The present study aims at investigating advanced subset simulation techniques, which are based on the theory of particle filter, for the assessment of the failure probability of a marine structure under extreme loading conditions. Three approaches are considered, namely the classical particle filter method, the subset simulation with a branching process and one using the minimum values of the samples as levels. They are, first, intensively applied on a simple example for which a known analytical solution is available, in order to investigate their parameter settings. Then, they are applied, with good performance, using their respective best parameter settings, to the assessment of failure probability of a FPSO subjected to extreme roll motion. INTRODUCTION The design of a marine structure requires to estimate the probability that the structure would fail under extreme load conditions. That issue is of great importance for the roll motion, which can significantly impacts the safety and the performance in operational conditions. It leads to the computation of the probability that the structural response, exceeds a given critical threshold for a given reference time (e.g. the timescale of a sea state, one year, ...): c F L X P Prob (1) where (X) denotes the structural response and Lc the critical threshold. Under some restrictive assumptions, namely when the loading is a random Gaussian process and the structural response, , is linear or quadratic, closed-form solutions or good analytical approximation are available. But in general, the structural response results from complex non-linear dynamic equations involving time consuming time-domain computations and the representation of the dynamic loadings as random processes involves a high number of random variables. In that case, there is no simple analytical solution and, for an accurate assessment of the failure probability, simulation-based reliability methods are more appropriate, but become unpractical as the number of simulations increases dramatically when low failure probabilities are estimated. An appealing way to reduce the number of simulations, required for small probabilities, is to use the subset simulation method. That method expresses the small failure probability, in virtue of the Bayes theorem, as the product of conditional probabilities, which are not so small and can be estimated with a reasonable number of simulations.