diff --git a/crates/ml-hyperopt/src/optimizer.rs b/crates/ml-hyperopt/src/optimizer.rs
index d8ec488a5..c72e7903e 100644
--- a/crates/ml-hyperopt/src/optimizer.rs
+++ b/crates/ml-hyperopt/src/optimizer.rs
@@ -354,12 +354,15 @@ impl ArgminOptimizer {
         let trials_used = trials.len();
         let remaining_trials = self.max_trials.saturating_sub(trials_used);
 
-        // For small budgets (< 2× swarm size), PSO can't complete a full iteration.
-        // Fall back to additional LHS samples which give better space coverage.
+        // Algorithm selection by budget:
+        // - Budget < 2× swarm → LHS (space-filling, optimal for small budgets)
+        // - Budget ≥ 2× swarm → PSO (swarm optimization, needs full iterations)
+        // PSO with a partial swarm is worse than random — 75% of particles
+        // have no cost value, so gbest is computed from incomplete data.
         if remaining_trials < self.n_particles * 2 {
             info!(
-                "Budget too small for PSO ({} remaining < {} particles × 2). Using additional LHS.",
-                remaining_trials, self.n_particles
+                "Using LHS for remaining {} trials (budget < {} particles × 2 = need PSO budget ≥ {})",
+                remaining_trials, self.n_particles, self.n_particles * 2
             );
             let extra_samples = Self::latin_hypercube_sampling(remaining_trials, &bounds, &mut rng);
             let mut model_guard = cost_fn.model.lock().map_err(|e| anyhow::anyhow!("lock: {e}"))?;
@@ -379,7 +382,7 @@ impl ArgminOptimizer {
         let max_iters_by_budget =
             ((remaining_trials as f64) / (self.n_particles as f64)).ceil() as usize;
         let max_iters = if remaining_trials < self.n_particles * 2 {
-            0 // Already handled by LHS fallback above
+            0 // LHS already handled above
         } else {
             max_iters_by_budget.min(self.max_iters_per_restart)
         };