--- title: 'MRC: Multi-Domain Applications in Research' url: https://www.emergentmind.com/topics/mrc type: topic --- # MRC: Multi-Domain Applications in Research In the cited arXiv literature, **MRC** denotes several distinct technical objects rather than a single unified concept. The most common usage is **machine reading comprehension**, a family of NLP tasks in which a model answers natural-language questions from a passage, often by extracting a span, selecting an option, or deciding that no answer exists [2107.08582][2210.13778][2503.07144]. In communications and signal processing, **MRC** denotes **maximum ratio combining** or **maximal-ratio combining**, a receiver strategy studied in massive MIMO, cooperative NOMA, and cell-free architectures [1703.10422][1412.6747][2012.13504]. More specialized usages include **Market Regime Council**, a cooperative multi-agent LLM system for online portfolio management [2605.24490], and the **Molonglo Reference Catalogue**, a radio-source catalogue used in H I absorption surveys [2310.14571]. ## 1. Principal expansions and disciplinary usage The abbreviation appears with materially different semantics across research areas. The following map captures the usages explicitly documented in the cited papers. | Expansion of MRC | Domain | Representative use | |---|---|---| | Machine Reading Comprehension | NLP / NLU | extractive QA, multiple-choice QA, unanswerable QA | | Maximum Ratio Combining / maximal-ratio combining | wireless communications | massive MIMO, NOMA, radio-stripe fronthaul | | Market Regime Council | multi-agent LLM decision systems | Shapley-based portfolio allocation | | Molonglo Reference Catalogue | radio astronomy | FLASH pilot survey target selection | Within NLP, machine reading comprehension has also become a modeling substrate for tasks not usually framed as question answering. The cited literature reformulates named-entity recognition, semantic role labeling, and dependency parsing as MRC-style prediction problems, suggesting that “MRC” in this domain often denotes both a benchmark family and a general interface for conditioning structured prediction on natural-language queries [2305.03970][2109.06660][2105.07654]. ## 2. Machine reading comprehension in natural language processing In its extractive form, machine reading comprehension is the task of answering natural-language questions by locating an appropriate span in a document. Formally, given a document $\mathcal{D}$ with tokens $\{w_1,\dots,w_L\}$ and a question $Q$, an extractive MRC model selects a contiguous answer span $a = w_{i:j} = (w_i,\dots,w_j)$ in $\mathcal{D}$ [2107.08582]. This formulation underlies standard benchmark evaluations such as SQuAD 1.1, NewsQA, TriviaQA, and SimpleNQ, typically scored with Exact Match and token-level F1 [2107.08582]. A major line of work concerns **unanswerability**. The Indonesian benchmark **IDK-MRC** was created because existing Indonesian MRC datasets lacked unanswerable questions. It combines automatic and manual unanswerable-question generation and, together with answerable questions, consists of **more than 10K questions in total** [2210.13778]. Its train/dev/test splits contain **5,042 / 382 / 422 answerable** and **4,290 / 382 / 422 unanswerable** questions, respectively, and its unanswerable set is balanced across six types: **negation**, **antonym**, **entity swap**, **question-tag swap**, **specific condition**, and **other** [2210.13778]. On the reported XLM-R comparison, training on IDK-MRC yields **UnAns EM = 88.29**, **UnAns F1 = 88.29**, **Overall EM = 74.86**, and **Overall F1 = 81.37**, versus **66.78 / 66.78 / 59.00 / 65.89** for TranslatedSQuAD and **0.90 / 0.90 / 33.01 / 39.74** for TyDiQA [2210.13778]. A second line concerns **benchmark completeness**. **MRCEval** introduces a three-level taxonomy of **13 distinct RC skills**: six under context comprehension, three under external knowledge comprehension, and four under reasoning [2503.07144]. It contains **2,103 multi-choice questions** and evaluates **28** open-source and proprietary models with accuracy as the sole metric [2503.07144]. The reported leaders are **Qwen-max-2025 (59.4%)**, **Gemini-2.0-flash (59.3%)**, **o3-mini (59.0%)**, and **DeepSeek-R1-Distill-Qwen-14B (56.4%)**, while counterfactual faithfulness remains a universal blind spot [2503.07144]. This benchmark positions MRC not merely as span extraction but as a broader operationalization of reading comprehension. ## 3. MRC as a reformulation framework for structured prediction Several papers use machine reading comprehension as a **conversion layer** for tasks that are not originally QA. In **NER-to-MRC**, Zhang et al. completely frame named-entity recognition as a machine reading comprehension problem [2305.03970]. The passage is the original token sequence $P = X$, the question is the fixed universal prompt **“What kind of entity is this?”**, and each option $O_j$ is the natural-language description of an entity type taken from the annotation guidelines [2305.03970]. The model input for option $j$ is the concatenation $[P;Q;O_j]$, processed in parallel across options with a shared **DeBERTa-v3** encoder, followed by **HRCA layers** that **review** the question, **read** the option with cross-attention to the question, and **find** the passage with cross-attention to the enriched option representation [2305.03970]. The prediction tensor is $M^{pred}\in\mathbb{R}^{k\times n_0\times 2}$, trained with per-token, per-option binary cross-entropy: $$ \ell_{i,j} = -\bigl[ m_{i,j}\log p_{i,j} +(1-m_{i,j})\log(1-p_{i,j}) \bigr], $$ with overall loss $$ \mathcal{L}_{\mathrm{overall}} = \sum_{j=1}^{n_0}\sum_{i=1}^{k}\ell_{i,j}. $$ Across **six benchmark datasets from three domains**, the system reports state-of-the-art performance without external data, including **65.61** F1 on **WNUT-16**, where the paper states an improvement of **+11.24%** over the best prior model [2305.03970]. In **semantic role labeling**, Wang et al. formalize **predicate disambiguation** as multiple-choice MRC and **argument labeling** as extractive MRC [2109.06660]. Candidate predicate senses are retrieved from PropBank frame files, each paired with its sense description $d_i^s$, and the input is $$ \texttt{[CLS] } d_i^s \texttt{ [SEP] } x_1 \dots \langle p\rangle p \langle /p\rangle \dots x_n \texttt{ [SEP]}. $$ A sigmoid-scored FFN predicts the correct sense, while argument labeling uses role-specific queries such as **“What are the X arguments of predicate Y with meaning Z?”** for core roles and **“What are the W modifiers of predicate Y?”** for non-core roles [2109.06660]. RoBERTa-based results include **97.3** predicate-disambiguation accuracy on CoNLL-2009 WSJ and **93.3 / 87.2** F1 on CoNLL-2009 WSJ/Brown for dependency SRL, with **90.0 / 85.1 / 88.3** F1 on CoNLL-2005 WSJ, CoNLL-2005 Brown, and CoNLL-2012 for span SRL [2109.06660]. In **dependency parsing**, Gan et al. model parsing as **MRC-based span-span prediction** [2105.07654]. Candidate subtrees are proposed as spans $(root,start,end)$, and span linking is formalized as MRC where the **context** is the full sentence, the **query** is a marked child span, and the **answer** is the parent span [2105.07654]. The query is encoded by inserting the markers **``, ``, ``, ``** around the child span in the sentence. Separate scores are predicted for the parent **head**, **start**, **end**, and **dependency label**, and a reverse child-query can be added for mutual consistency [2105.07654]. On PTB, the RoBERTa-based parser reaches **97.24% UAS / 95.49% LAS**; on CTB it reaches **93.14% UAS**; and it wins **11 of 12 UD languages** with average **+0.30 LAS** [2105.07654]. ## 4. Learning regimes for MRC: self-supervision, masking, and domain adaptation One research direction seeks to align pre-training more closely with the span-spotting structure of MRC. In **“Bridging the Gap between Language Model and Reading Comprehension: Unsupervised MRC via Self-Supervision”**, the pretext task **Spotting-MLM** masks an entire occurrence of an informative repeated span and requires the model to predict the start and end indices of another occurrence [2107.08582]. The masked query representation $\overline{x}_{mask}$ is obtained from the `[MASK]` token, and start and end probabilities are computed with bilinear scoring: $$ P_{start}(i\mid p_{mask}) = \mathrm{softmax}\bigl(x_i^\top W_s \overline{x}_{mask}\bigr),\qquad P_{end}(j\mid p_{mask}) = \mathrm{softmax}\bigl(x_j^\top W_e \overline{x}_{mask}\bigr). $$ The loss is the sum of negative log-likelihoods over gold start and end positions [2107.08582]. On four unsupervised benchmarks, the reported average score is **32.0 EM / 42.6 F1**, compared with **29.6 / 39.1** for Fabbri et al. (2020), and on SQuAD-Adv “AddSent” the model reaches **39.9 F1** [2107.08582]. A second direction studies whether **masking length distribution** should match downstream answer-length distribution. The Chinese study **“Analysing the Effect of Masking Length Distribution of MLM”** defines four MRC tasks—**short-span extraction**, **long-span extraction**, **short multiple-choice cloze**, and **long multiple-choice cloze**—and creates four corresponding Chinese datasets [2110.15712]. It then trains four MLMs whose masking-span-length distributions are set to $$ P_D(l)=\frac{x_l}{N_D}, $$ where $x_l$ is the number of answers of length $l$ in dataset $D$ [2110.15712]. In every reported case, the MLM whose mask-length distribution matches the target answer distribution achieves the highest downstream score, although the gains are described as **modest**—approximately **1–2% absolute**—and human performance remains far higher [2110.15712]. This suggests that MRC performance is sensitive not only to architecture but also to the structural statistics of pre-training corruption. A third direction addresses **domain shift**. **BioADAPT-MRC** adapts from general-purpose QA to biomedical QA without using synthetic or human-annotated biomedical target data [2202.13174]. The model comprises a shared **BioELECTRA** feature extractor $M_F$, an answer predictor $M_Q$, and a domain similarity discriminator $D$ with a triplet-loss head and an auxiliary QA head [2202.13174]. The combined objective is $$ \mathcal{L}_{\rm total} = \mathcal{L}_{\rm MRC} - \lambda \mathcal{L}_{\rm adv}, $$ implemented through a gradient-reversal layer [2202.13174]. Using **SQuAD-1.1** as source data and **BioASQ-7b/8b/9b** contexts as unlabeled targets, the model reports **0.45 / 0.64 / 0.53** on BioASQ-7b, **0.38 / 0.62 / 0.47** on BioASQ-8b, and **0.54 / 0.74 / 0.62** on BioASQ-9b for **SAcc / LAcc / MRR**, together with **0.21 EM / 0.40 F1** on emrQA [2202.13174]. ## 5. Maximum ratio combining in wireless communications In communications, **MRC** denotes **maximum ratio combining** or **maximal-ratio combining**, a receiver rule that weights branches by channel quality. In the dual-branch interference-aware model analyzed under spatial interference correlation and Nakagami fading, the post-combiner SINR is $$ \mathrm{SINR} = \frac{g_1}{I_1+\mathrm{SNR}^{-1}} + \frac{g_2}{I_2+\mathrm{SNR}^{-1}}, $$ where the two interference powers are correlated because they arise from the same Poisson field of interferers [1312.5938]. The paper derives an exact success-probability expression for dual-branch MRC, shows that assuming independent interference across antennas is highly optimistic, and shows that the common full-correlation approximation is slightly pessimistic but becomes exact as the interferer Nakagami parameter $m_I\to\infty$ [1312.5938]. In **uplink massive MIMO**, the cited stochastic-geometry analysis decomposes the interference under MRC into **intra-cell interference**, **inter-cell interference**, and **pilot-contamination interference** [1412.6747]. For practical but still large numbers of BS antennas, the key observation is that **intra-cell interference accounts for the dominant portion of the total interference** under MRC reception, whereas ZF substantially reduces this component while leaving inter-cell interference and pilot contamination almost unchanged [1412.6747]. The paper also notes that pilot-contamination interference and inter-cell interference have roughly equal mean powers, but pilot contamination has a much wider distribution range under strong shadowing [1412.6747]. Timing asynchrony fundamentally changes MRC behavior in large-scale MU-MIMO. In **“On the Performance of MRC Receiver with Unknown Timing Mismatch”**, neglecting unknown user delays causes the MRC achievable rate to saturate because residual ISI persists even as $M$ grows [1703.10422]. Under perfect CSI and power scaling $\rho_d = E_d/M$, ordinary MRC converges to $$ R_{l,p}^{\rm mrc}\to \log_2\!\Bigl( 1+\frac{E_d\beta_l E[g_0]^2}{E_d\beta_l\sum_{i\ne 0}E[g_i]^2+1} \Bigr), $$ which is capped at high $E_d$ by average ISI [1703.10422]. The proposed **MRC-ZF** post-processing removes average ISI using only the distribution of the timing mismatch, and the paper reports restored large-$M$ power scaling together with close agreement between theory and simulation [1703.10422]. In **vehicular communications at road intersections**, cooperative NOMA with MRC yields closed-form outage expressions and improves over both standard cooperative NOMA and MRC-OMA [2003.07299]. Under the reported settings, **MRC-NOMA reduces the outage of $D_2$ by about 34%** versus standard NOMA relay and exceeds **60%** outage gain over MRC-OMA for $\mathcal{R}_2\ge 1$ bit/s/Hz [2003.07299]. The paper also shows that outage peaks near the road intersection, where both perpendicular roads contribute interference [2003.07299]. In **cell-free massive MIMO with radio stripes**, MRC also appears in the fronthaul design. Each AP forwards only the local matched-filter outputs, giving a fronthaul load of $$ 2K(T_c-T_p) $$ real scalars per coherence [2012.13504]. The proposed **Q-LMMSE** reconstructs interference structure from the covariance of the MRC stream via a small $K\times K$ SVD and computes $$ \mathbf{S}_{\rm Q\mbox{-}LMMSE} = \mathbf{U}\mathbf{D}^{-1}\mathbf{U}^H \mathbf{S}_{\rm MRC}. $$ The paper reports that Q-LMMSE increases spectral efficiency by **347.1% over MRC** and **51.3% over N-LMMSE** in the correlated-channel case with $T_c=720$, while retaining the same MRC fronthaul [2012.13504]. ## 6. Other specialized meanings: Market Regime Council and the Molonglo Reference Catalogue In multi-agent LLM decision systems, **MRC** denotes **Market Regime Council** [2605.24490]. This framework treats specialist agents as players in a transferable-utility game, evaluates all non-empty coalitions, computes exact Shapley values, and uses those credits for online weighting [2605.24490]. With **$N=3$** agents, it runs three deliberation stages: private single-agent portfolios, pairwise debate portfolios, and a grand-coalition portfolio with a readout operator [2605.24490]. Coalition value is defined from exponentially weighted estimates of Sharpe and annualized return, $$ v(S;\mathcal{H}_t) = \gamma_\rho \sqrt{365}\, \frac{\hat\mu_{\mathrm{EW}(S)}}{\hat\sigma_{\mathrm{EW}(S)}} + \gamma_\mu \mu_{\mathrm{ann}(S)}, \qquad \gamma_\rho=0.4,\ \gamma_\mu=0.6, $$ and each period also applies a Bayesian adaptive mixture, regime-dependent multipliers, and a five-layer causal trace [2605.24490]. Over **1,037 trading days**, **13 crypto assets**, and **five seeds**, the reported performance is **CR = 440.1%**, **SR = 1.51**, **MDD = 34.1%**, and **IR = 0.47**, ranking first on **CR**, **SR**, and **IR** among active baselines and achieving the lowest **MDD** among active methods [2605.24490]. In radio astronomy, **MRC** denotes the **Molonglo Reference Catalogue**. The FLASH pilot survey searched for associated H I 21-cm absorption against bright radio sources from the **MRC 1-Jy sample** [2310.14571]. The final sample contained **62** MRC galaxies and quasars with **$0.42