Encoder-Level Entanglement

• Encoder-level entanglement is a concept where the encoder’s design and resource states govern the formation and measurability of nonlocal correlations.
• It impacts both machine learning and quantum information by either leveraging shared representations to enhance performance or by using controlled disentanglement for robustness.
• Design implications include optimizing multitask ASR, quantum coding, and simulation strategies, highlighting trade-offs between exploitable entanglement and necessary task separation.

Searching arXiv for recent and directly relevant papers on encoder-level entanglement across machine learning, quantum simulation, quantum coding, and communication. arxiv_search(query="encoder-level entanglement OR entanglement-assisted encoder OR disentangling quantum autoencoder", max_results=10) Encoder-level entanglement denotes a family of phenomena in which the encoder stage determines the form, accessibility, or harmfulness of correlations, rather than leaving those issues entirely to the output state or the decoder. The literature uses the expression in several non-equivalent ways. In dual-output second-language speech recognition, it refers to shared-encoder task coupling that can improve meaning recognition while degrading surface transcription, with English showing nearly identical task representations through most encoder layers (Cho et al., 4 Jun 2026). In quantum information, it can mean an encoding into an enlarged Hilbert space that makes anti-linear entanglement expressions directly measurable (Candia et al., 2013), the use of shared ebits that changes what a quantum encoder can implement (Wilde et al., 2010), or bulk-to-boundary maps whose internal bond states carry tunable entanglement (Qi et al., 15 Dec 2025). This suggests that encoder-level entanglement is not a single invariant, but a structural notion: correlations are governed upstream, at the level of shared representations, resource states, or encoder circuitry.

1. Scope and recurring mathematical forms

Across the cited works, encoder-level entanglement appears in several mathematically distinct objects: shared neural encoders, enlarged-space state maps, entanglement-assisted Clifford encoders, irreversible dimension-reducing channels, learned disentangling unitaries, and tensor-network contraction maps. What unifies them is not a common formalism, but a common localization of the key mechanism at the encoder stage.

Setting	Encoder-level object	Role
Dual-output L2 ASR	Shared encoder with two task-specific decoders	Shared representations can become task-entangled
Embedding quantum simulator	$$\mathcal{M}:\mathbb{C}_{2^N}\rightarrow\mathbb{R}_{2^{N+1}}$$	Makes anti-linear entanglement expressions measurable
EA quantum coding	Encoder acting on ancillas, logical qubits, and halves of ebits	Shared entanglement changes stabilizer structure
Entanglement-assisted communication	$$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$	Compresses an entangled subsystem into a qubit
Disentangling quantum autoencoder	$U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$	Learns a product representation for a restricted family
Spin-network holographic code	$$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$	Makes link entanglement part of the map itself

The distinction between these usages is important. In some works, encoder-level entanglement is detrimental because it collapses task distinctions inside a shared representation. In others, it is the intended resource: an encoder is strengthened precisely because it has access to shared ebits, partially entangled links, or a larger local subsystem before compression. A common misconception is therefore to treat encoder-level entanglement as either uniformly desirable or uniformly pathological. The corpus instead supports a conditional reading: its value depends on whether the encoder must preserve task separation, expose an otherwise inaccessible observable, or exploit a pre-shared nonlocal resource (Cho et al., 4 Jun 2026, Candia et al., 2013, Guo et al., 2023, Sireesh et al., 25 Feb 2025, Qi et al., 15 Dec 2025).

2. Multitask representation entanglement in learned encoders

The most explicit representational use of the term appears in dual-output second-language ASR, where one acoustic input must support both a surface transcription $y^{\text{surf}}$, the verbatim pronounced form, and a meaning transcription $y^{\text{mean}}$, the intended canonical written form. The paper contrasts single-output models, in which each task has its own encoder, decoder, and auxiliary CTC head, with a dual-output model in which a single shared encoder feeds two task-specific Transformer decoders and an auxiliary CTC head trained on the surface targets. The central empirical finding is asymmetric: dual-output multitask learning improves meaning recognition but degrades surface transcription, and the degradation is much worse in English than in Korean. For Korean Conformer models, surface CER changes from $11.14$ to $11.34$ while meaning CER improves from $1.60$ to $0.77$; for English, surface CER degrades from $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$0 to $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$1 while meaning CER improves from $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$2 to $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$3 (Cho et al., 4 Jun 2026).

The paper links this asymmetry to divergence between the two target transcriptions, measured by Levenshtein edit distance, computed at the character-level syllable for Korean and at the word level for English. The data are stratified into $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$4, $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$5, $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$6, and $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$7. Korean shows small and inconsistent effects, whereas English shows a monotonic trade-off: the surface gap $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$8 worsens from $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$9 at $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$0 to $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$1 at $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$2, while the meaning gap improves from $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$3 to $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$4. Because the two languages have broadly similar divergence distributions, the cross-lingual difference is not explained away by a simple imbalance argument.

Mechanistically, the paper operationalizes encoder entanglement through layer-wise Centered Kernel Alignment. The key comparison is between independently trained single-output surface and meaning encoders. In Korean, $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$5 drops from $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$6 at layer $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$7 to $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$8 at layer $U|\psi\rangle=\bigotimes_{j=1}^N |\phi_j\rangle$9, then remains in the range $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$0 to $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$1, which the paper interprets as substantial task separation from layer $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$2 onward. In English, the same comparison remains high through most of the network: $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$3 at layer $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$4, $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$5 at layer $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$6, $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$7 at layer $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$8, and $$V_{\{n\}}:\mathcal{H}_{\mathrm{bulk}}\to\mathcal{H}_B$$9 at layer $y^{\text{surf}}$0, separating substantially only at layer $y^{\text{surf}}$1 with $y^{\text{surf}}$2. This is the paper’s evidence that the English encoder produces nearly identical task representations, whereas the Korean encoder preserves more distinct ones.

Decoder analysis sharpens the localization of the bottleneck. In Korean, same-task decoder CKA exceeds cross-task CKA at every reported layer, so the dual-output decoders remain aligned with their corresponding single-output decoders. In English, the deeper meaning decoder behaves differently: at decoder layer $y^{\text{surf}}$3, $y^{\text{surf}}$4, whereas $y^{\text{surf}}$5. The paper interprets this as evidence that the meaning decoder adapts with a unique representation, effectively bypassing the entangled encoder, while the surface decoder remains constrained by the frame-aligned encoder and the surface-oriented CTC interface. The resulting design implication is explicit: standard shared-encoder multitask learning is not enough for dual-output L2 ASR, and promising directions include sparse decomposition, adversarial training, and gating mechanisms (Cho et al., 4 Jun 2026).

3. Enlarged-space and source-level entanglement encoding

In embedding quantum simulators, encoder-level entanglement is not about harmful representation collapse but about choosing an encoding so that entanglement monotones become directly measurable. The simulated state $y^{\text{surf}}$6 is mapped to an enlarged real state $y^{\text{surf}}$7 by splitting each complex amplitude into real and imaginary parts. The exact relation is $y^{\text{surf}}$8 with $y^{\text{surf}}$9. In this representation, complex conjugation becomes the physical operator $y^{\text{mean}}$0, so expressions of the form $y^{\text{mean}}$1 are converted into $y^{\text{mean}}$2, determined by two measurable expectation values, $y^{\text{mean}}$3 and $y^{\text{mean}}$4. The method requires only one additional qubit, reduces concurrence from $y^{\text{mean}}$5 tomography observables to $y^{\text{mean}}$6 measurements, and reduces the $y^{\text{mean}}$7-tangle from $y^{\text{mean}}$8 observables to $y^{\text{mean}}$9 (Candia et al., 2013).

A different source-level use appears in a sequentially excited quantum three-level system, where a biexciton-exciton-ground ladder encodes entanglement directly into photon-number occupations over energy and time modes. Under two-photon resonant excitation with an early and a late pulse, the emitted state is

$11.14$0

which lives in a $11.14$1-dimensional Hilbert space when written as four binary modes $11.14$2. The paper describes this as energy- and time-entangled light encoded in the photon-number basis, and measures mode-resolved second-order correlations and a partial HOM-based coherence probe. It is explicit, however, that only a partial retrieval of the entanglement structure is achieved: diagonal mode-occupation structure is well supported, but full density-matrix reconstruction and definitive multipartite entanglement certification are left open (Vajner et al., 2024).

These two examples place the encoder at different physical locations. In the enlarged-Hilbert-space simulator, the encoder is a mathematical embedding chosen to linearize anti-linear entanglement expressions. In the quantum-dot source, the encoder is the sequential control protocol itself, which maps internal cascade history into a structured multimode Fock superposition. The shared theme is that entanglement is determined upstream by how the state is encoded, not only by a later decoding or tomography procedure.

4. Entanglement-assisted quantum code encoders

In entanglement-assisted quantum turbo coding, encoder-level entanglement refers to the use of shared ebits inside the seed transformation of a quantum convolutional encoder. The paper studies a Clifford unitary acting on memory qubits, information qubits, ancilla qubits, and $11.14$3 halves of ebits, and shows that an entanglement-assisted quantum convolutional encoder can be both recursive and non-catastrophic. This is precisely the pair of properties that the unassisted stabilizer setting cannot achieve simultaneously. The paper further proves that an encoder acting only on information qubits, classical bits, gauge qubits, and ancilla qubits cannot simultaneously satisfy both. Its explanation is structural: ebits alter the syndrome structure because all four Pauli errors on Alice’s ebit half are distinguishable by Bell measurement, reducing the hidden degeneracy that causes catastrophic error propagation (Wilde et al., 2010).

The same structural effect appears in entanglement-assisted stabilizer-code synthesis. For an $11.14$4 code, Alice begins with $11.14$5 ancillas, $11.14$6 halves of ebits, and $11.14$7 information qubits. On Alice’s side, the raw simplified check matrix has $11.14$8 generators rather than $11.14$9, because each ebit contributes a symplectic pair. The resulting encoding complexity is

$11.34$0

compared with

$11.34$1

for an ordinary $11.34$2 stabilizer code. The paper interprets the additional $11.34$3 term as the algebraic cost of shared entanglement: the benefits from shared entanglement come at an additional cost of encoding complexity (Kuo et al., 2019).

For structured EA quantum QC-LDPC codes, the practical bottleneck shifts from feasibility to circuit simplification. In the EA-CSS construction, the required ebit count is

$11.34$4

over $11.34$5, and the stabilizer description is extended to $11.34$6 and $11.34$7 so that $11.34$8. Starting from the Sharma-Kumar-Garani encoder, the paper extracts the CNOT subsequence as an invertible binary matrix and optimizes its row-operation decomposition with a beam search guided by the Hamming-distance heuristic $11.34$9. For the tested EA quantum QC-LDPC code families, this yields CNOT-count reductions of $1.60$0–$1.60$1 relative to the SKG baseline, lower CNOT counts than Patel-Markov-Hayes synthesis on all tested instances, and stabilizer-tableau verification of correctness (Sodhani et al., 9 Jun 2026).

A related measurement-based alternative replaces deep unitary encoder and decoder circuits with stabilizer projections, single-qubit measurements, and teleportation. Starting from physical Bell pairs, the protocol projects them into logical Bell states, decodes one side by single-qubit measurements, and prepares a physical-logical Bell resource state

$1.60$2

which is then consumed by teleportation to encode arbitrary states. For surface codes, the paper argues that the resulting encoding and decoding can be constant depth, surpassing the limits of conventional logarithmic-depth encoders, while fault tolerance is recovered by distilling the physical-logical Bell resource state itself (Shi et al., 2024).

5. Compression, disentangling, and encoder redesign

One line of work places the decisive operation at the encoder by making it explicitly irreversible. In an entanglement-assisted prepare-and-measure task with a qubit channel, Alice initially shares a $1.60$3-dimensional entangled state and applies a completely positive trace-preserving map

$1.60$4

to compress her subsystem to a qubit before transmission. In the explicit $1.60$5 construction, Alice holds two qubits $1.60$6, applies a two-qubit unitary $1.60$7, and discards $1.60$8:

$1.60$9

Under exact state discrimination, any coherence-preserving protocol based on qubit entanglement and unitary encoding is bounded by

$0.77$0

whereas the irreversible $0.77$1 strategy attains

$0.77$2

The reported experiment achieves $0.77$3, above the corrected benchmark $0.77$4. The paper is explicit that this is task-specific: for ordinary state discrimination alone, dense coding with $0.77$5 remains optimal (Guo et al., 2023).

A complementary redesign is the disentangling quantum autoencoder. Instead of compressing an $0.77$6-qubit entangled state to a smaller latent register, it seeks a learned unitary $0.77$7 such that, for a restricted family $0.77$8,

$0.77$9

The training objective is built from one-qubit purities

$$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$00

so successful encoding means that each encoded qubit is locally pure and the global encoded state is a product state. The paper argues that this gives an exponential improvement in the number of copies needed to transport entangled states across qubit-loss or leakage channels. It also reports unusually favorable generalization for some structured families: for states evolved with the transverse-field Ising Hamiltonian, one training datum is sufficient, whereas for stabilizer states the training-set size scales linearly with the number of qubits. At the same time, the paper is explicit that no single unitary can map the full $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$01-qubit state space into single-qubit product states and that generic hardware-efficient ansätze suffer barren plateaus (Sireesh et al., 25 Feb 2025).

Taken together, these works show two opposite but related encoder strategies. One uses entanglement plus controlled loss of coherence to gain communication power; the other removes entanglement at the encoder to gain robustness under loss and leakage. A plausible implication is that encoder-level entanglement is best regarded as a design variable rather than a fixed resource: some tasks demand deliberate preservation or activation of nonlocal structure, while others benefit from learning a basis in which that structure disappears.

6. Spectral, geometric, and constituent-level structural viewpoints

In monitored random quantum circuits viewed as encoders, entanglement is not exhausted by entropy. Starting from a product input, the circuit alternates local two-qubit unitaries with single-qubit projections inserted randomly with probability $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$02. For contiguous bipartitions, the familiar transition is from volume-law to area-law entanglement at $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$03. Using random bipartitions and entanglement-spectrum statistics, the paper identifies a second transition inside the area-law phase: at $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$04 for Haar-random circuits, the entanglement spectrum changes from a regime with residual level repulsion and non-universal statistics interpolating between Wigner-Dyson and Poisson, to a pure Poisson regime. Tensor-network compression interprets this second transition as a percolation transition of entangled bonds in circuit spacetime. The key encoder-level lesson is that measurements can destroy extensive entanglement before they destroy spectral entanglement structure (Zhang et al., 2020).

In spin-network holographic coding, entanglement is moved even further upstream: internal link states become part of the bulk-to-boundary map itself. Standard gluing corresponds to the maximally entangled singlet state on a link; the enriched construction allows each link to carry a variable amount of entanglement, labeled by a discrete variable $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$05, interpolating between product states and the fully entangled singlet. The effective bulk Hilbert space is spanned by tensor products of the corresponding link states, and contraction with fixed intertwiners defines

$$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$06

In a suitable large-spin regime, the map is a co-isometry in expectation value, and after restricting to a code subspace with links that are effectively glued or open, with small fluctuations around that pattern, the map becomes an exact isometry. The paper emphasizes that intermediate link entanglement not only tunes the amount of entanglement but also the extent to which gauge invariance at internal links is preserved or broken, admitting an interpretation in terms of emergent edge-mode-like degrees of freedom (Qi et al., 15 Dec 2025).

A constituent-level precursor to these structural views appears in symmetric pure states of $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$07 identical two-level atoms. There the collective pseudo-spin variances are decomposed into independent single-atom fluctuations plus residual correlation terms. For symmetric pure states, separability corresponds to

$$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$08

and the paper introduces

$$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$09

with $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$10 iff the state is unentangled and $$\Lambda_x:\mathcal{L}(\mathbb{C}^d)\to\mathcal{L}(\mathbb{C}^2)$$11 iff it is entangled. This is not an encoder in the architectural sense used by later papers. It nevertheless suggests a constituent-level template that recurs elsewhere: entanglement can be isolated by subtracting what would remain if the elementary units were independent, leaving only the residual correlation structure (Deb, 2011).

7. Common themes, misconceptions, and design implications

Several broad themes recur across these otherwise disparate literatures. First, encoder-level entanglement is frequently a statement about an upstream bottleneck. In dual-output ASR, the shared encoder can collapse task distinctions and force the meaning decoder, but not the surface decoder, to adapt downstream (Cho et al., 4 Jun 2026). In entanglement-assisted turbo coding, shared ebits change the local state diagram of the convolutional encoder itself, enabling a property pair that ancillas, classical bits, and gauge qubits cannot supply (Wilde et al., 2010). In measurement-based stabilizer encoding, entanglement is loaded into resource states so that the long unitary encoder is replaced by projection and teleportation (Shi et al., 2024).

Second, the literature repeatedly separates the existence of entanglement from its operational accessibility. Enlarged-space embeddings do not change the target monotone; they change whether the relevant anti-linear quantity is directly measurable (Candia et al., 2013). Sequentially excited three-level emitters generate a structured multimode entangled state, but the paper is explicit that only partial experimental retrieval of that structure is achieved (Vajner et al., 2024). Monitored-circuit work similarly shows that entanglement entropy can miss residual spectral organization that survives deeper into the area-law regime (Zhang et al., 2020).

Third, “more entanglement” is not the invariant design principle. Some encoders benefit from pre-shared ebits or higher-dimensional entangled resources; others are improved by deliberate disentangling into single-qubit product states. This suggests that the relevant variable is not entanglement magnitude alone, but whether the encoder realizes the right structured sharing for the task at hand. The design language used in the corpus makes this explicit: mitigate encoder-level entanglement in multitask L2 ASR (Cho et al., 4 Jun 2026), exploit irreversible compression in entanglement-assisted communication (Guo et al., 2023), optimize ebit-aware stabilizer encoders rather than removing ebits (Sodhani et al., 9 Jun 2026), or learn a basis in which a restricted entangled family becomes fully factorized (Sireesh et al., 25 Feb 2025).

A final misconception is to treat the term as a settled technical standard. The papers do not support that reading. Instead, they support a more precise editorial usage: encoder-level entanglement refers to situations in which the encoder architecture, resource state, or encoded representation geometry determines whether nonlocal correlations are harmful, measurable, compressible, suppressible, or code-defining. Under that usage, the term names a structural locus of control rather than a single cross-disciplinary quantity.