diff --git a/config/src/config/quorum_store_config.rs b/config/src/config/quorum_store_config.rs
index 6e442023b1ffa..8a30b918bdae8 100644
--- a/config/src/config/quorum_store_config.rs
+++ b/config/src/config/quorum_store_config.rs
@@ -47,6 +47,7 @@ pub struct QuorumStoreConfig {
     pub channel_size: usize,
     pub proof_timeout_ms: usize,
     pub batch_generation_poll_interval_ms: usize,
+    // No longer being used
     pub batch_generation_min_non_empty_interval_ms: usize,
     pub batch_generation_max_interval_ms: usize,
     /// The maximum number of transactions that the batch generator puts in a batch.
@@ -91,6 +92,9 @@ pub struct QuorumStoreConfig {
     pub num_workers_for_remote_batches: usize,
     pub batch_buckets: Vec<u64>,
     pub allow_batches_without_pos_in_proposal: bool,
+    /// If batch generator pulls lesser than `minimum_batch_size` transactions from mempool, batch generator will
+    /// wait for more transactions to be available in mempool before creating a batch.
+    pub minimum_batch_size: u64,
 }
 
 impl Default for QuorumStoreConfig {
@@ -129,6 +133,7 @@ impl Default for QuorumStoreConfig {
             num_workers_for_remote_batches: 10,
             batch_buckets: DEFAULT_BUCKETS.to_vec(),
             allow_batches_without_pos_in_proposal: true,
+            minimum_batch_size: 20,
         }
     }
 }
diff --git a/consensus/src/quorum_store/batch_generator.rs b/consensus/src/quorum_store/batch_generator.rs
index b163709a74b5e..860d39e8b131a 100644
--- a/consensus/src/quorum_store/batch_generator.rs
+++ b/consensus/src/quorum_store/batch_generator.rs
@@ -23,7 +23,7 @@ use aptos_types::{transaction::SignedTransaction, PeerId};
 use futures_channel::mpsc::Sender;
 use rayon::prelude::*;
 use std::{
-    collections::{btree_map::Entry, BTreeMap, HashMap},
+    collections::{btree_map::Entry, BTreeMap, HashMap, VecDeque},
     sync::Arc,
     time::{Duration, Instant},
 };
@@ -402,6 +402,8 @@ impl BatchGenerator {
         let mut dynamic_pull_txn_per_s = (self.config.back_pressure.dynamic_min_txn_per_s
             + self.config.back_pressure.dynamic_max_txn_per_s)
             / 2;
+        // The number of transactions pulled from mempool in the past second
+        let mut pulled_txns_window: VecDeque<(Instant, u64)> = VecDeque::new();
 
         loop {
             let _timer = counters::BATCH_GENERATOR_MAIN_LOOP.start_timer();
@@ -448,24 +450,40 @@ impl BatchGenerator {
                     } else {
                         counters::QS_BACKPRESSURE_PROOF_COUNT.observe(0.0);
                     }
+
                     let since_last_non_empty_pull_ms = std::cmp::min(
                         tick_start.duration_since(last_non_empty_pull).as_millis(),
                         self.config.batch_generation_max_interval_ms as u128
                     ) as usize;
-                    if (!self.back_pressure.proof_count
-                        && since_last_non_empty_pull_ms >= self.config.batch_generation_min_non_empty_interval_ms)
-                        || since_last_non_empty_pull_ms == self.config.batch_generation_max_interval_ms {
 
-                        let dynamic_pull_max_txn = std::cmp::max(
-                            (since_last_non_empty_pull_ms as f64 / 1000.0 * dynamic_pull_txn_per_s as f64) as u64, 1);
+                    // If there is txn/proof backpressure, pull once every 250ms.
+                    // If there is no backpressure, pull every 25ms.
+                    if (!self.back_pressure.proof_count && !self.back_pressure.txn_count)
+                        || since_last_non_empty_pull_ms >= self.config.batch_generation_max_interval_ms {
+
+                        while let Some(&(timestamp, _)) = pulled_txns_window.front() {
+                            if timestamp.elapsed() >= Duration::from_secs(1) {
+                                pulled_txns_window.pop_front();
+                            } else {
+                                break;
+                            }
+                        }
+
+                        let pulled_in_last_sec = pulled_txns_window.iter().map(|(_, txns)| txns).sum();
                         let pull_max_txn = std::cmp::min(
-                            dynamic_pull_max_txn,
+                            std::cmp::max(dynamic_pull_txn_per_s.saturating_sub(pulled_in_last_sec), 1),
                             self.config.sender_max_total_txns as u64,
                         );
                         let batches = self.handle_scheduled_pull(pull_max_txn).await;
-                        if !batches.is_empty() {
-                            last_non_empty_pull = tick_start;
+                        let num_txns_in_batches: u64 = batches.iter().map(|b| b.batch_info().num_txns()).sum();
+                        info!("pulled_txns_window: {:?}, current_time: {:?}, pulled_in_last_sec: {}, pull_max_txn: {}, batch size: {}, num_txns_in_batches: {}", pulled_txns_window, Instant::now(), pulled_in_last_sec, pull_max_txn, batches.len(), num_txns_in_batches);
 
+                        // Ignore the batch if batch is smaller than minimum batch size and pull again in 25ms
+                        // If a batch hasn't been created for a while, send it out even if it's smaller than minimum batch size
+                        if num_txns_in_batches as u64 >= self.config.minimum_batch_size
+                            || (!batches.is_empty() && since_last_non_empty_pull_ms >= self.config.batch_generation_min_non_empty_interval_ms) {
+                            last_non_empty_pull = tick_start;
+                            pulled_txns_window.push_back((tick_start, num_txns_in_batches as u64));
                             let persist_start = Instant::now();
                             let mut persist_requests = vec![];
                             for batch in batches.clone().into_iter() {