diff --git a/klausur-service/backend/cv_vocab_pipeline.py b/klausur-service/backend/cv_vocab_pipeline.py
index 6f65d67..0fd82f1 100644
--- a/klausur-service/backend/cv_vocab_pipeline.py
+++ b/klausur-service/backend/cv_vocab_pipeline.py
@@ -5657,6 +5657,9 @@ async def llm_review_entries_streaming(
 
         prompt = _build_llm_prompt(table_lines)
 
+        logger.info("LLM review streaming: batch %d — sending %d entries to %s",
+                    batch_start // batch_size, len(batch_entries), model)
+
         t0 = time.time()
         async with httpx.AsyncClient(timeout=300.0) as client:
             resp = await client.post(
@@ -5665,7 +5668,8 @@ async def llm_review_entries_streaming(
                     "model": model,
                     "messages": [{"role": "user", "content": prompt}],
                     "stream": False,
-                    "options": {"temperature": 0.1, "num_predict": 4096},
+                    "think": False,   # qwen3: disable chain-of-thought
+                    "options": {"temperature": 0.1, "num_predict": 8192},
                 },
             )
             resp.raise_for_status()
@@ -5673,7 +5677,11 @@ async def llm_review_entries_streaming(
         batch_ms = int((time.time() - t0) * 1000)
         total_duration_ms += batch_ms
 
+        logger.info("LLM review streaming: response %dms, length=%d chars", batch_ms, len(content))
+        logger.debug("LLM review streaming raw (first 500): %.500s", content)
+
         corrected = _parse_llm_json_array(content)
+        logger.info("LLM review streaming: parsed %d entries, applying diff...", len(corrected))
         batch_changes, batch_corrected = _diff_batch(batch_entries, corrected)
 
         # Merge back