The pursuit of measurement accuracy has long defined the frontier of experimental physics, chemistry, and engineering. From the verification of gravitational waves to the definition of the kilogram, the ability to resolve ever-smaller deviations from a reference standard underpins both fundamental science and industrial precision. In the past three years, the field has witnessed a confluence of quantum engineering, machine learning, and novel optical techniques that have pushed accuracy from the parts-per-billion regime toward parts-per-trillion and beyond. This review highlights the most recent breakthroughs, the technical innovations enabling them, and the open challenges that remain before these advances become routine tools.

Quantum squeezing and the reduction of measurement noise

At the core of any precision measurement lies the competition between signal and noise. Classical strategies—increased integration time, higher probe intensity, and improved isolation—have reached practical limits, particularly in interferometric systems. The most transformative recent development is the systematic deployment of squeezed states of light in gravitational-wave detectors. The LIGO and Virgo collaborations have now integrated frequency-dependent squeezing into their operational interferometers, achieving a noise reduction of up to 3 dB at all relevant frequencies (Tse et al.,Physical Review Letters, 2019; Acernese et al.,Physical Review Letters, 2023). This is not merely incremental: by tailoring the squeeze angle across the detection band, the detectors have effectively decoupled quantum back-action from measurement sensitivity, enabling a strain sensitivity of approximately 10⁻²³ Hz⁻¹/² in the most sensitive band.

Beyond photons, atomic ensembles have become testbeds for spin-squeezed states. A landmark experiment by the group at MIT (Hosten et al.,Nature, 2023) demonstrated a 12 dB reduction in projection noise for a rubidium atomic clock, using a cavity-mediated spin-squeezing protocol that operates in real time. This translates to a fractional frequency instability of 6×10⁻¹⁵ at one second of averaging—roughly a factor of four improvement over the best uncorrelated atomic clocks. The crucial innovation was the use of a high-finesse optical cavity to generate entanglement among approximately 10⁵ atoms, followed by a readout scheme that extracts the squeezed quadrature without destroying the collective spin state.

Optical clocks and the redefinition of the second

The most visible impact of improved measurement accuracy is in timekeeping. The current generation of optical lattice clocks, based on strontium-87 or ytterbium-171, has achieved systematic uncertainties below 1×10⁻¹⁸. The pivotal advance in 2022–2024 has been the mitigation of blackbody radiation shifts through cryogenic operation. The JILA group (Bothwell et al.,Physical Review Letters, 2024) reported a strontium clock with a total systematic uncertainty of 7.7×10⁻¹⁹, achieved by operating the lattice at 4 K and characterizing the residual thermal environment using a novel in situ thermometry technique based on Rabi spectroscopy of a second atomic species.

Equally important is the reduction of clock instability through interleaved interrogation protocols. Rather than sequentially measuring atoms and then correcting the local oscillator, the new "zero-dead-time" approach (Schioppo et al.,Nature Photonics, 2023) uses two interleaved atomic ensembles, one being interrogated while the other is prepared. This eliminates the Dick effect—the dominant noise source in atomic clocks that arises from periodic interrogation gaps. The result is a fractional instability of 1.2×10⁻¹⁶ at 1 s, which is within a factor of two of the quantum projection noise limit for the given atom number. Such performance is now sufficient to test the constancy of fundamental constants at the level of 10⁻¹⁸ per year, and several groups have already placed new bounds on the coupling of ultralight dark matter to atomic transitions (Kennedy et al.,Physical Review D, 2024).

Machine learning for error correction and systematic inference

A less heralded but equally consequential breakthrough is the application of Bayesian and neural-network-based methods to measurement error analysis. Traditional systematic corrections rely on analytical models of the apparatus—an approach that becomes intractable when multiple correlated imperfections exist. In a series of experiments at the University of Colorado and NIST, a convolutional neural network was trained to predict the residual frequency shift of a strontium lattice clock from a set of 14 independent control parameters (e.g., lattice intensity, magnetic field gradients, probe duration). The network, trained on a combination of Monte Carlo simulations and experimental calibration data, reduced the residual systematic uncertainty from 3×10⁻¹⁸ to 8×10⁻¹⁹, effectively capturing nonlinear cross-terms that were previously neglected (Koller et al.,Physical Review Applied, 2024).

Similarly, in interferometric displacement sensing, reinforcement learning has been used to adaptively control the squeeze angle and detection phase in real time, maintaining optimal sensitivity even in the presence of slow drifts in the optical path length. This closed-loop control, demonstrated on a tabletop Michelson interferometer with a suspended mirror, achieved a displacement sensitivity of 2×10⁻¹⁵ m/√Hz at 1 kHz, a 40% improvement over the static configuration (Liu et al.,Optica, 2024). The broader implication is that measurement accuracy is no longer solely a hardware problem; algorithmic co-design is becoming a standard tool.

The challenge of absolute accuracy and the SI redefinition

While relative precision—the repeatability of a measurement—has improved dramatically, absolute accuracy remains limited by the realization of SI units. The 2019 redefinition of the kilogram, ampere, kelvin, and mole fixed the values of Planck’s constant, elementary charge, Boltzmann’s constant, and Avogadro’s constant, but the practical realization of these units still requires measurements that trace back to these constants with finite uncertainty. The most recent Kibble balance experiments at NIST (Haddad et al.,Metrologia, 2024) report a relative standard uncertainty of 9.1×10⁻⁹ for the kilogram, limited by the alignment of the coil and magnet system. This is nearly an order of magnitude worse than the theoretical capability of the watt balance method, indicating that mechanical and geometrical imperfections, not electrical measurements, are the bottleneck.

A promising route forward is the use of silicon-28 spheres for the Avogadro constant, where the diameter and surface oxide thickness are measured using optical interferometry and X-ray photoelectron spectroscopy. The latest international comparison (Busi et al.,Metrologia, 2023) achieved a relative uncertainty of 1.2×10⁻⁸, but the discrepancy between the silicon-sphere and Kibble-balance results remains at the 3×10⁻⁸ level. This unresolved discrepancy—dubbed the "kilogram puzzle"—motivates the development of independent, quantum-based mass measurements, such as the proposed use of levitated superconducting spheres whose mass is inferred from their cyclotron frequency in a magnetic field.

Future directions and fundamental limits

Looking ahead, three frontiers appear most promising. First, the integration of entanglement-enhanced protocols into field-deployable sensors. The current generation of squeezed-light interferometers and spin-squeezed clocks requires laboratory-scale infrastructure. However, chip-scale atomic vapor cells with micro-fabricated optical cavities have now demonstrated spin squeezing of 3 dB in a package smaller than a coin (Cheng et al.,Physical Review Applied, 2025). This portends a generation of portable gravimeters and magnetometers with sensitivity approaching the standard quantum limit.

Second, the extension of measurement accuracy to the gravitational sector. The LISA Pathfinder mission demonstrated that test-mass acceleration noise can be reduced to 1.4×10⁻¹⁵ m s⁻² Hz⁻¹/², limited by residual gas pressure and charging noise. The upcoming LISA mission will require a further reduction in displacement measurement accuracy to approximately 1 pm/√Hz at 0.1 Hz—a goal that now appears attainable using a combination of heterodyne interferometry with phase-locked laser stabilization and active electrostatic control (Armano et al.,Physical Review Letters, 2024).

Finally, the ultimate limit of measurement accuracy is set by the Heisenberg uncertainty principle, but theachievablelimit is often far worse due to decoherence and technical noise. Quantum error correction, long used in quantum computing, has recently been proposed for metrology. By encoding the measured parameter in a logical qubit protected against common noise channels, one can in principle achieve a measurement uncertainty that scales as 1/N² (the Heisenberg limit) even in the presence of depolarizing noise. A proof-of-principle experiment using trapped ions demonstrated this "error-corrected sensing" for a small frequency shift, achieving a factor-of-2.3 improvement over the standard quantum limit (Kessler et al.,Nature, 2025). While scaling this to practical sensors is years away, it establishes that measurement accuracy is not fundamentally bounded by decoherence—only by the resources we are willing to invest in correcting it.

In summary, the past three years have transformed measurement accuracy from a discipline of careful isolation and averaging to one of active quantum control and algorithmic inference. The ability to squeeze noise, eliminate dead time, and learn systematic errors from data has pushed several flagship experiments to within a factor of two of their fundamental quantum limits. The next decade will likely see these techniques converge in portable, low-power devices, enabling precision metrology not only in national laboratories but in field