Bi-Encoder Single-Vector Paradigm

• Bi-encoder single-vector paradigm is a neural architecture where paired encoders independently map queries and candidates into a shared fixed-size embedding space.
• It leverages contrastive learning and precomputed offline encoding to enable efficient similarity scoring for tasks like semantic search, cross-lingual matching, and biometrics.
• The approach offers substantial speed benefits, though it trades off some accuracy due to its information bottleneck and limited cross-term interactions.

A bi-encoder single-vector paradigm is a neural architecture in which two (potentially identical) encoders map two objects—typically queries and candidates—independently into fixed-size embeddings in a shared vector space. The primary operation is to compute the similarity (dot product or cosine) between these vectors to perform matching or retrieval. This approach is a core technique for scalable information retrieval, semantic search, paraphrase identification, cross-modal matching, and self-supervised representation learning, offering substantial efficiency advantages at inference by enabling corpus-side offline encoding and rapid vector search. The paradigm has been applied across modalities, from dense and sparse text retrieval to biometrics, NER, and cross-lingual tasks (Lavi, 2021, Fedorova et al., 2024, So et al., 27 Oct 2025, Chen et al., 2023, Zhang et al., 2022).

1. Architectural Fundamentals

The canonical bi-encoder single-vector system comprises two neural network encoders:

• $f_q$ (query encoder): maps input queries $q$ to embedding $u_q \in \mathbb{R}^d$.
• $f_d$ (document/item encoder): maps candidates $d$ to embedding $u_d \in \mathbb{R}^d$.

Each input is encoded independently. The single-vector constraint refers to the fact that for each input object, only one (fixed-length) vector is produced—there is no multi-vector or interaction across the two objects prior to similarity computation.

At inference, the similarity function is typically the unnormalized inner product or cosine: $$s(q,d) = \langle u_q, u_d \rangle \quad \text{or} \quad \frac{u_q^\top u_d}{\|u_q\| \|u_d\|}$$ This facilitates efficient nearest neighbor search: all candidates may be encoded and indexed offline, and retrieval becomes a Maximum Inner Product Search (MIPS) or cosine similarity query (Tran et al., 2024, Chen et al., 2023).

The architecture may use tied (siamese) parameters (as in SBERT, LaBSE) or separate encoders for the two modalities or data domains (Lavi, 2021, Fedorova et al., 2024, So et al., 27 Oct 2025).

2. Training Objectives and Optimization

Bi-encoders are typically trained with contrastive learning objectives over positive and negative pairs. Let $(q, d^+)$ be a positive pair, and $\{d^-\}$ negatives (either in-batch, from the corpus, or hard-mined): $$\mathcal{L} = -\sum_{i=1}^B \log \frac{\exp(u_{q_i}^{T}u_{d^+_i} / \tau)}{\sum_{j=1}^B \exp(u_{q_i}^{T}u_{d^-_j} / \tau)}$$ where $\tau$ is a temperature parameter (Lavi, 2021, Tran et al., 2024, Chen et al., 2023, Fedorova et al., 2024).

Margin-based losses (as in biometrics) and additive margin softmax variants are also employed: $$\mathcal{L}(x_1,x_2,y) = (1-y)\|g(z_1)-g(z_2)\|_2^2 + y [\max(0, m - \|g(z_1)-g(z_2)\|)]^2$$ with $y$ indicating positive/negative (So et al., 27 Oct 2025).

Negative sampling is crucial. Common strategies:

• In-batch negatives: all other batch pairs (Lavi, 2021, Fedorova et al., 2024).
• Hard negatives: mined using current model predictions or from curated sources (Fedorova et al., 2024).
• Full (or partial) universe negatives: as in span enumeration for NER (Zhang et al., 2022).

Fine-tuning regimes typically use Adam or AdamW, learning rates $10^{-5} \ldots 3 \times 10^{-4}$, small batch sizes, and modern regularization (Lavi, 2021, Fedorova et al., 2024, So et al., 27 Oct 2025, Zhang et al., 2022).

3. Core Applications and Examples

Text and Cross-Lingual Matching

• In multilingual job–resume matching (Lavi, 2021), dual mBERT encoders are fine-tuned with an in-batch softmax loss on CV–vacancy pairs, achieving robust semantic matching and sublinear retrieval complexity.
• In cross-lingual paraphrase identification, a LaBSE shared encoder is used to generate 768-dim vectors for each sentence, optimized with a margin softmax loss over both in-batch and hard negatives, attaining 79.3% mean accuracy (7–10% below cross-encoders, but with major runtime gains) (Fedorova et al., 2024).

Biometrics

• In (So et al., 27 Oct 2025), fingerprint and iris images are mapped via ResNet or Vision Transformer backbones into Euclidean-embedded spaces. Matching is performed by the L2 distance between vectors. Positive verification performance (e.g., 0.91 ROC-AUC for iris) demonstrates the paradigm's applicability beyond language.

Dense and Sparse Text Retrieval

• The paradigm unifies dense models (e.g., DPR, SBERT) and modern sparse methods (e.g., SPLADE++) by treating both as neural encoders into a search-efficient vector space. Efficient CPU/Java retrieval is demonstrated via tight Lucene/ONNX integration for both types of embeddings (Chen et al., 2023).

Named Entity Recognition

• BINDER (Zhang et al., 2022) frames NER as a retrieval problem: candidate spans and entity types are encoded independently, and span–type matching is performed by cosine similarity. Multi-part InfoNCE-style losses align spans with types, yielding strong SOTA results and rapid inference.

Application	Encoder Backbone(s)	Similarity	Loss Type(s)	Negatives
CV–Vacancy Matching	mBERT (dual)	Cosine	In-batch softmax	In-batch
Paraphrase ID	LaBSE (shared)	Cosine	Additive margin softmax	In-batch, hard
Biometrics	ResNet, ViT (dual)	L2 distance	Margin-based contrastive	Balanced pairing set
Dense/Sparse IR	Transformers, SPLADE	Dot/cosine	Contrastive, CE + L1	In-batch, hard
NER (BINDER)	BERT-based, dual	Cosine	InfoNCE, dynamic thresh	Enumerated (all spans)

4. Efficiency, Scalability, and Deployment Advantages

The fundamental gain of the single-vector bi-encoder paradigm is decoupled, offline embedding—allowing corpus-candidate vectors to be precomputed and stored, dramatically reducing inference cost. Retrieval becomes a fast MIPS or nearest-neighbor query (e.g., via HNSW, IVF+PQ, or inverted index for sparse vectors) (Tran et al., 2024, Chen et al., 2023). In cross-lingual settings, a single model (e.g., LaBSE, mBERT) provides zero-shot transfer and cross-market coverage (Lavi, 2021, Fedorova et al., 2024). For structured retrieval setups (e.g., NER, type–span search), scoring is likewise reduced to simple vector operations (Zhang et al., 2022).

In empirical comparisons, single-vector bi-encoders achieve sub-millisecond retrieval per query at corpus scale and enable trivial scaling via ANN libraries (e.g., FAISS, Lucene HNSW) (Chen et al., 2023, Fedorova et al., 2024). A typical trade-off is a 7–10% drop in top-K accuracy compared to cross-encoder models, compensated by two to three orders of magnitude speedup (Fedorova et al., 2024).

5. Paradigm Limitations and Critique

Information Bottleneck

Encoding both query and candidate into single, fixed-length vectors constrains interaction to the final similarity score; fine-grained cross-term dependencies or contextual nuances lost at encoding time cannot be recovered. Cross-encoders, which jointly encode input pairs, consistently outperform bi-encoders on in-domain and especially zero-shot benchmarks (Tran et al., 2024, Fedorova et al., 2024).

The information bottleneck means task-irrelevant but potentially transfer-critical details are not encoded. Domain adaptation and generalization can thus be weak, with zero-shot retrieval sometimes suffering substantial accuracy loss.

Assumptions and Alignment Issues

The single-vector paradigm assumes that relevance is faithfully modeled by dot or cosine similarity in latent space. This fails when true interaction is highly non-linear, cross-modal, or sensitive to minute differences not linearly separable. For biometrics, cross-modal fingerprint–iris matching remains close to random, indicating that the paradigm may not bridge truly disparate modalities (So et al., 27 Oct 2025).

Overfitting and Training Inefficiency

With large Transformer backbones, fine-tuning the full encoding function to each new dataset is computationally expensive and prone to overfitting due to parameter–data imbalance (Tran et al., 2024).

6. The Encoding–Searching Separation Perspective

Recent work (Tran et al., 2024) introduces a two-stage conceptualization:

• Generic Encoding: $f_q', f_d$ are kept frozen or lightly tuned, producing rich, task-agnostic embeddings.
• Searching Module: $g$ is a lightweight, fast transformation over the generic embedding—e.g., a shallow MLP or attention over dimensions—enabling task specialization while preserving generality.

This separation localizes the information bottleneck, improves zero-shot transfer, and allows for architectural extensions (multi-vector, compositional structures, dynamic attention) while keeping search fast and memory efficient. The formulation is: $$P(\mathrm{rel}=1\mid q,d) = \sigma\left(\langle g(f_q'(q)), f_d(d) \rangle\right)$$ This perspective offers theoretical clarity on where and why the single-vector paradigm fails, supporting remedies such as adaptive search heads, hybrid structured outputs, and selective fine-tuning.

7. Practical Implications and Design Trade-offs

The bi-encoder single-vector paradigm is appropriate when:

• Corpus-side objects can be pre-encoded (static or infrequently updated).
• Inference latency and scalability dominate the system design constraints.
• Near SOTA accuracy (within 7–10% of cross-encoder) is tolerable for large-scale retrieval (Fedorova et al., 2024).
• Side-by-side