MACA: Metadata-Aware Cross-Model Alignment

• MACA is a teacher–student framework that uses metadata (topic, intent, entities) to disambiguate semantically similar but operationally distinct FAQ items.
• It distills the offline judgments of a metadata-aware LLM reranker into a compact dense retriever, enabling efficient retrieval for short, underspecified queries.
• MACA employs cross-model margin loss and hard-negative selection to enhance top-rank disambiguation, outperforming standard dense retrievers in enterprise FAQ search.

Metadata-Aware Cross-Model Alignment (MACA) is a teacher–student retrieval framework for enterprise FAQ search in which a metadata-aware LLM reranker is used offline to supervise a compact dense retriever for deployment-time use without online LLM calls. It is designed for short, underspecified queries such as “foreign transaction fee refund” or “recent check status,” where surface semantic similarity alone is often insufficient because closely related FAQs differ in structured attributes such as topic, sub-topic, intent, or entity. In this formulation, “cross-model alignment” means aligning the student retriever’s ranking behavior with the ranking preferences and score margins of a stronger metadata-aware LLM reranker, rather than aligning hidden states directly (Gudipudi et al., 1 Jan 2026).

1. Definition, scope, and problem setting

MACA addresses enterprise retrieval over FAQ-style corpora where queries are short, ambiguous, and frequently omit crucial qualifiers. The motivating failure mode is not merely lexical mismatch. The central problem is metadata-sensitive retrieval: multiple candidate FAQs may be semantically similar at the surface-text level while differing in topic, sub-topic, intent, or entity. The framework therefore treats metadata as a mechanism for disambiguating semantically close but operationally distinct candidates, such as debit-card versus credit-card activation or foreign transaction fee versus foreign ATM fee (Gudipudi et al., 1 Jan 2026).

The metadata used in MACA is induced automatically and consists of topic, sub-topic, intent, and entities. Each FAQ item has a question, answer, and taxonomy labels induced by an LLM-based taxonomy creation process, and query metadata is also inferred. This design allows MACA to distinguish between items that look similar in text space but diverge in structured meaning. A common misconception is that MACA is merely a reranking heuristic. In the paper, it is instead a two-phase framework whose central claim is that metadata-aware judgments from an audited teacher can be distilled into a fast retriever (Gudipudi et al., 1 Jan 2026).

The framework is motivated by three limitations of standard alternatives. Standard dense retrievers rely primarily on semantic similarity in text and make limited use of structured metadata attached to candidate FAQs. Manual labeling is expensive and inconsistent when relevance depends on latent metadata. Online LLM reranking can resolve subtle distinctions, but per-query LLM inference is costly, latency-sensitive, and operationally brittle in production. MACA is therefore structured so that the teacher operates offline, whereas only the student retriever is used at inference time (Gudipudi et al., 1 Jan 2026).

2. Teacher construction and trustworthiness protocol

MACA begins with candidate generation from two dense retrieval views: query-to-question, denoted $q \to Q$, and query-to-answer, denoted $q \to A$. These rankings are fused with Reciprocal Rank Fusion, and the fused top-$K$ candidates are reranked using a metadata-aware LLM prompt. In the reported setup, the dense retrieval stage uses OpenAI text-embedding-3-large, and the offline teacher reranker is gpt-4o (Gudipudi et al., 1 Jan 2026).

Before reranking, MACA induces a taxonomy automatically using an LLM. The taxonomy spans

$$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$

The prompt then asks the teacher to infer query-level metadata and score each candidate using both text and metadata alignment. The prompt summary specifies the metadata score as

$$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$

and the relevance score as

$$\mathrm{relevance} = 0.65 \times \mathrm{coverage} + 0.35 \times \mathrm{metadata}.$$

Candidates are also assigned four-point TREC-style labels: exact, partial, less relevant, and irrelevant, mapped to $3,2,1,0$ (Gudipudi et al., 1 Jan 2026).

A distinguishing feature of MACA is that the teacher is not assumed trustworthy a priori. Prompt variants including CoT, ReAct, Reflexion, MACA, and MAFA are audited using performance, consistency under candidate permutation, and robustness to paraphrases. The consistency metric is

$$C@k = \frac{1}{|Q|}\sum_{q\in Q} \frac{1}{P_q} \max_{S}\#\{\,p : \mathrm{topk}_p(q)=S\,\},$$

and the robustness metric is

$$R@k = \frac{1}{|Q|}\sum_{q\in Q} \frac{1}{M_q} \max_{S}\#\{\,i : \mathrm{topk}(q^{(i)})=S\,\}.$$

These definitions operationalize trustworthiness as stable ranking behavior under candidate reordering and query paraphrasing, rather than as an abstract property of the LLM itself (Gudipudi et al., 1 Jan 2026).

3. Distillation signals and the MetaFusion objective

After prompt selection, MACA uses a MACA Judge to convert teacher outputs into student-training supervision. The judge sees the teacher top-$k_T$ list with teacher scores $q \to A$0, the student top-$q \to A$1 list with student scores $q \to A$2, metadata, and teacher labels. It outputs a teacher-selected positive $q \to A$3, a hard negative $q \to A$4, and a teacher margin $q \to A$5 (Gudipudi et al., 1 Jan 2026).

The hard negative is chosen from a metadata-aware near-miss set,

$q \to A$6

where overlap tests for shared topic or intent and allows sub-topic or entity differences. The selected hard negative is

$q \to A$7

This construction makes MACA explicitly target semantically close but metadata-mismatched competitors (Gudipudi et al., 1 Jan 2026).

The teacher margin is defined as

$q \to A$8

with a label-aware floor $q \to A$9 and clipping to $K$0. The student margin is

$K$1

and the cross-model margin loss is

$K$2

The total MetaFusion objective is

$K$3

The paper explicitly states that $K$4 is standard in-batch Multiple Negatives Ranking Loss, but it does not provide its explicit formula (Gudipudi et al., 1 Jan 2026).

An important clarification is that MACA’s metadata-awareness is concentrated in the teacher, the judge, and the construction of positives, hard negatives, and margins. The student is not described as having a separate metadata encoder or explicit metadata fusion architecture. It uses off-the-shelf retrieval encoders—all-MiniLM-L6-v2, all-mpnet-base-v2, and msmarco-distilbert-base-tas-b—and learns metadata-sensitive distinctions through supervision rather than through a bespoke metadata input pathway (Gudipudi et al., 1 Jan 2026).

4. Position within metadata-aware alignment research

MACA belongs to a broader family of methods that use side information to shape alignment, but it occupies a distinctive position because it aligns a compact retriever to a metadata-aware LLM reranker rather than learning a unified encoder over raw multimodal inputs. Related work shows several neighboring design patterns: metadata-conditioned contrastive learning, shared latent-space objectives with structural regularization, object-level cross-architecture supervision, and sparse-anchor transport across pretrained latent spaces (Yuan et al., 14 Feb 2026, Zhang et al., 5 Mar 2025, Camuffo et al., 17 Sep 2025, Gholamzadeh et al., 18 May 2025).

Method	Alignment unit	Metadata or side-information role
MACA (Gudipudi et al., 1 Jan 2026)	Query–FAQ ranking behavior	Topic, sub-topic, intent, and entities shape teacher scoring, hard negatives, and margins
sleep2vec (Yuan et al., 14 Feb 2026)	Synchronized biosignal segments	Age, gender, site, and subject-night history weight negatives and soften pseudo-negatives
AlignXpert (Zhang et al., 5 Mar 2025)	Modality embeddings in a shared latent space	No explicit metadata-aware mechanism; similarity, regularization, and stress define alignment
MOCHA (Camuffo et al., 17 Sep 2025)	Object regions across teacher and student architectures	Boxes, class labels, and pairwise region relations act as metadata-like supervision
Model alignment using inter-modal bridges (Gholamzadeh et al., 18 May 2025)	Latent samples across pretrained models	Paired anchors and bridge costs define couplings for conditional flow alignment

These comparisons clarify what MACA is and is not. It is not a contrastive multimodal pretraining method of the type used in sleep2vec, where metadata reshapes the negative set directly through a loss such as DASH-InfoNCE (Yuan et al., 14 Feb 2026). It is also not a shared-space optimization objective like AlignXpert, whose formulation is “maximize similarity $K$5 regularization $K$6 stress” but is not metadata-aware as written (Zhang et al., 5 Mar 2025). Nor is it an object-level distillation framework like MOCHA, where ground-truth boxes, class labels, and teacher region embeddings provide structured side information for cross-architecture alignment (Camuffo et al., 17 Sep 2025). Instead, MACA is a retrieval-specific instance of metadata-aware cross-model supervision: metadata changes which distinctions the teacher considers important, and those distinctions are distilled into the student (Gudipudi et al., 1 Jan 2026).

A plausible implication is that MACA can be viewed as one point in a larger design space. The related papers suggest at least three general mechanisms for metadata-aware alignment: metadata-conditioned hardness over negatives (Yuan et al., 14 Feb 2026), explicit structure-preserving shared-space objectives (Zhang et al., 5 Mar 2025), and sparse-anchor couplings with bridge costs between pretrained latent spaces (Gholamzadeh et al., 18 May 2025). MACA instantiates a fourth: metadata-aware teacher auditing and margin transfer for efficient retrieval (Gudipudi et al., 1 Jan 2026).

5. Empirical results and observed behavior

On the teacher side, MACA outperforms prompt-level baselines on both evaluation corpora. On the proprietary consumer banking set, MACA reports Acc@1/3/5 of $K$7, compared with MAFA at $K$8. On BankFAQs, MACA reports $K$9, compared with MAFA at $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$0. The abstract summarizes these differences as about 5 points Acc@1 gain on the proprietary set and about 3 points on BankFAQs (Gudipudi et al., 1 Jan 2026).

The student results show that most of the performance gain over pretrained retrievers comes from metadata-aware ranking supervision, with MetaFusion adding smaller but consistent improvements at top rank. On the proprietary corpus, MiniLM improves from $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$1 Acc@1 in the pretrained baseline to $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$2 with MNRL and to $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$3 with MetaFusion. MPNet improves from $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$4 to $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$5 to $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$6, and TAS-B from $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$7 to $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$8 to $$(\text{topic}, \text{sub-topic}, \text{intent}, \text{entities}).$$9. On BankFAQs, gains are smaller because the task is easier, but top-rank improvements remain consistent: MiniLM rises from $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$0 to $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$1 to $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$2, MPNet from $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$3 to $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$4 to $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$5, and TAS-B from $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$6 to $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$7 to $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$8 (Gudipudi et al., 1 Jan 2026).

The paper emphasizes that MetaFusion mainly affects strict top-rank disambiguation rather than broader recall. On the proprietary corpus, the gain from MNRL to MetaFusion is $$\mathrm{metadata} = 0.40\,[\mathrm{intent}] + 0.25\,[\mathrm{topic}] + 0.20\,[\mathrm{subtopic}] + 0.15 \times \mathrm{Jaccard},$$9 for MiniLM, $$\mathrm{relevance} = 0.65 \times \mathrm{coverage} + 0.35 \times \mathrm{metadata}.$$0 for MPNet, and $$\mathrm{relevance} = 0.65 \times \mathrm{coverage} + 0.35 \times \mathrm{metadata}.$$1 for TAS-B, while Acc@5 remains nearly unchanged. This is consistent with the intended role of RCMA: preserving teacher-implied margins against metadata-confusable near misses rather than globally reshaping the whole ranking distribution (Gudipudi et al., 1 Jan 2026).

The qualitative example highlighted in the paper is entity drift. A baseline retriever returns the same debit-card FAQ for both “activate my debit card” and “activate my credit card,” whereas MACA returns the correct debit-card and credit-card FAQs at rank 1. This behavior illustrates the specific kind of distinction MACA is designed to preserve: semantically close candidates that differ in latent metadata fields (Gudipudi et al., 1 Jan 2026).

6. Misconceptions, limitations, and extensions

A common misconception is that MACA makes the deployed retriever metadata-aware by explicit architectural design. In the reported framework, metadata-awareness is primarily transferred through supervision: metadata-aware prompting, metadata-aware judge selection, and metadata-aware teacher margins. The student architecture itself is an off-the-shelf encoder, and the paper does not describe metadata concatenation, a metadata encoder, or a metadata-fusion head in the retriever (Gudipudi et al., 1 Jan 2026).

Another misconception concerns calibration. The paper repeatedly uses the term “calibrated,” but it does not define a separate probabilistic calibration procedure such as temperature scaling. What is explicitly specified is prompt selection based on accuracy, consistency, and robustness, together with a teacher-margin floor $$\mathrm{relevance} = 0.65 \times \mathrm{coverage} + 0.35 \times \mathrm{metadata}.$$2 and clipping to $$\mathrm{relevance} = 0.65 \times \mathrm{coverage} + 0.35 \times \mathrm{metadata}.$$3. The paper also states that richer notions of calibration and uncertainty-aware labeling remain future work (Gudipudi et al., 1 Jan 2026).

The limitations are correspondingly clear. The experiments are in banking-domain FAQ retrieval; extension beyond banking is identified as future work. Richer metadata such as user or channel context is not yet included. Exact numerical values for $$\mathrm{relevance} = 0.65 \times \mathrm{coverage} + 0.35 \times \mathrm{metadata}.$$4 are not reported. Code release is pending organizational approval. The student does not explicitly incorporate metadata as a separate input module, so MACA’s metadata-awareness is mostly in teacher and loss construction rather than in student architecture (Gudipudi et al., 1 Jan 2026).

Related papers suggest several concrete extension paths. A plausible extension is metadata-conditioned hardness modeling over negatives, as in DASH-InfoNCE, where age, gender, site, and history reshape the denominator and soften pseudo-negatives (Yuan et al., 14 Feb 2026). Another is to add shared-space structural penalties, such as similarity maximization plus stress preservation, following the design pattern of AlignXpert, although that paper is not metadata-aware as written (Zhang et al., 5 Mar 2025). A third is to use sparse-anchor bridge costs and conditional flow matching to align latent spaces when only limited trusted correspondences are available (Gholamzadeh et al., 18 May 2025). These directions suggest that MACA, in its published form, is best understood as a retrieval-specific instantiation of metadata-aware cross-model supervision rather than a complete general theory of metadata-aware alignment.