diff --git a/skyrl/train/trainer.py b/skyrl/train/trainer.py
index 91e38d3263..2124a74431 100644
--- a/skyrl/train/trainer.py
+++ b/skyrl/train/trainer.py
@@ -248,14 +248,24 @@ async def train(self):
                     with Timer("postprocess_generator_output", self.all_timings):
                         generator_output = self.postprocess_generator_output(generator_output, uids)
 
-                    # 2. print example just for debugging
-                    vis = self.tokenizer.decode(generator_output["response_ids"][0])
-                    log_example(
-                        logger,
-                        prompt=generator_input["prompts"][0],
-                        response=vis,
-                        reward=generator_output["rewards"][0],
-                    )
+                    # 2. print all trajectories for debugging
+                    # Prefer trajectory_ids from generator_output (step-wise mode expands them per step),
+                    # fall back to generator_input for standard mode.
+                    trajectory_ids = generator_output.get("trajectory_ids") or generator_input.get("trajectory_ids", [])
+                    input_prompts = generator_input.get("prompts", [])
+                    for i, (response_ids, reward) in enumerate(
+                        zip(generator_output["response_ids"], generator_output["rewards"])
+                    ):
+                        vis = self.tokenizer.decode(response_ids)
+                        tid = trajectory_ids[i] if i < len(trajectory_ids) else None
+                        tid_str = f"instance={tid.instance_id}, rep={tid.repetition_id}" if tid else f"idx={i}"
+                        prompt = input_prompts[i] if i < len(input_prompts) else "<prompt unavailable>"
+                        log_example(
+                            logger,
+                            prompt=prompt,
+                            response=f"[Trajectory {tid_str}]\n{vis}",
+                            reward=reward,
+                        )
 
                     # 3. Convert GeneratorOutput to TrainingInputBatch
                     with Timer("convert_to_training_input", self.all_timings):