1: \begin{abstract}

2: Bayesian phylogenetic inference is often conducted via local or sequential search over topologies and branch lengths using algorithms such as random-walk Markov chain Monte Carlo (\textsc{Mcmc}) or Combinatorial Sequential Monte Carlo (\textsc{Csmc}). %These methods sample tree topologies and branch lengths, 

3: However, when \textsc{Mcmc} is used for evolutionary parameter learning, convergence requires long runs with inefficient exploration of the state space. We introduce Variational Combinatorial Sequential Monte Carlo (\textsc{Vcsmc}), % a novel Variational Inference method that simultaneously performs both parameter inference and model learning. \textsc{Vcsmc} is 

4: a powerful framework that establishes 

5: variational sequential search to learn distributions over intricate combinatorial structures.

6: %\textsc{Csmc} to construct a variational objective defined on the composite space of phylogenetic trees. 

7: We then develop nested \textsc{Csmc}, an efficient proposal distribution for \textsc{Csmc} and prove that nested \textsc{Csmc} is an exact approximation to the (intractable) locally optimal proposal. We use nested \textsc{Csmc} to define a second objective, \textsc{Vncsmc} which yields tighter lower bounds than \textsc{Vcsmc}. %. 

8: We show that \textsc{Vcsmc} and \textsc{Vncsmc} are computationally efficient and 

9: explore higher probability spaces than existing methods on a range of tasks.

10: \end{abstract}

11: