Quantum parameter estimation has many applications, from gravitational wave detection to quantum key distribution. We present the first experimental demonstration of the time-symmetric technique of quantum smoothing. We consider both adaptive and non-adaptive quantum smoothing, and show that both are better than their well-known time-asymmetric counterparts (quantum filtering). For the problem of estimating a stochastically varying phase shift on a coherent beam, our theory predicts that adaptive quantum smoothing (the best scheme) gives an estimate with a mean-square error up to 2 √ 2 times smaller than that from non-adaptive quantum filtering (the standard quantum limit). The experimentally measured improvement is 2.24 ± 0.14.PACS numbers: 42.50. Dv, 42.50.Xa, 03.65.Ta, 06.90.+v Quantum parameter estimation (QPE) is the problem of estimating an unknown classical parameter (or process) which plays a role in the preparation (or dynamics) of a quantum system [1,2], and is central to many fields including gravitational wave interferometry [5], quantum computing [3], and quantum key distribution [4]. The fundamental limit to the precision of the estimate in QPE is set by quantum mechanics [1,2]. Thus one of the key issues in QPE is the development of practical methodologies which allow measurements to approach or exceed the standard quantum limit (SQL) for a given measurement coupling [6,7,8,9,10,11,12]. Because of its wide-ranging technological relevance, the prime example of QPE is estimating an optical phase shift [13,14,15,16,17,18,19,20].Apart from some theoretical papers [19,20], work in this area of QPE has concentrated upon the problem of estimating a fixed, but unknown phase shift, which can be thought of as preparing the quantum state with an average phase equal to this parameter. It was shown theoretically [15] that for this problem adaptive homodyne measurements coupled with an optimal estimation filter can yield an estimate with mean-square error smaller than the standard quantum limit (as set by perfect heterodyne detection). This was demonstrated experimentally in Ref.[16] using very weak coherent states (for which the factor of improvement is at most 2). More recent theory and experiment have shown that interferometric measurements with photon counting can also be improved using adaptive techniques [17,18].A far richer, and in many cases more experimentally relevant, problem of quantum phase estimation arises when the phase evolves dynamically under the influence of an unknown classical stochastic process [19,20]. The general problem of estimating a classical process dynamically coupled to a quantum system under continuous measurement has recently been considered by Tsang [21], who introduced three main categories of quantum estimation: prediction or filtering, smoothing, and retrodiction. Of those, prediction or filtering is a causal estimation technique that can be used in real-time applications [24]. Smoothing and retrodiction are acausal and so cannot be used in real time, but they can be use...