diff --git a/tensorrt_llm/executor/result.py b/tensorrt_llm/executor/result.py
index 145a0275dd0a..d9fe12404c5a 100644
--- a/tensorrt_llm/executor/result.py
+++ b/tensorrt_llm/executor/result.py
@@ -1023,9 +1023,19 @@ def _handle_response(self, response: "GenerationExecutor.Response"):
                 is not None) and self.tokenizer is not None:
             for beam_output in self.outputs:
                 beam_output._last_text_len = len(beam_output.text)
-                if hasattr(
-                        self.tokenizer, 'decode_incrementally'
-                ) and self._streaming and not self.sampling_params.use_beam_search:
+                # A stop word matched by token id is trimmed from token_ids in
+                # _handle_sequence. If some of its ids came in earlier streamed
+                # responses, the incremental decoder has already turned them
+                # into text, or holds the lead bytes of a character that the
+                # final flush would emit as U+FFFD. Decode the kept ids afresh
+                # instead, as the non-streaming path does.
+                trimmed_streamed_ids = (len(beam_output.token_ids)
+                                        < beam_output._last_token_ids_len)
+                use_incremental = (self._streaming
+                                   and not self.sampling_params.use_beam_search
+                                   and not trimmed_streamed_ids and hasattr(
+                                       self.tokenizer, 'decode_incrementally'))
+                if use_incremental:
                     beam_output.text, beam_output._incremental_states = self.tokenizer.decode_incrementally(
                         beam_output.token_ids_diff,
                         prev_text=beam_output.text,
@@ -1036,6 +1046,13 @@ def _handle_response(self, response: "GenerationExecutor.Response"):
                 else:
                     beam_output.text = self.tokenizer.decode(
                         beam_output.token_ids, **kwargs)
+                    if trimmed_streamed_ids:
+                        # Drop the pending flush, and move the watermark to the
+                        # kept ids so that later responses for sibling
+                        # candidates (n > 1) do not decode this output again.
+                        beam_output._incremental_states = None
+                        beam_output._last_token_ids_len = len(
+                            beam_output.token_ids)
 
                 is_generating = not self._done
                 is_finished_with_stop_or_length = (
diff --git a/tests/integration/test_lists/test-db/l0_cpu.yml b/tests/integration/test_lists/test-db/l0_cpu.yml
index edf634c48fcf..bd5c4ee2d27a 100644
--- a/tests/integration/test_lists/test-db/l0_cpu.yml
+++ b/tests/integration/test_lists/test-db/l0_cpu.yml
@@ -206,6 +206,7 @@ l0_cpu:
   - unittest/executor/test_stats_serializer.py
   - unittest/executor/test_spec_dec_perf_metrics.py
   - unittest/executor/test_stop_word_side_output_trim.py
+  - unittest/executor/test_stop_word_streaming_text.py
   - unittest/executor/test_ray_stub.py
   - unittest/executor/test_generation_request_array_prompt.py
   - unittest/grpc/openengine/test_coordinator.py
diff --git a/tests/unittest/executor/test_stop_word_streaming_text.py b/tests/unittest/executor/test_stop_word_streaming_text.py
new file mode 100644
index 000000000000..d4104ac456fd
--- /dev/null
+++ b/tests/unittest/executor/test_stop_word_streaming_text.py
@@ -0,0 +1,272 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""A streamed request that stops on a multi-token stop word must end with the non-streamed text.
+
+When the sampler matches a stop word by token id, ``_handle_sequence`` trims
+the stop ids from ``token_ids``. In a streamed request some of those ids may
+have arrived in earlier responses and already been decoded incrementally, so
+the final ``.text`` has to be decoded again from the kept ids: otherwise it
+keeps the start of the stop string (``"<answer>42</answer"``), or ends in
+U+FFFD when the decoder still holds the lead bytes of a character that only
+the stop word completes (a Thai digit such as ``"๒"``). CPU-only; no engine
+needed.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+from tokenizers import Tokenizer
+from tokenizers.decoders import ByteFallback, Fuse, Sequence
+from tokenizers.models import WordLevel
+from transformers import PreTrainedTokenizerFast
+
+from tensorrt_llm.bindings import executor as tllm
+from tensorrt_llm.executor.result import DetokenizedGenerationResultBase
+from tensorrt_llm.sampling_params import SamplingParams
+from tensorrt_llm.tokenizer import TransformersTokenizer
+
+pytestmark = pytest.mark.cpu_only
+
+# "๑" and "๒" are E0 B9 91 and E0 B9 92 in UTF-8: one byte token each, so the
+# decoder holds <0xE0><0xB9> until the last byte completes the character.
+VOCAB = {
+    "<unk>": 0,
+    "<0xE0>": 1,
+    "<0xB9>": 2,
+    "<0x91>": 3,
+    "<0x92>": 4,
+    "ข้อ": 5,
+    " ": 6,
+    "<answer>": 7,
+    "42": 8,
+    "</": 9,
+    "answer": 10,
+    ">": 11,
+}
+THAI_DIGIT_TWO = [1, 2, 4]
+# "ข้อ ๑ ข้อ ๒"
+THAI_GENERATED = [5, 6, 1, 2, 3, 6, 5, 6, *THAI_DIGIT_TWO]
+ANSWER_CLOSE = [9, 10, 11]
+# "<answer>42</answer>"
+ANSWER_GENERATED = [7, 8, *ANSWER_CLOSE]
+
+
+@pytest.fixture(scope="module")
+def tokenizer() -> TransformersTokenizer:
+    backend = Tokenizer(WordLevel(VOCAB, unk_token="<unk>"))
+    backend.decoder = Sequence([ByteFallback(), Fuse()])
+    return TransformersTokenizer(PreTrainedTokenizerFast(tokenizer_object=backend))
+
+
+@pytest.fixture(params=["HF", "TRTLLM"], autouse=True)
+def detokenization_backend(request, monkeypatch: pytest.MonkeyPatch) -> str:
+    from tensorrt_llm.tokenizer import tokenizer as tokenizer_module
+
+    monkeypatch.setattr(tokenizer_module, "TLLM_INCREMENTAL_DETOKENIZATION_BACKEND", request.param)
+    return request.param
+
+
+def _response(token_ids, is_final, finish_reason, sequence_index=0):
+    result = SimpleNamespace(
+        is_final=is_final,
+        context_phase_params=None,
+        decoding_iter=1,
+        avg_decoded_tokens_per_iter=None,
+        finish_reasons=[finish_reason],
+        output_token_ids=[list(token_ids)],
+        sequence_index=sequence_index,
+        cum_log_probs=None,
+        log_probs=None,
+        generation_logits=None,
+        per_pos_drafted=None,
+        per_pos_accepted=None,
+        spec_dec_totals=None,
+        time_breakdown_metrics=None,
+        request_perf_metrics=None,
+        additional_context_outputs=None,
+        additional_generation_outputs=None,
+        context_logits=None,
+    )
+    return SimpleNamespace(has_error=lambda: False, result=result)
+
+
+def _sampling_params(stop, stop_ids, **kwargs):
+    sampling_params = SamplingParams(stop=stop, **kwargs)
+    # Stands in for SamplingParams._setup tokenizing the stop string.
+    sampling_params._stop_word_ids = [stop_ids]
+    sampling_params._stream_interval = 1
+    return sampling_params
+
+
+def _run(tokenizer, sampling_params, chunks, finish_reason, streaming=True):
+    """Feed one response per chunk; the last one finishes the request with finish_reason."""
+    result = DetokenizedGenerationResultBase(
+        0, sampling_params, tokenizer=tokenizer, streaming=streaming
+    )
+    deltas = []
+    for index, chunk in enumerate(chunks):
+        is_final = index == len(chunks) - 1
+        result._handle_response(
+            _response(
+                chunk, is_final, finish_reason if is_final else tllm.FinishReason.NOT_FINISHED
+            )
+        )
+        deltas.append(result.outputs[0].text_diff)
+    return result.outputs[0], "".join(deltas)
+
+
+def _chunked(token_ids, size):
+    return [token_ids[i : i + size] for i in range(0, len(token_ids), size)]
+
+
+# The last field is the joined text_diff values. For "</answer>" they still end in
+# "</answer": earlier responses already streamed the decoded stop-word prefix, and
+# deltas cannot take text back (#21318). Change it here once streaming holds back a
+# possible stop-word prefix.
+CASES = [
+    pytest.param("๒", THAI_DIGIT_TWO, THAI_GENERATED, "ข้อ ๑ ข้อ ", "ข้อ ๑ ข้อ ", id="thai_digit"),
+    pytest.param(
+        "</answer>",
+        ANSWER_CLOSE,
+        ANSWER_GENERATED,
+        "<answer>42",
+        "<answer>42</answer",
+        id="answer_tag",
+    ),
+]
+
+
+@pytest.mark.parametrize("chunk_size", [1, 2])
+@pytest.mark.parametrize("stop, stop_ids, generated, expected, expected_joined", CASES)
+def test_streamed_text_drops_stop_word_ids_streamed_earlier(
+    tokenizer, stop, stop_ids, generated, expected, expected_joined, chunk_size
+):
+    sampling_params = _sampling_params(stop, stop_ids)
+    output, joined = _run(
+        tokenizer, sampling_params, _chunked(generated, chunk_size), tllm.FinishReason.STOP_WORDS
+    )
+    reference, _ = _run(
+        tokenizer,
+        _sampling_params(stop, stop_ids),
+        [generated],
+        tllm.FinishReason.STOP_WORDS,
+        streaming=False,
+    )
+
+    assert reference.text == expected
+    assert output.text == expected
+    assert output.token_ids == reference.token_ids
+    assert output.finish_reason == "stop"
+    assert output.stop_reason == stop
+    assert joined == expected_joined
+
+
+@pytest.mark.parametrize("chunk_size", [1, 2])
+def test_streamed_deltas_do_not_end_in_replacement_character(tokenizer, chunk_size):
+    """The held lead bytes of the stop character must not be flushed into the last delta."""
+    _, joined = _run(
+        tokenizer,
+        _sampling_params("๒", THAI_DIGIT_TWO),
+        _chunked(THAI_GENERATED, chunk_size),
+        tllm.FinishReason.STOP_WORDS,
+    )
+    assert "\N{REPLACEMENT CHARACTER}" not in joined
+    assert joined == "ข้อ ๑ ข้อ "
+
+
+def test_stop_word_inside_final_response_keeps_streamed_text(tokenizer):
+    """Stop ids that all arrive in the final response never reached the decoder."""
+    chunks = [THAI_GENERATED[:-3], THAI_DIGIT_TWO]
+    output, joined = _run(
+        tokenizer, _sampling_params("๒", THAI_DIGIT_TWO), chunks, tllm.FinishReason.STOP_WORDS
+    )
+    assert output.text == joined == "ข้อ ๑ ข้อ "
+
+
+def test_include_stop_str_in_output_keeps_stop_word(tokenizer):
+    output, joined = _run(
+        tokenizer,
+        _sampling_params("๒", THAI_DIGIT_TWO, include_stop_str_in_output=True),
+        _chunked(THAI_GENERATED, 1),
+        tllm.FinishReason.STOP_WORDS,
+    )
+    assert output.token_ids == THAI_GENERATED
+    assert output.text == joined == "ข้อ ๑ ข้อ ๒"
+
+
+def test_length_finish_flushes_held_bytes(tokenizer):
+    """Without a stop-word trim the final flush still emits what a full decode gives."""
+    generated = THAI_GENERATED[:-1]
+    output, _ = _run(
+        tokenizer,
+        _sampling_params("๒", THAI_DIGIT_TWO),
+        _chunked(generated, 1),
+        tllm.FinishReason.LENGTH,
+    )
+    assert output.finish_reason == "length"
+    assert output.text == tokenizer.decode(generated)
+
+
+def test_trimmed_candidate_is_decoded_once_when_siblings_stream_on(tokenizer, monkeypatch):
+    """With n > 1, a candidate that stopped on a trimmed stop word is not decoded again.
+
+    Candidate 0 streams to its stop word first; candidate 1 then streams on and
+    each of its responses walks every output. Candidate 0 must keep the
+    non-streamed text and an empty text_diff without another full decode.
+    """
+    sampling_params = _sampling_params(
+        "๒", THAI_DIGIT_TWO, n=2, best_of=2, temperature=0.8, top_p=0.9
+    )
+    result = DetokenizedGenerationResultBase(
+        0, sampling_params, tokenizer=tokenizer, streaming=True
+    )
+    full_decodes = []
+    decode = tokenizer.decode
+
+    def counting_decode(token_ids, *args, **kwargs):
+        full_decodes.append(list(token_ids))
+        return decode(token_ids, *args, **kwargs)
+
+    monkeypatch.setattr(tokenizer, "decode", counting_decode)
+
+    sibling = [5, 6, 7, 8]
+    steps = [
+        (0, [token_id], tllm.FinishReason.STOP_WORDS if index == len(THAI_GENERATED) - 1 else None)
+        for index, token_id in enumerate(THAI_GENERATED)
+    ] + [
+        (1, [token_id], tllm.FinishReason.LENGTH if index == len(sibling) - 1 else None)
+        for index, token_id in enumerate(sibling)
+    ]
+    first, second = result._outputs
+    first_diffs_after_stop = []
+    for step, (sequence_index, chunk, finish_reason) in enumerate(steps):
+        is_final = step == len(steps) - 1
+        result._handle_response(
+            _response(
+                chunk,
+                is_final,
+                finish_reason or tllm.FinishReason.NOT_FINISHED,
+                sequence_index=sequence_index,
+            )
+        )
+        if sequence_index == 1:
+            first_diffs_after_stop.append(first.text_diff)
+
+    assert first.text == "ข้อ ๑ ข้อ "
+    assert first.finish_reason == "stop"
+    assert first_diffs_after_stop == [""] * len(sibling)
+    assert full_decodes == [THAI_GENERATED[:-3]]
+    assert second.text == "ข้อ <answer>42"
+    assert second.finish_reason == "length"