diff --git a/rdagent/components/coder/factor_coder/CoSTEER/__init__.py b/rdagent/components/coder/factor_coder/CoSTEER/__init__.py
index e3a81409..9fa2f84e 100644
--- a/rdagent/components/coder/factor_coder/CoSTEER/__init__.py
+++ b/rdagent/components/coder/factor_coder/CoSTEER/__init__.py
@@ -98,4 +98,5 @@ class FactorCoSTEER(TaskGenerator[FactorExperiment]):
         if self.new_knowledge_base_path is not None:
             pickle.dump(factor_knowledge_base, open(self.new_knowledge_base_path, "wb"))
         self.knowledge_base = factor_knowledge_base
+        factor_experiment.based_experiments = exp.based_experiments
         return factor_experiment
diff --git a/rdagent/components/coder/factor_coder/CoSTEER/evolving_strategy.py b/rdagent/components/coder/factor_coder/CoSTEER/evolving_strategy.py
index fcda9ad8..94e1af9b 100644
--- a/rdagent/components/coder/factor_coder/CoSTEER/evolving_strategy.py
+++ b/rdagent/components/coder/factor_coder/CoSTEER/evolving_strategy.py
@@ -149,7 +149,7 @@ class FactorEvolvingStrategy(MultiProcessEvolvingStrategy):
                     queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
                 )
             )
-            session = APIBackend(use_chat_cache=False).build_chat_session(
+            session = APIBackend(use_chat_cache=True).build_chat_session(
                 session_system_prompt=system_prompt,
             )
 
@@ -249,7 +249,7 @@ class FactorEvolvingStrategyWithGraph(MultiProcessEvolvingStrategy):
                 )
             )
 
-            session = APIBackend(use_chat_cache=False).build_chat_session(
+            session = APIBackend(use_chat_cache=True).build_chat_session(
                 session_system_prompt=system_prompt,
             )
 
@@ -276,7 +276,7 @@ class FactorEvolvingStrategyWithGraph(MultiProcessEvolvingStrategy):
                         )
                         .strip("\n")
                     )
-                    session_summary = APIBackend(use_chat_cache=False).build_chat_session(
+                    session_summary = APIBackend(use_chat_cache=True).build_chat_session(
                         session_system_prompt=error_summary_system_prompt,
                     )
                     for _ in range(10):  # max attempt to reduce the length of error_summary_user_prompt
diff --git a/rdagent/components/coder/model_coder/CoSTEER/__init__.py b/rdagent/components/coder/model_coder/CoSTEER/__init__.py
index be9ede38..a5818987 100644
--- a/rdagent/components/coder/model_coder/CoSTEER/__init__.py
+++ b/rdagent/components/coder/model_coder/CoSTEER/__init__.py
@@ -83,4 +83,5 @@ class ModelCoSTEER(TaskGenerator[ModelExperiment]):
         if self.new_knowledge_base_path is not None:
             pickle.dump(model_knowledge_base, open(self.new_knowledge_base_path, "wb"))
         self.knowledge_base = model_knowledge_base
+        model_experiment.based_experiments = exp.based_experiments
         return model_experiment
diff --git a/rdagent/scenarios/qlib/factor_proposal.py b/rdagent/scenarios/qlib/factor_proposal.py
index 7be92420..b66d752d 100644
--- a/rdagent/scenarios/qlib/factor_proposal.py
+++ b/rdagent/scenarios/qlib/factor_proposal.py
@@ -77,4 +77,6 @@ class QlibFactorHypothesis2Experiment(FactorHypothesis2Experiment):
             tasks.append(FactorTask(factor_name, description, formulation, variables))
         exp = FactorExperiment(tasks)
         exp.based_experiments = [t[1] for t in trace.hist if t[2]]
+        if len(exp.based_experiments) == 0:
+            exp.based_experiments.append(FactorExperiment(sub_tasks=[]))
         return exp