diff --git a/crates/ml/src/cuda_pipeline/c51_loss_kernel.cu b/crates/ml/src/cuda_pipeline/c51_loss_kernel.cu
index 6af86fc31..a4a9e0ccf 100644
--- a/crates/ml/src/cuda_pipeline/c51_loss_kernel.cu
+++ b/crates/ml/src/cuda_pipeline/c51_loss_kernel.cu
@@ -476,15 +476,26 @@ extern "C" __global__ void c51_loss_batched(
          * When Q-values differentiate, this converges to argmax.
          * When Q-values are flat, this averages → stable learning signal. */
 
-        /* Compute softmax weights from expected Q per action */
+        /* Compute softmax weights from expected Q per action.
+         * Temperature = max(Q_gap/3, 0.001). When Q-values differentiate,
+         * tau shrinks → weights sharpen toward argmax. When flat, tau=0.001
+         * still produces exp(0.06/0.001)=exp(60) → effectively argmax for
+         * ANY non-zero Q-difference. This prevents the uniform-averaging
+         * fixed point where all actions get equal weight → zero gradient
+         * to push Q-values apart. */
         __shared__ float action_weights[MAX_BRANCH_SIZE];
         if (tid == 0) {
             float max_eq = eq_per_action[0];
-            for (int a = 1; a < n_d; a++)
+            float min_eq = eq_per_action[0];
+            for (int a = 1; a < n_d; a++) {
                 max_eq = fmaxf(max_eq, eq_per_action[a]);
+                min_eq = fminf(min_eq, eq_per_action[a]);
+            }
+            float q_gap_local = max_eq - min_eq;
+            float tau = fmaxf(q_gap_local / 3.0f, 0.001f);
             float sum_exp = 0.0f;
             for (int a = 0; a < n_d; a++) {
-                action_weights[a] = expf(eq_per_action[a] - max_eq);
+                action_weights[a] = expf((eq_per_action[a] - max_eq) / tau);
                 sum_exp += action_weights[a];
             }
             for (int a = 0; a < n_d; a++)