diff --git a/llama_stack/apis/eval/eval.py b/llama_stack/apis/eval/eval.py
index 5ae779ca7..50fb922fe 100644
--- a/llama_stack/apis/eval/eval.py
+++ b/llama_stack/apis/eval/eval.py
@@ -70,7 +70,6 @@ class Eval(Protocol):
     async def run_eval(
         self,
         task_id: str,
-        task_def: EvalTaskDef,
         task_config: EvalTaskConfig,
     ) -> Job: ...
 
diff --git a/llama_stack/distribution/routers/routers.py b/llama_stack/distribution/routers/routers.py
index 7fc65800f..8edf950b2 100644
--- a/llama_stack/distribution/routers/routers.py
+++ b/llama_stack/distribution/routers/routers.py
@@ -267,12 +267,10 @@ class EvalRouter(Eval):
     async def run_eval(
         self,
         task_id: str,
-        task_def: EvalTaskDef,
         task_config: AppEvalTaskConfig,
     ) -> Job:
         return await self.routing_table.get_provider_impl(task_id).run_eval(
             task_id=task_id,
-            task_def=task_def,
             task_config=task_config,
         )
 
diff --git a/llama_stack/providers/inline/meta_reference/eval/eval.py b/llama_stack/providers/inline/meta_reference/eval/eval.py
index d20cf30d2..57d1d0124 100644
--- a/llama_stack/providers/inline/meta_reference/eval/eval.py
+++ b/llama_stack/providers/inline/meta_reference/eval/eval.py
@@ -51,22 +51,20 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
         # TODO: assume sync job, will need jobs API for async scheduling
         self.jobs = {}
 
+        self.eval_tasks = {}
+
     async def initialize(self) -> None: ...
 
     async def shutdown(self) -> None: ...
 
+    async def register_eval_task(self, task_def: EvalTaskDef) -> None:
+        self.eval_tasks[task_def.identifier] = task_def
+
     async def list_eval_tasks(self) -> List[EvalTaskDef]:
         # NOTE: In order to be routed to this provider, the eval task def must have
         # a EvalTaskDef with identifier defined as DEFAULT_EVAL_TASK_IDENTIFIER
         # for app eval where eval task benchmark_id is not pre-registered
-        eval_tasks = [
-            EvalTaskDef(
-                identifier=DEFAULT_EVAL_TASK_IDENTIFIER,
-                dataset_id="",
-                scoring_functions=[],
-            )
-        ]
-        return eval_tasks
+        return list(self.eval_tasks.values())
 
     async def validate_eval_input_dataset_schema(self, dataset_id: str) -> None:
         dataset_def = await self.datasets_api.get_dataset(dataset_identifier=dataset_id)
@@ -94,9 +92,9 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
     async def run_eval(
         self,
         task_id: str,
-        task_def: EvalTaskDef,
         task_config: EvalTaskConfig,
     ) -> Job:
+        task_def = self.eval_tasks[task_id]
         dataset_id = task_def.dataset_id
         candidate = task_config.eval_candidate
         scoring_functions = task_def.scoring_functions
diff --git a/llama_stack/providers/tests/eval/test_eval.py b/llama_stack/providers/tests/eval/test_eval.py
index 794026e63..242be50ca 100644
--- a/llama_stack/providers/tests/eval/test_eval.py
+++ b/llama_stack/providers/tests/eval/test_eval.py
@@ -9,7 +9,11 @@ import pytest
 
 from llama_models.llama3.api import SamplingParams
 
-from llama_stack.apis.eval.eval import AppEvalTaskConfig, EvalTaskDef, ModelCandidate
+from llama_stack.apis.eval.eval import (
+    AppEvalTaskConfig,
+    EvalTaskDefWithProvider,
+    ModelCandidate,
+)
 from llama_stack.providers.tests.datasetio.test_datasetio import register_dataset
 
 
@@ -28,7 +32,7 @@ class Testeval:
         _, eval_tasks_impl, _, _, _, _ = eval_stack
         response = await eval_tasks_impl.list_eval_tasks()
         assert isinstance(response, list)
-        assert len(response) >= 1
+        assert len(response) == 0
 
     @pytest.mark.asyncio
     async def test_eval_evaluate_rows(self, eval_stack):
@@ -48,9 +52,17 @@ class Testeval:
             "meta-reference::llm_as_judge_8b_correctness",
             "meta-reference::equality",
         ]
+        task_id = "meta-reference::app_eval"
+        task_def = EvalTaskDefWithProvider(
+            identifier=task_id,
+            dataset_id="test_dataset_for_eval",
+            scoring_functions=scoring_functions,
+            provider_id="meta-reference",
+        )
+        await eval_tasks_impl.register_eval_task(task_def)
 
         response = await eval_impl.evaluate_rows(
-            task_id="meta-reference::app_eval",
+            task_id=task_id,
             input_rows=rows.rows,
             scoring_functions=scoring_functions,
             task_config=AppEvalTaskConfig(
@@ -76,15 +88,16 @@ class Testeval:
             "meta-reference::subset_of",
         ]
 
-        task_id = "meta-reference::app_eval"
+        task_id = "meta-reference::app_eval-2"
+        task_def = EvalTaskDefWithProvider(
+            identifier=task_id,
+            dataset_id="test_dataset_for_eval",
+            scoring_functions=scoring_functions,
+            provider_id="meta-reference",
+        )
+        await eval_tasks_impl.register_eval_task(task_def)
         response = await eval_impl.run_eval(
             task_id=task_id,
-            task_def=EvalTaskDef(
-                # NOTE: this is needed to make the router work for all app evals
-                identifier=task_id,
-                dataset_id="test_dataset_for_eval",
-                scoring_functions=scoring_functions,
-            ),
             task_config=AppEvalTaskConfig(
                 eval_candidate=ModelCandidate(
                     model="Llama3.2-3B-Instruct",