference Architecture for the Industrial Compute Continuum" (Dinh-Tuan, 9 Sep 2025), or mapped onto neighboring canonical terms such as CRAC, CRA, or a learned conflict-resolution agent in other fields. The resulting landscape spans clinical speech assessment, multilingual coreference resolution, conformal uncertainty quantification for segmentation, industrial cloud-native architectures, cellular IoT random access, cognitive radio information theory, and O-RAN control conflict mitigation. This plurality makes disambiguation a prerequisite for technical discussion.

1. Nomenclature and disambiguation

Across the cited works, CRACI is best treated as a polysemous research label rather than a single acronym with a stable expansion. In some papers it is the exact canonical name; in others it is explicitly described as a query-side or interpretive mapping to a different term. A common misconception is therefore that CRACI always refers to one method family. The literature indicates the opposite: the same string indexes several unrelated research programs with distinct objects, datasets, metrics, and mathematical machinery (Jeong et al., 22 Jun 2026, Bourgois et al., 16 May 2026, Luo et al., 10 Apr 2025, Dinh-Tuan, 9 Sep 2025).

Usage in literature	Canonical designation in source	Domain
CRACI	CRAC	Cross-lingual dysarthria severity assessment (Jeong et al., 22 Jun 2026)
CRACI	CRAC 2026	Multilingual coreference resolution shared task (Bourgois et al., 16 May 2026)
CRACI	CRA / CCRA / CCRA-S	Conformal risk adaptation for segmentation (Luo et al., 10 Apr 2025)
CRACI	CRACI	Industrial compute continuum architecture (Dinh-Tuan, 9 Sep 2025)
CRACI	Conflict Resolution agent concept, implemented as ACCoRD	O-RAN xApp conflict mitigation (Adamczyk et al., 21 May 2026)
CRACI	CRA with NOMA layering	Cellular IoT random access (Choi, 2018)
CRACI	Interpretive label for a unified CRCC inner bound	Cognitive radio information theory (Monemizadeh et al., 2011)

Only one of these sources uses CRACI as the paper-title acronym itself: the industrial architecture work (Dinh-Tuan, 9 Sep 2025). In the dysarthria paper, CRACI is the cross-lingual instantiation of CRAC; in the coreference paper, it corresponds to CRAC as the workshop/shared-task acronym; in the conformal segmentation paper, it is a convenient label for the intent of CRA/CCRA/CCRA-S rather than the authors’ terminology; and in the O-RAN, CRA, and CRCC summaries it is a functional or interpretive expansion rather than the canonical term in the original title.

2. Cross-lingual retrieval-augmented dysarthria severity assessment

In clinical speech processing, CRACI denotes the cross-lingual instantiation of CRAC, "Cross-lingual Retrieval-Augmented Classification," for dysarthria severity assessment. The method addresses the scarcity of labeled pathological speech by using an align-retrieve-fuse pipeline that leverages speech from a different language rather than naively pooling multilingual data. A frozen Whisper-small encoder produces frame-level hidden states, mean pooling yields a content feature $e \in \mathbb{R}^{768}$, and a two-layer MLP with ReLU maps $768 \rightarrow 384 \rightarrow 128$ to an $L_2$-normalized search vector $z \in \mathbb{R}^{128}$. Supervised contrastive learning, with mixed Korean-Italian and task-mixed mini-batches and temperature $\tau = 0.15$, constructs a severity-focused embedding space in which samples sharing a severity label are treated as positives regardless of language or task. Retrieval is then performed with FAISS over an opposite-language database using cosine similarity on normalized $z$ vectors, with top-$k = 5$ references, and the retrieved content features are fused with the query via an 8-head cross-attention block followed by an MLP classifier with hidden sizes $[512,256]$ and a 3-way softmax (Jeong et al., 22 Jun 2026).

The evaluation protocol is speaker-independent and three-class, with severity labels Healthy Control, Mild-to-Moderate, and Severe. Korean data comprise post-stroke dysarthria in a clinical rehabilitation setting; Italian data comprise ALS dysarthria from a public dataset. Each subject contributes six recordings: MPT vowels /a/, /i/, /u/ and DDK syllables /pa/, /ta/, /ka/. At inference, recording-level softmax outputs are averaged across the six tasks and the subject-level class is obtained by argmax. The reported primary metric is balanced accuracy,

$$\mathrm{BA}=\frac{1}{C}\sum_{c=1}^{C}\frac{\mathrm{TP}_c}{\mathrm{TP}_c+\mathrm{FN}_c}.$$

The reported subject-level results are 87.3% balanced accuracy on Korean and 86.7% on Italian. These correspond to gains of 8.4 percentage points over the monolingual baseline on Korean and 20.0 percentage points over the monolingual baseline on Italian; relative to pooled multilingual training, gains are 10.9 and 6.7 percentage points, respectively. Ablations show that alignment alone gives mixed results, while retrieval without alignment fails, indicating that severity-focused representation learning and retrieval are complementary. Qualitative t-SNE analysis further places the clearest class separation in the fused representation $f=[e_q;c]$, suggesting that retrieved opposite-language references contribute information not present in the raw Whisper content feature alone.

3. CRAC 2026 and multilingual coreference resolution

In computational linguistics, CRACI corresponds not to a model family but to CRAC, the Workshop on Computational Models of Reference, Anaphora and Coreference. The relevant system paper describes a first-place LLM-track submission to the CRAC 2026 shared task for multilingual coreference resolution on a unified benchmark derived from CorefUD. The task requires systems to identify mentions, including zero anaphora when present in a dataset, and cluster them by referent using head-based mention matching with singletons excluded from scoring. Final ranking is by the macro-average CoNLL F1 over the test datasets, where CoNLL F1 is the average of MUC, $768 \rightarrow 384 \rightarrow 128$0, and $768 \rightarrow 384 \rightarrow 128$1 (Bourgois et al., 16 May 2026).

The system uses Gemma-3-27B-it with QLoRA, with LoRA rank $768 \rightarrow 384 \rightarrow 128$2, scaling $768 \rightarrow 384 \rightarrow 128$3, dropout $768 \rightarrow 384 \rightarrow 128$4, and adapters applied to all linear layers. Fine-tuning follows a two-stage adaptation strategy. First, a multilingual base adapter is trained for one epoch on the concatenation of all datasets. Second, dataset-specific adapters continue from the base adapter for 1–5 additional epochs, with the final submission using up to 3. This specialization is motivated by annotation-guideline heterogeneity and substantially improves performance on low-resource and historical datasets; on the 1B proxy model, dataset-specific adaptation improves average CoNLL F1 by 8.15 points over the multilingual baseline.

A central design decision is headword-only representation with minimal XML-inspired tags. Mention heads are marked by a single tag after the headword, zero anaphora are tagged after the verbal head, and local reindexing maps visible cluster IDs to a contiguous local range before projecting them back to document-global IDs. Documents are processed iteratively in sentence batches with previously annotated context fed forward; the final 27B system uses 6 sentences per pass, 1,024 training-context tokens, and 3,072 inference-context tokens. The paper states that 3,072 tokens cover 99.84% of measured last-antecedent distances. The official test-set result is 74.32 average CoNLL F1, ranking first in the LLM track and third overall. The same paper emphasizes that retrieval augmentation is not used in this setting; chunking, long context, local reindexing, and dataset-specific adapters suffice for long-document robustness.

4. Conditional conformal risk adaptation and segmentation uncertainty

In uncertainty quantification for image segmentation, the term CRACI is not used by the paper itself, but it closely matches the intent of CRA, CCRA, and CCRA-S: methods that retain split-conformal marginal validity while improving image-wise, or "conditional," risk behavior. The primary loss is false negative rate,

$768 \rightarrow 384 \rightarrow 128$5

and the objective is to control $768 \rightarrow 384 \rightarrow 128$6 while reducing the variability of per-image loss across heterogeneous images. CRA constructs adaptive segmentation sets by selecting the smallest set of pixels whose predicted probability mass reaches a specified fraction of total mass. The paper then establishes a weighted-quantile framework linking conformal risk control and conformal prediction for arbitrary monotone score functions, yielding an exact split-conformal threshold without grid search. For FNR with $768 \rightarrow 384 \rightarrow 128$7, the threshold is defined by

$768 \rightarrow 384 \rightarrow 128$8

CCRA replaces raw probabilities with isotonic-regression-calibrated probabilities, and CCRA-S further stratifies images by an image-level confidence proxy, using total calibrated probability mass to obtain stratum-specific thresholds (Luo et al., 10 Apr 2025).

The empirical study uses polyp segmentation with PraNet on ETIS, CVC-ClinicDB, CVC-ColonDB, EndoScene, and Kvasir, trained on 1000 images with 798 held out for evaluation. The calibration/test split is 70%/30%, repeated over 100 trials. At $768 \rightarrow 384 \rightarrow 128$9, the reported mean coverage gaps are 0.157 for CRC, 0.143 for CRA, 0.113 for CCRA, and 0.101 for CCRA-S; corresponding marginal coverages are 0.900, 0.901, 0.901, and 0.908. At $L_2$0, coverage gap decreases from 0.264 for CRC to 0.138 for CCRA-S. The qualitative interpretation supplied by the paper is that adaptive cumulative-mass sets, calibrated probabilities, and stratified thresholds progressively reduce conditional variability while preserving marginal guarantees. Exact per-image conditional guarantees remain impossible in the distribution-free setting, and the paper explicitly frames its contribution as improved conditional stability rather than exact conditional validity.

5. CRACI as a cloud-native reference architecture for the industrial compute continuum

The exact title acronym CRACI is defined as a Cloud-Native Reference Architecture for the Industrial Compute Continuum. It is proposed as an alternative to rigid hierarchical models such as ISA-95 and RAMI 4.0, which are described as prone to data silos, top-down integration patterns, and limited support for cloud-native deployment and governance. CRACI instead adopts a decoupled, event-driven, data-centric architecture organized over the industrial compute continuum: shop-floor edge resources, regional or on-premises cores, and public or private cloud. Its core architectural elements are a Digital Twin Hub within Asset Representation Services, a pub/sub communication backbone, semantic enrichment via Semantic Context Services, and distributed Orchestration Services, all overlaid by four foundational pillars: Trust, Governance & Policy, Observability, and Lifecycle Management (Dinh-Tuan, 9 Sep 2025).

The deployment axis includes Physical Assets, Adaptation Services, Edge Platform Services, and Operational Intelligence on the shop floor; Semantic Context Services, Asset Representation Services, and Orchestration Services in the regional core; and Strategic Intelligence in the cloud, plus Unified API Management Gateway, User Interaction Services, and a Data Space Interface for external exchange. The architecture is explicitly cloud-native and technology-agnostic. The paper cites microservices, containerization, IaC, CI/CD, GitOps, and distributed orchestration as first-class operating principles; it also discusses OPC UA and MQTT, Kafka as an example event bus, and NOA-inspired OT/IT secure gateways. The Digital Twin Hub is positioned as a single source of truth for asset state, enabling both OLTP and OLAP consumers to subscribe to shared streams without hardwired hierarchical traversal.

Validation is twofold: comparative theoretical analysis against ISA-95, RAMI 4.0, NOA, FIWARE, AIOTI, LASFA, SITAM, and 8C, and quantitative evaluation through MAIA in a low-latency robotics scenario with autonomous mobile robots. The reported quantitative results are end-to-end latency below 50 ms across all tests on a commodity server, latency below 20 ms for up to 100 robots, an increase in the share of time spent in inter-service communication from 67.62% for 1 robot to 86.97% for 150 robots, and reductions of up to 80.74% in artifact size and 78.79% in memory consumption through tuned base images and JVM parameters. The paper presents these findings as evidence that the main bottleneck at scale is the communication fabric rather than per-service internal processing, and it explicitly notes that CRACI may be over-engineered for simpler static systems with hard real-time constraints.

6. Other specialized uses: random access, cognitive radio, and O-RAN conflict resolution

Further uses of CRACI appear as interpretive or functional labels in three additional domains. In cellular IoT random access, the summary of "NOMA-based Compressive Random Access using Gaussian Spreading" frames CRACI as a Compressive Random Access for Cellular IoT approach. The underlying model is layered CRA with Gaussian spreading, power-domain NOMA, and SIC over $L_2$1 layers with received power levels $L_2$2. Activity detection is analyzed via MAP under Gaussian approximations and implemented with a low-complexity CAVI procedure of per-iteration complexity $L_2$3. The large-system design rule is to choose per-layer powers satisfying $L_2$4 and $L_2$5, with recursion based on a target SNR $L_2$6. In the reported example with $L_2$7, $L_2$8, $L_2$9, and $z \in \mathbb{R}^{128}$0, average missed-detection/false-alarm counts decrease from 6.71 for $z \in \mathbb{R}^{128}$1 to 1.90 for $z \in \mathbb{R}^{128}$2 and 0.66 for $z \in \mathbb{R}^{128}$3, while very high $z \in \mathbb{R}^{128}$4 becomes counterproductive because interference departs from the Gaussian approximation (Choi, 2018).

In network information theory, CRACI is used as an interpretive label for the unified achievable-rate characterization of the cognitive radio channel with common message. The CRCC model contains a common message $z \in \mathbb{R}^{128}$5, private messages $z \in \mathbb{R}^{128}$6 and $z \in \mathbb{R}^{128}$7, and a cognitive sender with non-causal knowledge of the primary message. The paper develops a unified inner bound using auxiliaries $z \in \mathbb{R}^{128}$8, together with superposition coding, binning, and simultaneous decoding. Theorem 1 gives the achievable region for the modified CRCC in terms of $z \in \mathbb{R}^{128}$9; Theorem 2 reduces it by Fourier-Motzkin elimination to inequalities in $\tau = 0.15$0. The summary emphasizes that this inner bound subsumes several known regions, including the Hodtani CRC region, the Han-Kobayashi region for the IC, the ICC region, the CMACC region, and the strong-interference common-message case, where the derived inner bound matches known capacity results under the stated strong-interference conditions (Monemizadeh et al., 2011).

In O-RAN conflict mitigation, the paper "ACCoRD: Actor-Critic Conflict Resolution with Deep learning for O-RAN xApps" states that it does not use the term CRACI, but that ACCoRD’s CR Agent is functionally equivalent to a learned CRACI concept in the Near-RT RIC. The setting involves indirect conflicts between MRO and MLB xApps over TTT, hysteresis, and CIO. ACCoRD uses a feed-forward actor-critic ANN trained with PPO-Clip; per-decision actions are NO_MODIFICATION, REJECTION_WITH_COOLDOWN, INCREASE_1, and DECREASE_1, with reward defined from changes in ping-pong handovers, radio link failures, and call blockages. Reported average penalties over 36 runs show that rule-based prioritization is slightly better in the small-UE scenario, but ACCoRD is best in the medium and large scenarios, with average penalties 402.84 and 549.87 versus 412.82 and 568.56 for the best rule-based baseline. The paper’s functional implication is that a learned Near-RT conflict-resolution agent can outperform fixed priority rules when conflict frequency and traffic load increase (Adamczyk et al., 21 May 2026).

Taken together, these usages show that CRACI has become a cross-domain index term whose meaning is determined almost entirely by local context. In speech assessment it denotes cross-lingual retrieval augmentation; in coreference it points to the CRAC shared task; in conformal segmentation it abbreviates conditional risk adaptation ideas; in industrial systems it is a named reference architecture; and in several other areas it serves as a convenient label for conceptually adjacent CRA, CRCC, or conflict-resolution frameworks rather than a canonical acronym.