diff --git a/README.md b/README.md
index 12a29d1..846d742 100644
--- a/README.md
+++ b/README.md
@@ -132,6 +132,30 @@ python infer_refine.py critic-infer.csv VDebugger/VDebugger-refiner-generalist-1
 ```
 Then you can execute the programs in `critic-refine-infer.csv` as in step 2 of [Generation and Execution of Visual Programs](https://github.com/shirley-wu/vdebugger/tree/main?tab=readme-ov-file#generation-and-execution-of-visual-programs)
 
+## Run VDebugger for Multiple Iterations
+
+To run VDebugger for `T` iterations (`T` > 1), you first need to generate the initial programs and collect their execution feedback as in step 1 and 3 in [Generation and Execution of Visual Programs](https://github.com/shirley-wu/vdebugger/tree/main?tab=readme-ov-file#generation-and-execution-of-visual-programs). Then, you need to repeat the steps below for `T` times:
+1. Infer critic, as in [Inference of VDebugger](https://github.com/shirley-wu/vdebugger/tree/main?tab=readme-ov-file#inference-of-vdebugger);
+2. Infer refiner, as in [Inference of VDebugger](https://github.com/shirley-wu/vdebugger/tree/main?tab=readme-ov-file#inference-of-vdebugger);
+3. Collect execution feedback for the new programs generated by refiner, as in step 2 in [Generation and Execution of Visual Programs](https://github.com/shirley-wu/vdebugger/tree/main?tab=readme-ov-file#generation-and-execution-of-visual-programs). The next iteration will be run on top of the feedback collected in this step.
+
+Then after `T` iterations, evaluate the final programs as in step 2 in [Generation and Execution of Visual Programs](https://github.com/shirley-wu/vdebugger/tree/main?tab=readme-ov-file#generation-and-execution-of-visual-programs).
+
+Since the major computational overhead comes from program execution (i.e. step 3 in each iteration), you can use the helper scripts `remove_dup.py` and `merge_csv.py` in [vdebugger/interative_helper](vdebugger/iterative_helper) to reduce the redundant execution:
+* Before step 3 in each iteration, remove the programs that are duplicate as last iteration, by executing:
+```bash
+python remove_dup.py PROGRAM_CSV_FROM_LAST_ITERATION PROGRAM_CSV_FOR_THIS_ITERATION
+```
+which will produce a dedup file `PROGRAM_CSV_FOR_THIS_ITERATION_DEDUP`
+* Then collect execution feedback for the resulted `PROGRAM_CSV_FOR_THIS_ITERATION_DEDUP`.
+* After collecting their feedback, merge the execution feedback from the last iteration and the current iteration:
+```bash
+python merge_csv.py EXECUTION_FEEDBACK_FROM_LAST_ITERATION EXECUTION_FEEDBACK_FOR_THIS_ITERATION_DEDUP
+```
+which will produce a file `EXECUTION_FEEDBACK_FOR_THIS_ITERATION_MERGED` containing merged execution results. For the next iteration, use the merged execution results.
+
+There will still be some repeated computation within step 1 and 2 in this iteration, but that will be tolerable. If you are concerned, you can modify the scripts by yourself to avoid the computation.
+
 ## Training of VDebugger
 
 If you want to reproduce our training of VDebugger, please use `vdebugger/training_scripts/train_{critic, refiner}.sh`. You will need to install `deepspeed==0.14.0`.
diff --git a/vdebugger/iterative_helper/merge_csv.py b/vdebugger/iterative_helper/merge_csv.py
new file mode 100644
index 0000000..3f348c4
--- /dev/null
+++ b/vdebugger/iterative_helper/merge_csv.py
@@ -0,0 +1,21 @@
+import argparse
+import os
+
+import pandas as pd
+
+parser = argparse.ArgumentParser()
+parser.add_argument('orig')
+parser.add_argument('new')
+args = parser.parse_args()
+
+orig = pd.read_csv(args.orig)
+new = pd.read_csv(args.new)
+new.loc[new['code'] == '[', 'result'] = orig.loc[new['code'] == '[', 'result']
+if 'traced' in new:
+    new.loc[new['code'] == '[', 'traced'] = orig.loc[new['code'] == '[', 'traced']
+new.loc[new['code'] == '[', 'code'] = orig.loc[new['code'] == '[', 'code']
+
+out_fname = args.new.replace('.csv', '.merged.csv')
+print("Dump to", out_fname)
+assert not os.path.exists(out_fname), "File {} exists".format(out_fname)
+new.to_csv(out_fname)
diff --git a/vdebugger/iterative_helper/remove_dup.py b/vdebugger/iterative_helper/remove_dup.py
new file mode 100644
index 0000000..c26c52e
--- /dev/null
+++ b/vdebugger/iterative_helper/remove_dup.py
@@ -0,0 +1,19 @@
+import argparse
+import os
+
+import pandas as pd
+
+parser = argparse.ArgumentParser()
+parser.add_argument('orig')
+parser.add_argument('new')
+args = parser.parse_args()
+
+orig = pd.read_csv(args.orig)
+new = pd.read_csv(args.new)
+new.loc[new['code'] == orig['code'], 'code'] = '['
+print("Remaining valid code", (new['code'] != '[').sum())
+
+out_fname = args.new.replace('.csv', '.remove-dup.csv')
+print("Dump to", out_fname)
+assert not os.path.exists(out_fname), "File {} exists".format(out_fname)
+new.to_csv(out_fname)