diff --git a/docling/backend/html_backend.py b/docling/backend/html_backend.py
index 82c44f2253..fa5779e326 100644
--- a/docling/backend/html_backend.py
+++ b/docling/backend/html_backend.py
@@ -124,6 +124,16 @@ def _warn_headers_without_origin() -> None:
 # Using Unicode Private Use Area to avoid conflicts with actual content
 _BR_SENTINEL = "\ue000"
 
+
+def _resolve_line_breaks(text: str) -> str:
+    """Replace the sentinels of <br> tags with a line break.
+
+    For text that is one paragraph (heading, caption, list item): consecutive <br>
+    tags give one line break.
+    """
+    return re.sub(rf" *{_BR_SENTINEL}[ {_BR_SENTINEL}]*", "\n", text).strip()
+
+
 DEFAULT_IMAGE_WIDTH = 128
 DEFAULT_IMAGE_HEIGHT = 128
 _MHTML_SYNTHETIC_BASE = "thismessage:/"
@@ -384,7 +394,7 @@ def to_single_text_element(self) -> AnnotatedText:
             current_code = c if c else current_code
 
         return AnnotatedText(
-            text=current_text.strip(),
+            text=_resolve_line_breaks(current_text),
             hyperlink=current_h,
             formatting=current_f,
             code=current_code,
@@ -2871,9 +2881,9 @@ def _add_list_item_with_content(
             tag, ignore_list=True, find_parent_annotation=True
         )
         min_parts = parts.simplify_text_elements()
-        item_text = re.sub(
-            r"\s+|\n+", " ", "".join([el.text for el in min_parts])
-        ).strip()
+        item_text = _resolve_line_breaks(
+            re.sub(r"\s+|\n+", " ", "".join([el.text for el in min_parts]))
+        )
 
         if not item_text:
             return None
@@ -2895,7 +2905,12 @@ def _add_list_item_with_content(
             with self._use_inline_group(min_parts, doc):
                 compacted_parts = self._compact_adjacent_single_char_parts(min_parts)
                 for annotated_text, source_tag_ids in compacted_parts:
-                    text_part = re.sub(r"\s+|\n+", " ", annotated_text.text).strip()
+                    text_part = _resolve_line_breaks(
+                        re.sub(r"\s+|\n+", " ", annotated_text.text)
+                    )
+                    if not text_part:
+                        # The part only held <br> tags
+                        continue
                     clean_text = HTMLDocumentBackend._clean_unicode(text_part)
 
                     # Apply extra formatting if provided
@@ -2940,7 +2955,7 @@ def _add_list_item_with_content(
         else:
             # Simple content - single text element
             annotated_text = min_parts[0]
-            text = re.sub(r"\s+|\n+", " ", annotated_text.text).strip()
+            text = _resolve_line_breaks(re.sub(r"\s+|\n+", " ", annotated_text.text))
             clean_text = HTMLDocumentBackend._clean_unicode(text)
             prov = self._make_text_prov(
                 text=clean_text,
@@ -3460,7 +3475,12 @@ def _handle_block(self, tag: Tag, doc: DoclingDocument) -> list[RefItem]:  # noq
             language_hint = self._code_language_hint(tag)
             with self._use_inline_group(annotated_texts, doc) as inline_ref:
                 for annotated_text in annotated_texts:
-                    text_orig = annotated_text.text.strip()
+                    # In preformatted text, each <br> is a newline
+                    text_orig = re.sub(
+                        rf" ?{_BR_SENTINEL} ?", "\n", annotated_text.text
+                    ).strip()
+                    if not text_orig:
+                        continue
                     text_clean = HTMLDocumentBackend._clean_unicode(text_orig)
                     prov = self._make_prov(
                         text=text_clean,
@@ -3883,7 +3903,7 @@ def _extract_text_excluding_tag_obj_ids(
     ) -> str:
         def _extract(node: PageElement) -> list[str]:
             if isinstance(node, NavigableString):
-                return [str(node)]
+                return [str(node).replace(_BR_SENTINEL, "\n")]
             if isinstance(node, Tag):
                 if id(node) in excluded_obj_ids:
                     return []
diff --git a/tests/test_backend_html.py b/tests/test_backend_html.py
index 40502e5a91..4ba99db06c 100644
--- a/tests/test_backend_html.py
+++ b/tests/test_backend_html.py
@@ -2274,6 +2274,52 @@ def test_html_newline_handling():
     )
 
 
+def test_html_line_break_outside_paragraphs():
+    # A <br> outside a paragraph or a table cell, e.g. in a heading or a list
+    # item, left the internal sentinel (U+E000) in the text. It is now a line
+    # break, as in a paragraph. The text of a checkbox stays on one line.
+    src = (
+        b"<html><body>"
+        b"<h1>Title<br>line</h1>"
+        b"<h2>Heading<br>line</h2>"
+        b"<ul><li>Item<br>line</li><li><b>Bold</b><br><i>italic</i></li></ul>"
+        b"<dl><dt>Term<br>line</dt><dd>Definition<br>line</dd></dl>"
+        b"<table><caption>Caption<br>line</caption>"
+        b"<tr><td>A</td><td>B</td></tr></table>"
+        b"<figure><img src='a.png'><figcaption>Figure<br>line</figcaption></figure>"
+        b"<pre>Code<br>line</pre>"
+        b"<pre><b>bold</b><br><i>code</i></pre>"
+        b"<ul><li><input type='checkbox' checked>Task<br>line</li></ul>"
+        b"</body></html>"
+    )
+    in_doc = InputDocument(
+        path_or_stream=BytesIO(src),
+        format=InputFormat.HTML,
+        backend=HTMLDocumentBackend,
+        filename="t.html",
+    )
+    doc = HTMLDocumentBackend(in_doc=in_doc, path_or_stream=BytesIO(src)).convert()
+
+    assert [item.text for item in doc.texts] == [
+        "Title\nline",
+        "Heading\nline",
+        "Item\nline",
+        "",  # the list item with formatting: its text is in the two runs below
+        "Bold",
+        "italic",
+        "Term\nline",
+        "Definition\nline",
+        "Caption\nline",
+        "Figure\nline",
+        "Code\nline",
+        "bold",
+        "code",
+        "Task line",
+    ]
+    assert not any(_BR_SENTINEL in item.orig for item in doc.texts)
+    assert "## Heading line" in doc.export_to_markdown()
+
+
 def test_gfm_task_list_renders_checkbox_with_text():
     # <li><input type=checkbox>text</li> is the canonical GFM task list. The
     # text belongs to the checkbox item; it used to become a separate list