diff --git a/paper/figure3_performance/SF_calc_comparison/results_20260609_090622/figure3_sf_comparison_cpu.png b/paper/figure3_performance/SF_calc_comparison/results_20260609_090622/figure3_sf_comparison_cpu.png
new file mode 100644
index 00000000..7424a333
Binary files /dev/null and b/paper/figure3_performance/SF_calc_comparison/results_20260609_090622/figure3_sf_comparison_cpu.png differ
diff --git a/paper/figure3_performance/SF_calc_comparison/results_20260609_090622/figure3_sf_comparison_gpu.png b/paper/figure3_performance/SF_calc_comparison/results_20260609_090622/figure3_sf_comparison_gpu.png
new file mode 100644
index 00000000..0b837729
Binary files /dev/null and b/paper/figure3_performance/SF_calc_comparison/results_20260609_090622/figure3_sf_comparison_gpu.png differ
diff --git a/paper/figure3_performance/data/refinement_cycle/results_20260618_141803/summary.csv b/paper/figure3_performance/data/refinement_cycle/results_20260618_141803/summary.csv
new file mode 100644
index 00000000..db15a6fe
--- /dev/null
+++ b/paper/figure3_performance/data/refinement_cycle/results_20260618_141803/summary.csv
@@ -0,0 +1,7 @@
+device,n_threads,agg_fwd_no_grad_mean,agg_fwd_no_grad_min,agg_fwd_no_grad_max,agg_fwd_no_grad_speedup,agg_fwd_graph_mean,agg_fwd_graph_min,agg_fwd_graph_max,agg_fwd_graph_speedup,agg_bwd_only_mean,agg_bwd_only_min,agg_bwd_only_max,agg_bwd_only_speedup,agg_fwd_bwd_mean,agg_fwd_bwd_min,agg_fwd_bwd_max,agg_fwd_bwd_speedup,target_xray_mean,target_geometry_total_mean,target_adp_total_mean,n_iterations,n_atoms,n_reflections,d_min
+cpu,1,0.113103,0.112087,0.116177,1.000,0.117169,0.114671,0.121472,1.000,0.109615,0.103882,0.129210,1.000,0.220804,0.218538,0.224117,1.000,0.168295,0.050306,0.003425,10,3051,23356,2.050076961517334
+cpu,2,0.079371,0.078412,0.080419,1.425,0.082772,0.081475,0.083990,1.416,0.071112,0.068595,0.091207,1.541,0.151549,0.150123,0.153226,1.457,0.112285,0.036993,0.003153,10,3051,23356,2.050076961517334
+cpu,4,0.068186,0.066976,0.069046,1.659,0.071653,0.071009,0.073620,1.635,0.056851,0.051149,0.077308,1.928,0.121116,0.118407,0.125663,1.823,0.081113,0.036427,0.002929,10,3051,23356,2.050076961517334
+cpu,8,0.054157,0.052808,0.056465,2.088,0.056912,0.055289,0.059030,2.059,0.041248,0.037923,0.061957,2.657,0.095952,0.093675,0.098929,2.301,0.066538,0.027685,0.002832,10,3051,23356,2.050076961517334
+cpu,16,0.044224,0.043969,0.044662,2.557,0.048189,0.047425,0.049536,2.431,0.035492,0.032328,0.054858,3.088,0.080881,0.079916,0.082032,2.730,0.055828,0.024657,0.002731,10,3051,23356,2.050076961517334
+gpu,0,0.008740,0.008616,0.008845,12.941,0.009387,0.009282,0.009504,12.482,0.005838,0.005728,0.006215,18.776,0.015329,0.015230,0.015458,14.404,0.007933,0.007074,0.003660,10,3051,23356,2.050076961517334
diff --git a/paper/figure3_performance/output/figure3b_profiling.png b/paper/figure3_performance/output/figure3b_profiling.png
index fa796f86..86fc8cc7 100644
Binary files a/paper/figure3_performance/output/figure3b_profiling.png and b/paper/figure3_performance/output/figure3b_profiling.png differ
diff --git a/paper/figure3_performance/refinement_cycle_benchmark/benchmark_worker.py b/paper/figure3_performance/refinement_cycle_benchmark/benchmark_worker.py
index 94324947..6254c26b 100644
--- a/paper/figure3_performance/refinement_cycle_benchmark/benchmark_worker.py
+++ b/paper/figure3_performance/refinement_cycle_benchmark/benchmark_worker.py
@@ -101,7 +101,7 @@ def run_benchmark(n_iterations: int, n_warmup: int, device_str: str = "cpu") ->
         data_file=MTZ_FILE,
         pdb=PDB_FILE,
         device=device,
-        target_mode="bhattacharyya",
+        target_mode="ml",
     )
 
     # Create and configure loss state with default weights