Adaptive Quantum Computers

• Adaptive quantum computers are systems that dynamically integrate quantum operations with classical feedback, modifying circuits based on real-time measurements and hardware constraints.
• They employ techniques such as hardware-aware compilation, adaptive ansatz growth, and measurement-driven control to optimize quantum state preparation and error mitigation.
• Their architecture shifts part of the computational burden to classical processing, enabling compressed quantum depth and improved performance across a variety of quantum tasks.

Adaptive quantum computers are quantum computing systems in which the quantum process is not treated as a fixed, closed circuit, but is instead modified by intermediate classical computation, hardware-aware compilation, task-dependent ansatz growth, or feedback derived from measurements and device characterization. In the most explicit formalization, they are quantum computers that interact with a standard computer during the computation; in broader usage, the term also encompasses compilers, intermediate representations, and quantum-native modules that adapt to measurements, hardware constraints, and problem structure (Neumann, 10 Sep 2025, Sim et al., 2018). Across current literature, adaptivity appears at several layers simultaneously: alternating quantum/classical execution, state preparation and variational simulation, compilation and partitioning, noise estimation and control, and measurement-driven many-body dynamics (Neumann, 8 Sep 2025, Budiutama et al., 20 Aug 2025).

1. Conceptual scope and defining features

Adaptive quantum computation is defined most directly by the presence of an explicit interaction between quantum and classical computation during execution. The formal model introduced for this purpose is

$\mathrm{LAQCC}(\mathcal Q,\mathcal C,d),$

where quantum layers $Q\in\mathcal Q$ alternate with classical layers $C\in\mathcal C$ for $d$ rounds; after each quantum layer, measurements feed a classical computation that may control subsequent quantum operations. The central instance studied is

$\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$

namely constant-depth quantum layers, logarithmic-depth classical layers, and a constant number of alternations (Neumann, 10 Sep 2025).

A narrower but closely related definition appears in work on adaptive quantum algorithms for realistic hardware, where the essential mechanism is the alternation of quantum steps, measurement, classical processing, and conditionally chosen later operations. In that setting, the classical computation is no longer limited to syndrome processing for error correction; it is repurposed to drive the computation itself (Neumann, 8 Sep 2025). This makes adaptive quantum computers distinct from both purely unitary circuits and generic hybrid workflows in which the classical side merely optimizes parameters offline.

The literature also uses “adaptive” in broader systems and representation-theoretic senses. Compiler work on qudits argues that the compilation strategy should adapt to the hardware’s actual level structure, connectivity, and calibration cost rather than impose a rigid qubit-style decomposition (Mato et al., 2022). Quantum machine-learning work introduces trainable unitary layers that adaptively interpolate between known transforms, rather than learning unconstrained deep variational circuits from scratch (Budiutama et al., 20 Aug 2025). Variational simulation and quantum chemistry papers use “adaptive” for ansätze or basis sets that are grown or reparameterized according to the target problem, rather than fixed a priori (Grimsley et al., 2018, Kwon et al., 2022).

A recurring practical assumption is bounded alternation or fixed-step termination. This appears explicitly in the LAQCC formalization, which focuses on computations that terminate after a fixed number of steps because such protocols are easier to implement on current hardware (Neumann, 10 Sep 2025). A similar boundedness condition underlies many hardware demonstrations of mid-circuit feedback, adaptive compilation, and dynamic refresh, where classical latency, measurement overhead, and storage time remain first-order constraints (Zhang et al., 23 Apr 2025, Pokharel et al., 22 Sep 2025).

2. Formal models, computational claims, and state-preparation advantages

The strongest theoretical claims for adaptive quantum computers concern computational power under restricted depth and noise-sensitive state preparation. In the decoding setting, adaptive quantum computation was shown to outperform standard computation for retrieving information from corrupted digital data. The proof uses a structure-versus-randomness decomposition: low-rank or low-analytic-rank decoders are too structured to recover arbitrary messages reliably, while high-rank decoders behave pseudorandomly under biased noise. Within this framework, constant-depth classical decoders in $NC^0[\oplus]$ fail with probability bounded away from zero for sufficiently large message length, whereas adaptive quantum constructions achieve separations for Hadamard-code-related tasks and stronger noisy list-decoding regimes (Neumann, 10 Sep 2025).

The same formal program yields adaptive state-preparation procedures for several canonical states. Uniform superposition over $q$ states can be prepared with $O((\log_2 q)^2)$ qubits; the GHZ state can be prepared in constant depth using $2n-1$ qubits; the $W$-state admits a constant-depth adaptive construction using $Q\in\mathcal Q$0 qubits; and Dicke-state preparation is obtained either in constant depth for small $Q\in\mathcal Q$1 or with $Q\in\mathcal Q$2 quantum/classical alternations for general $Q\in\mathcal Q$3 (Neumann, 10 Sep 2025). These results articulate a central theme of adaptive quantum computing: quantum depth can be compressed if classical feedforward, additional workspace, and measurement-conditioned correction are allowed.

A more hardware-oriented analysis compares adaptive and non-adaptive GHZ and $Q\in\mathcal Q$4-state preparation under a worst-case error model in which every error is a Haar-random unitary. In that model, adaptive GHZ preparation uses constant quantum depth and $Q\in\mathcal Q$5 qubits, and can be exponentially better than standard all-to-all or linear protocols when two-qubit gates are not substantially more error-prone than prolonged idling. For GHZ preparation, the advantage condition is expressed asymptotically as $Q\in\mathcal Q$6 against all-to-all circuits and $Q\in\mathcal Q$7 against linear nearest-neighbor circuits; for $Q\in\mathcal Q$8-states, the corresponding condition is $Q\in\mathcal Q$9 (Neumann, 8 Sep 2025).

Theoretical advantage does not imply present-day superiority. On IBM Brisbane, the worst-case model predicted about a $C\in\mathcal C$0 success-probability advantage for adaptive GHZ preparation at $C\in\mathcal C$1, with estimated runtimes

$C\in\mathcal C$2

Empirically, however, the adaptive protocol did not outperform the full non-adaptive protocol: for small $C\in\mathcal C$3, the standard method slightly outperformed the adaptive one, and for larger $C\in\mathcal C$4 both degraded substantially (Neumann, 8 Sep 2025). This is one of the clearest demonstrations that adaptive quantum computing is not synonymous with immediate hardware advantage; its gains depend on the relative costs of idling, measurement, classical feedback latency, routing, and gate decomposition.

3. Compilation, intermediate representations, and the adaptive systems stack

At the compilation level, adaptivity means that the mapping from logical operations to hardware operations is optimized against the actual physical substrate rather than a fixed textbook decomposition. For qudit systems, adaptive compilation is built on an energy coupling graph whose nodes are physical energy levels and whose edges encode allowed or efficient transitions. The target unitary is decomposed as

$C\in\mathcal C$5

with each $C\in\mathcal C$6 a two-level rotation and $C\in\mathcal C$7 a diagonal phase matrix. The compiler performs an adaptive depth-first-search/backtracking exploration over alternative Givens-rotation eliminations, routes states dynamically through the graph, uses ancillary levels when beneficial, and propagates phase shifts virtually. Relative to fixed QR-style decomposition on random Clifford unitaries in dimensions $C\in\mathcal C$8, $C\in\mathcal C$9, and $d$0, the average cost is significantly lower, the maximum cost is always lower, improvements can reach a factor of $d$1, and runtime per unitary remained under $d$2 second in the reported evaluation (Mato et al., 2022).

A distinct line of work formulates adaptive circuit adaptation as a satisfiability modulo theories problem. The circuit is partitioned into two-qubit blocks, candidate substitutions and decompositions are precomputed, and the solver chooses a globally compatible set of substitutions using Boolean variables $d$3, block start times $d$4, durations $d$5, and log-fidelity variables $d$6. The target platform is semiconducting spins, where distinct native two-qubit gates—SWAP, CPHASE/CZ, and CROT/CNOT-like realizations—have different duration and fidelity profiles. On a noisy simulator, this global adaptation improved Hellinger fidelity by up to $d$7 and reduced qubit idle time by up to $d$8 relative to alternative adaptation techniques (Brandhofer et al., 2023). The significance is not merely better decomposition quality, but the explicit treatment of idle-time decoherence and operational-mode selection as first-class compilation objectives.

Photonic one-way quantum computing requires a different adaptive layer: the intermediate representation itself must encode hardware resource bounds. OneAdapt introduces an IR that constrains temporal-edge lengths to an adaptive limit $d$9, allows skewed temporal edges within Hamming distance $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$0, and uses dynamic refreshing plus 2D-bounded temporal routing to balance 1D depth, 2D size, fusion-device demand, and measurement wait time. Restricting temporal edges to $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$1 implies that only $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$2 fusion devices per chiplet are needed; with PL ratio fixed at $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$3, the longest storage time is $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$4 RSLs. When integrated with fault-tolerant scheduling based on surface-code logical blocks, dynamic refresh reduced consumed RSL depth by $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$5 on average relative to a static periodic strategy (Zhang et al., 23 Apr 2025).

Adaptive circuits also alter partitioning and distributed execution. A hypergraphic representation has been proposed in which conditional gates, measurements, classical control dependencies, and grouped classical-operation sections are encoded as hyperedges or super-hyperedges. Partitioning then uses an FM-style gain

$\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$6

with movement decisions modified to preserve measurement-conditioned structure and reduce communication overhead across QPUs (Cambiucci et al., 12 Apr 2025). At the workflow level, the algo2qpu framework treats adaptive hybrid quantum-classical algorithms as end-to-end deployment problems: choose an algorithm, instantiate it, compile to device-compatible gates, execute through cloud services, postprocess measurements, and re-enter a classical feedback loop for parameter updates until convergence (Sim et al., 2018). This suggests that adaptivity is not confined to the circuit semantics; it extends to the orchestration layer of the quantum software stack.

4. Adaptive ansätze, transforms, and problem-tailored quantum representations

A large part of the adaptive-quantum-computing literature concerns replacing fixed circuit templates with representations that grow or interpolate according to the target problem. The canonical example is ADAPT-VQE, in which the ansatz is assembled one operator at a time from an operator pool by choosing the generator with the largest energy-gradient magnitude at each iteration. The stopping criterion is the $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$7 norm of the gradient vector, $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$8, with thresholds $\mathrm{LAQCC}=\mathrm{LAQCC}(QNC^0,NC^1,O(1)),$9. In numerical studies on LiH, BeH$NC^0[\oplus]$0, and linear H$NC^0[\oplus]$1, ADAPT-VQE used far fewer parameters than UCCSD and achieved average PES errors of $NC^0[\oplus]$2, $NC^0[\oplus]$3, and $NC^0[\oplus]$4 kcal/mol, respectively, for $NC^0[\oplus]$5, whereas UCCSD gave $NC^0[\oplus]$6, $NC^0[\oplus]$7, and $NC^0[\oplus]$8 kcal/mol (Grimsley et al., 2018). The key point is that the ansatz is quasi-optimal with respect to the molecule rather than tied to a fixed excitation truncation.

The same adaptive logic has been extended to post-Born–Oppenheimer simulation in the nuclear-electronic orbital framework. NEO-FNO-ADAPT-VQE combines adaptive ansatz growth with frozen natural orbital truncation. Electronic and protonic virtual orbitals with occupations below $NC^0[\oplus]$9 and $q$0, respectively, are discarded, shrinking the H$q$1/D$q$2 problem from $q$3 electronic and $q$4 nuclear spin orbitals to $q$5 electronic and $q$6 nuclear spin orbitals. After parity mapping and $q$7 tapering, the Hamiltonian becomes a $q$8-qubit problem. Relative to NEO-FNO-UCCSD with about $q$9–$O((\log_2 q)^2)$0 CNOTs, the fermionic and qubit ADAPT variants reduce the CNOT count to $O((\log_2 q)^2)$1 and $O((\log_2 q)^2)$2, respectively, while staying within $O((\log_2 q)^2)$3 mHartree of the NEO-FNO-FCI reference (Nykänen et al., 2023). This is a concrete instance of adaptive compression of both the orbital space and the circuit.

Adaptivity can also be embedded directly into a trainable quantum layer. The Adaptive Interpolating Quantum Transform defines a unitary $O((\log_2 q)^2)$4 satisfying

$O((\log_2 q)^2)$5

so that learning is confined to a structured subspace of unitaries rather than the full unitary group. In the principal construction, AIQT interpolates between the Hadamard transform and the quantum Fourier transform by replacing each QFT controlled rotation with

$O((\log_2 q)^2)$6

At $O((\log_2 q)^2)$7, the controlled phases reduce to identities and the transform becomes Hadamard-like; at $O((\log_2 q)^2)$8, the exact QFT is recovered. Used as a preprocessing layer ahead of a shallow QNN, AIQT-QNN achieved the lowest final loss, faster convergence, sharper class-probability transitions near phase boundaries, and better scaling with qubit count than QFT-QNN and a baseline QNN on a $O((\log_2 q)^2)$9-qubit quantum phase-classification task. A TFIM-based AIQT variant achieved lower validation loss than the QFT-based AIQT, slightly higher validation accuracy, and exceeded $2n-1$0 classification accuracy, indicating that the best interpolation depends on task structure (Budiutama et al., 20 Aug 2025).

Other forms of problem-tailored adaptation operate on basis design and sampled subspaces rather than gate pools. Geometry-dependent adaptive basis sets for H$2n-1$1 optimize Gaussian exponents and contraction coefficients as functions of the H–H distance while keeping the basis size fixed, allowing A-STO-3G$2n-1$2 to approach double-zeta quality without increasing qubit count (Kwon et al., 2022). A basis-adaptive algorithm for the XXZ chain starts from a small set of classical basis states, evolves them for a single Trotter step with $2n-1$3, samples new bitstrings, filters them by $2n-1$4 and reflection symmetry, and diagonalizes the Hamiltonian in the resulting reduced space. On IBM Heron, it reached sub-percent ground-state energy error up to $2n-1$5 qubits, with reported errors of $2n-1$6, $2n-1$7, and $2n-1$8 at $2n-1$9, $W$0, and $W$1, outperforming SKQD in the tested regime (Biswas et al., 14 Dec 2025). For one-dimensional many-body states, normal MPS preparation has likewise been made adaptive through ADAPT-AQC and improved AQC-Tensor initialization, enabling a $W$2-site XXZ quench study on hardware with quench circuits of CZ depth $W$3 and $W$4, compared with substantially deeper staircase-style preparations (Jaderberg et al., 12 Mar 2025).

5. Estimation, feedback, and adaptive control of noisy quantum dynamics

Adaptive quantum computers are also characterized by closed-loop estimation and control. One line of work treats NISQ noise as a time-varying process whose effective channel parameters should be inferred from the circuit’s own measurement outputs rather than assumed from stale calibration data. In this model,

$W$5

with error parameter $W$6 inferred from binary outcomes $W$7 through Bayesian updating,

$W$8

For a $W$9-qubit uniform-superposition circuit, the method estimated a $Q\in\mathcal Q$00-parameter error vector from $Q\in\mathcal Q$01 shots per execution and used the MAP estimate to perform readout mitigation and coherent angle compensation. Adaptive mitigation reduced the Hellinger distance to the ideal distribution relative to the unmitigated case, whereas static mitigation could become worse when device noise drifted away from the assumed calibration (Dasgupta et al., 2022). The principal limitation identified there is scalability: inference becomes statistically inefficient if the channel model is too high-dimensional.

A second estimation-and-control framework, Noise Mapping for Quantum Architectures, uses qubits as sensors to construct spatial noise maps. NMQA employs a two-layer particle filter: $Q\in\mathcal Q$02-particles represent hypotheses for the spatial field, while $Q\in\mathcal Q$03-particles represent neighborhood sizes that determine which qubits share a common noise process. An adaptive controller then schedules the next measurement in the region of greatest posterior uncertainty. On simulated spatially inhomogeneous fields, NMQA reduced the number of measurements required to reach a target error metric by up to $Q\in\mathcal Q$04; on trapped-$Q\in\mathcal Q$05 experiments, the reduction was up to $Q\in\mathcal Q$06 (Gupta et al., 2019). This places adaptive quantum computing in continuity with autonomous calibration and “quantum firmware,” where the machine actively learns its own noise landscape.

The most explicit hardware realization of adaptive control as computation is the recent work on adaptive monitored quantum circuits. There, local mid-circuit measurements and resets are combined with conditional feedback on IBM’s ibm_fez superconducting processor to implement a non-unitary Bernoulli-map-like dynamics on systems of $Q\in\mathcal Q$07 to $Q\in\mathcal Q$08 qubits. Each step applies either a scrambling two-qubit gate with probability $Q\in\mathcal Q$09 or a control/reset operation with probability $Q\in\mathcal Q$10, with measurement-conditioned $Q\in\mathcal Q$11 gates resetting qubits to $Q\in\mathcal Q$12. The experiment applied nearly $Q\in\mathcal Q$13 entangling gates and nearly $Q\in\mathcal Q$14 non-unitary mid-circuit operations, and observed a quantum-to-classical dynamical phase transition with entanglement transition near $Q\in\mathcal Q$15 and control transition near $Q\in\mathcal Q$16. Finite-size scaling gave $Q\in\mathcal Q$17 and $Q\in\mathcal Q$18 for the magnetization-based transition, while dynamical criticality was consistent with $Q\in\mathcal Q$19 (Pokharel et al., 22 Sep 2025). In this setting, measurement is not a terminal readout but an active resource that competes with scrambling to stabilize a fixed point.

6. Empirical status, limitations, and research trajectory

The contemporary literature presents adaptive quantum computers as promising but unevenly realized. On the positive side, adaptive compilation has already shown hardware-relevant gains in qudits, semiconducting spins, and photonic MBQC (Mato et al., 2022, Brandhofer et al., 2023, Zhang et al., 23 Apr 2025). Adaptive ansätze and representations have delivered substantial circuit compression in chemistry, many-body simulation, and quantum machine learning (Grimsley et al., 2018, Nykänen et al., 2023, Budiutama et al., 20 Aug 2025). Mid-circuit measurement plus reset has been demonstrated at a scale—up to $Q\in\mathcal Q$20 qubits and roughly $Q\in\mathcal Q$21 total operations—that begins to resemble the operational structure required for fault tolerance (Pokharel et al., 22 Sep 2025).

At the same time, several recurring limitations delimit present performance. Adaptive algorithms frequently reduce quantum depth only by increasing measurement count, qubit count, classical control complexity, or routing overhead (Neumann, 8 Sep 2025, Neumann, 10 Sep 2025). The SMT-based adaptation framework relies on precomputed substitution costs and is restricted to two-qubit blocks, which keeps the optimization tractable but narrows the search space (Brandhofer et al., 2023). Adaptive Bayesian mitigation is explicitly described as facing an open scaling problem once the channel model grows beyond compact structured parameterizations (Dasgupta et al., 2022). The basis-adaptive XXZ algorithm is benchmarked mainly in symmetry-rich regimes and depends on physically motivated initial states (Biswas et al., 14 Dec 2025). Noise mapping presupposes spatially correlated noise processes from which local information sharing is beneficial (Gupta et al., 2019).

A common misconception is that adaptivity is equivalent to generic variational optimization. The record is broader. Some adaptive quantum computers are defined by alternating classical and quantum layers with explicit feedforward (Neumann, 10 Sep 2025); others by hardware-aware compilation and IR design (Mato et al., 2022, Zhang et al., 23 Apr 2025); others by trainable quantum-native transforms (Budiutama et al., 20 Aug 2025); and others by autonomous characterization and real-time control (Gupta et al., 2019, Pokharel et al., 22 Sep 2025). Another misconception is that adaptivity always improves practical performance. Theoretical separations and asymptotic state-preparation gains do not yet imply robust hardware wins across the board; the GHZ study on IBM Brisbane is a direct counterexample, where adaptive protocols were theoretically favored but did not empirically surpass the standard implementation (Neumann, 8 Sep 2025).

Taken together, the evidence suggests that adaptive quantum computers are best understood as a multi-layer architecture rather than a single algorithmic trick. The unifying principle is the relocation of part of the computational burden from a fixed quantum circuit into dynamic choices made by classical processing, structured parameterizations, or hardware-aware control. A plausible implication is that scalable quantum systems will rely less on brute-force deep variational circuits or static compilation pipelines, and more on bounded-depth quantum primitives embedded within adaptive loops for synthesis, routing, estimation, and stabilization (Sim et al., 2018, Budiutama et al., 20 Aug 2025). The present literature supports that architectural shift, while also showing that its practical advantage remains contingent on control-stack maturity, measurement performance, and the cost of classical intervention.