diff --git a/src/gen/inference/importance.cljc b/src/gen/inference/importance.cljc
index 1fb04bb..208db94 100644
--- a/src/gen/inference/importance.cljc
+++ b/src/gen/inference/importance.cljc
@@ -1,39 +1,87 @@
 (ns gen.inference.importance
-  (:require [clojure.math :as math]
+  (:require [gen.choicemap :as choicemap]
             [gen.distribution.kixi :as dist]
             [gen.generative-function :as gf]))
 
 ;; This implementation comes from `fastmath.core`, ported here for cljc
 ;; purposes.
 
-(defn- logsumexp
-  "log(exp(x1)+...+exp(xn))."
-  ^double [xs]
-  (loop [[^double x & rst] xs
-         r 0.0
-         alpha ##-Inf]
-    (if (<= x alpha)
-      (let [nr (+ r (math/exp (- x alpha)))]
-        (if-not (seq rst)
-          (+ (math/log nr) alpha)
-          (recur rst nr alpha)))
-      (let [nr (inc (* r (math/exp (- alpha x))))]
-        (if-not (seq rst)
-          (+ (math/log nr) x)
-          (recur rst nr (double x)))))))
+(defn ^:nodoc logsumexp
+  "log(exp(x1)+...+exp(xn)). This version can handle infinities."
+  ([x] x)
+  ([x y]
+   (if (<= y x)
+     (+ x (Math/log (inc (Math/exp (- y x)))))
+     (+ y (Math/log (inc (Math/exp (- x y)))))))
+  ([x y & more]
+   (reduce logsumexp (logsumexp x y) more)))
 
-(defn resampling [gf args observations n-samples]
-  ;; https://github.com/probcomp/Gen.jl/blob/master/src/inference/importance.jl#L77...L95
-  (let [result (gf/generate gf args observations)
-        model-trace (volatile! (:trace result))
-        log-total-weight (volatile! (:weight result))]
-    (dotimes [_ (dec n-samples)]
-      (let [candidate (gf/generate gf args observations)
-            candidate-model-trace (:trace candidate)
-            log-weight (:weight candidate)]
-        (vswap! log-total-weight #(logsumexp [log-weight %]))
-        (when (dist/bernoulli (math/exp (- log-weight @log-total-weight)))
-          (vreset! model-trace candidate-model-trace))))
-    (let [log-ml-estimate (- @log-total-weight (math/log n-samples))]
-      {:trace @model-trace
-       :weight log-ml-estimate})))
+(defn- proposal-fn
+  ([gen-fn args constraints]
+   (if (empty? constraints)
+     (fn []
+       (gf/propose gen-fn args))
+     (let [constraints (choicemap/choicemap constraints)]
+       (fn []
+         (choicemap/merge
+          constraints
+          (gf/propose gen-fn args)))))))
+
+(defn sampling
+  ([gf args {:keys [observations n-samples proposal proposal-args]
+             :or   {proposal-args []
+                    observations  choicemap/EMPTY}}]
+   ;; TODO do this so it returns a lazy-seq of samples.
+   )
+  ([gf args observations n-samples]
+   (sampling gf args {:observations observations
+                      :n-samples    n-samples}))
+  ([gf args observations proposal proposal-args n-samples]
+   (sampling gf args {:observations  observations
+                      :n-samples     n-samples
+                      :proposal      proposal
+                      :proposal-args proposal-args})))
+
+(defn resampling
+  "https://github.com/probcomp/Gen.jl/blob/master/src/inference/importance.jl#L77...L95
+
+  Run sampling importance resampling, returning a single trace.
+
+  TODO do we actually want a sequence of traces encountered, so that we can
+  visualize the path?"
+  ([gf args {:keys [observations n-samples proposal proposal-args]
+             :or   {proposal-args []
+                    observations  choicemap/EMPTY
+                    n-samples     10}}]
+   (let [constraint-fn (if proposal
+                         (proposal-fn proposal proposal-args observations)
+                         (fn [] {:choices observations
+                                :weight  0.0}))
+         {proposal-choices :choices
+          proposal-weight :weight} (constraint-fn)
+         result        (gf/generate gf args proposal-choices)]
+     (loop [i                (dec n-samples)
+            model-trace      (:trace result)
+            log-total-weight (- (:weight result) proposal-weight)]
+       (if (zero? i)
+         (let [log-ml-estimate (- log-total-weight (Math/log n-samples))]
+           {:trace  model-trace
+            :weight log-ml-estimate})
+         (let [{proposal-choices :choices
+                proposal-weight :weight} (constraint-fn)
+               candidate       (gf/generate gf args proposal-choices)
+               candidate-trace (:trace candidate)
+               log-weight      (- (:weight candidate) proposal-weight)
+               new-weight      (logsumexp log-total-weight log-weight)
+               new-trace       (if (dist/bernoulli (Math/exp (- log-weight new-weight)))
+                                 candidate-trace
+                                 model-trace)]
+           (recur (unchecked-dec-int i) new-trace new-weight))))))
+  ([gf args observations n-samples]
+   (resampling gf args {:observations observations
+                        :n-samples    n-samples}))
+  ([gf args observations proposal proposal-args n-samples]
+   (resampling gf args {:observations  observations
+                        :n-samples     n-samples
+                        :proposal      proposal
+                        :proposal-args proposal-args})))