diff --git a/ludwig/schema/llms/generation.py b/ludwig/schema/llms/generation.py
index a12c9990ee3..e1d96e21f94 100644
--- a/ludwig/schema/llms/generation.py
+++ b/ludwig/schema/llms/generation.py
@@ -19,7 +19,10 @@ class LLMGenerationConfig(schema_utils.BaseMarshmallowConfig):
     max_new_tokens: Optional[int] = schema_utils.PositiveInteger(
         default=32,
         allow_none=True,
-        description="The maximum number of new tokens to generate, ignoring the number of tokens in the input prompt.",
+        description="The maximum number of new tokens to generate, ignoring the number of tokens in the input prompt. "
+        "If not set, this is dynamically determined by Ludwig based on either the `max_sequence_length` of the ouput "
+        "feature, the global_max_sequence_length specified in preprocessing (if specified), or the "
+        "maximum context length supported by the model (in the order specified).",
         parameter_metadata=LLM_METADATA["generation"]["max_new_tokens"],
     )