diff --git a/examples/multi_agent/agent_system.py b/examples/multi_agent/agent_system.py
index 449023512d..1ed29317e4 100644
--- a/examples/multi_agent/agent_system.py
+++ b/examples/multi_agent/agent_system.py
@@ -38,14 +38,24 @@ async def generate_response(args, prompt, key):
 
         # Extract new response tokens
         if "output_token_logprobs" in output["meta_info"]:
-            new_response_tokens = [item[1] for item in output["meta_info"]["output_token_logprobs"]]
+            output_token_logprobs = output["meta_info"]["output_token_logprobs"]
+            new_response_tokens = [item[1] for item in output_token_logprobs]
+            new_response_log_probs = [item[0] for item in output_token_logprobs]
         else:
             # abort
             new_response_tokens = []
+            new_response_log_probs = []
 
         # Update sample with tokens directly - avoiding re-tokenization
         sample.tokens = sample.tokens + new_response_tokens
         sample.response_length += len(new_response_tokens)
+        if sample.rollout_log_probs is None:
+            sample.rollout_log_probs = []
+        sample.rollout_log_probs += new_response_log_probs
+        assert len(sample.rollout_log_probs) == sample.response_length, (
+            f"rollout logprob length mismatch: {len(sample.rollout_log_probs)} logprobs "
+            f"vs {sample.response_length} response tokens"
+        )
         sample.response = output["text"]
 
         match output["meta_info"]["finish_reason"]["type"]:
diff --git a/examples/multi_agent/run-qwen3-30B-A3B-multi-agent.sh b/examples/multi_agent/run-qwen3-30B-A3B-multi-agent.sh
index d41f6bec21..d065b9101f 100644
--- a/examples/multi_agent/run-qwen3-30B-A3B-multi-agent.sh
+++ b/examples/multi_agent/run-qwen3-30B-A3B-multi-agent.sh
@@ -88,6 +88,7 @@ GRPO_ARGS=(
    --entropy-coef 0.00
    --eps-clip 0.2
    --eps-clip-high 0.28
+   --use-rollout-logprobs
 )
 
 OPTIMIZER_ARGS=(