diff --git a/blog/dist/part-1.html b/blog/dist/part-1.html
index 0582edcb..20b1f70f 100644
--- a/blog/dist/part-1.html
+++ b/blog/dist/part-1.html
@@ -465,11 +465,13 @@ <h2 id="a-frozen-llm-joins-the-ladder">A frozen LLM joins the
 <p>Everything so far is a purpose-built forecasting method. What happens
 if we hand the same task to a general LLM — no fine-tuning, no
 forecasting head, just the raw model?</p>
-<p>The technique is the LLM-Process (LLMP): serialize the return history
-— and, optionally, the same covariate panel — into a text prompt, and
-ask the model to emit the full quantile grid directly, as numbers. We
-then score those quantiles with CRPS exactly like every other method. No
-special pleading: same origins, same cutoff, same referee.</p>
+<p>The technique is the LLM-Process (LLMP), introduced by <a
+href="https://arxiv.org/abs/2405.12856">Requeima et al. (2024)</a>:
+serialize the return history — and, optionally, the same covariate panel
+— into a text prompt, and ask the model to emit the full quantile grid
+directly, as numbers. We then score those quantiles with CRPS exactly
+like every other method. No special pleading: same origins, same cutoff,
+same referee.</p>
 <p>The result that matters is that this works at all. A frozen,
 general-purpose model, prompted with a column of numbers, emits a
 genuine predictive distribution — calibrated enough to compete. In the
@@ -502,6 +504,13 @@ <h2 id="the-rung-that-has-to-read">The rung that has to read</h2>
 of the same inputs breaks through it. The next rung has to read. In Part
 2, we give the forecaster the news — and then face the harder problem of
 how you trust one that does.</p>
+<hr />
+<p><strong>Reference</strong></p>
+<p>Requeima, J., Bronskill, J., Choi, D., Turner, R. E., &amp; Duvenaud,
+D. (2024). <a href="https://arxiv.org/abs/2405.12856">LLM Processes:
+Numerical Predictive Distributions Conditioned on Natural Language</a>.
+<em>Advances in Neural Information Processing Systems, 37</em>,
+109609–109671.</p>
 
 </article>
 <footer class="post-footer">
diff --git a/blog/part-1/post.md b/blog/part-1/post.md
index f8af6328..4eec00b9 100644
--- a/blog/part-1/post.md
+++ b/blog/part-1/post.md
@@ -183,7 +183,8 @@ Everything so far is a purpose-built forecasting method. What happens if we hand
 the same task to a general LLM — no fine-tuning, no forecasting head, just the
 raw model?
 
-The technique is the LLM-Process (LLMP): serialize the return history — and,
+The technique is the LLM-Process (LLMP), introduced by [Requeima et al.
+(2024)](https://arxiv.org/abs/2405.12856): serialize the return history — and,
 optionally, the same covariate panel — into a text prompt, and ask the model to
 emit the full quantile grid directly, as numbers. We then score those quantiles
 with CRPS exactly like every other method. No special pleading: same origins,
@@ -222,3 +223,12 @@ That is the ceiling of numbers-only forecasting, and no rearrangement of the
 same inputs breaks through it. The next rung has to read. In Part 2, we give the
 forecaster the news — and then face the harder problem of how you trust one that
 does.
+
+---
+
+**Reference**
+
+Requeima, J., Bronskill, J., Choi, D., Turner, R. E., & Duvenaud, D. (2024).
+[LLM Processes: Numerical Predictive Distributions Conditioned on Natural
+Language](https://arxiv.org/abs/2405.12856). *Advances in Neural Information
+Processing Systems, 37*, 109609–109671.