diff --git a/backend/app/api/v1/literature.py b/backend/app/api/v1/literature.py
index b3b7f3d..3bfec9e 100644
--- a/backend/app/api/v1/literature.py
+++ b/backend/app/api/v1/literature.py
@@ -7,7 +7,7 @@ from datetime import datetime, timezone
 from fastapi import APIRouter, Depends, HTTPException, Query
 from fastapi.responses import PlainTextResponse
 from pydantic import BaseModel
-from sqlalchemy import case, func, literal, or_, select, text
+from sqlalchemy import case, exists, func, literal, or_, select, text
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from app.core.permissions import get_current_user
@@ -289,10 +289,11 @@ async def search_literature(
                 ).limit(100)
             )).scalars().all()
             if _tag_matches:
-                _tag_lit_subq = select(GlobalLiteratureTag.literature_id).where(
-                    GlobalLiteratureTag.tag_id.in_(list(_tag_matches))
+                _tag_exist = exists().where(
+                    GlobalLiteratureTag.literature_id == GlobalLiterature.id,
+                    GlobalLiteratureTag.tag_id.in_(list(_tag_matches)),
                 )
-                search_cond = or_(search_cond, GlobalLiterature.id.in_(_tag_lit_subq))
+                search_cond = or_(search_cond, _tag_exist)
         # COUNT：使用 LIMIT 10001 截断，避免大结果集的精确计数
         # 结果集 ≤10000 时返回精确值，否则返回 10001+
         MAX_EXACT = 10000
diff --git a/backend/app/services/query_expansion.py b/backend/app/services/query_expansion.py
index f68571b..021f401 100644
--- a/backend/app/services/query_expansion.py
+++ b/backend/app/services/query_expansion.py
@@ -20,7 +20,7 @@ import logging
 import re
 from uuid import UUID
 
-from sqlalchemy import or_ as _or_, select
+from sqlalchemy import exists, or_ as _or_, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from app.models.literature import GlobalTag, GlobalTagTreeNumber, GlobalLiteratureTag, GlobalLiterature
@@ -53,13 +53,11 @@ async def expand_atm(db: AsyncSession, query: str):
     if not expanded_ids:
         return None
 
-    # 不用 .distinct()：IN (subquery) 语义本身就按值去重，显式 DISTINCT 只会让 PG
-    # 先物化+排序海量 literature_id（'cancer' 这类词展开出 150+ 个 tag，命中的
-    # literature_id 可达数百万），再对 id IN (百万集合) 扫描。去掉后可直接走
-    # ix_glt_tag 索引半连接，year_counts 全量扫描和主查询都受益。
-    return GlobalLiterature.id.in_(
-        select(GlobalLiteratureTag.literature_id)
-        .where(GlobalLiteratureTag.tag_id.in_(expanded_ids))
+    # 不用 .distinct()：EXISTS 关联子查询按 literature_id 探测 PK 索引（每外行一次
+    # 索引查找），不走 ix_glt_tag 全量物化。
+    return exists().where(
+        GlobalLiteratureTag.literature_id == GlobalLiterature.id,
+        GlobalLiteratureTag.tag_id.in_(expanded_ids),
     )
 
 
diff --git a/backend/app/services/search_engine.py b/backend/app/services/search_engine.py
index b4aac32..85ca304 100644
--- a/backend/app/services/search_engine.py
+++ b/backend/app/services/search_engine.py
@@ -2190,19 +2190,20 @@ class AdvancedSearchEngine:
             return None
 
         uids = list(mesh_tag_ids)
-        # 不用 func.distinct：IN (subquery) 语义本身就按值去重，显式 DISTINCT 只让
-        # PG 多一次物化排序；去掉后可直接走 ix_glt_tag 索引半连接（year_counts 等
-        # 全量扫描场景更敏感）。
+        # 不用 .distinct()：EXISTS 关联子查询按 literature_id 探测 PK 索引（每外行一次
+        # 索引查找），不走 ix_glt_tag 全量物化。IN (subquery) 在 OR 过滤器里会被 planner
+        # 反复重扫（'cancer' 展开命中的 literature_id 达数十万，Materialize 循环数十次
+        # 直接 15s 超时），EXISTS 关联每次外行只做一次 PK 探测。
         if major_only:
-            subq = select(GlobalLiteratureTag.literature_id).where(
+            return exists().where(
+                GlobalLiteratureTag.literature_id == GlobalLiterature.id,
                 GlobalLiteratureTag.tag_id.in_(uids),
                 GlobalLiteratureTag.is_major == True,
             )
-        else:
-            subq = select(GlobalLiteratureTag.literature_id).where(
-                GlobalLiteratureTag.tag_id.in_(uids),
-            )
-        return GlobalLiterature.id.in_(subq)
+        return exists().where(
+            GlobalLiteratureTag.literature_id == GlobalLiterature.id,
+            GlobalLiteratureTag.tag_id.in_(uids),
+        )
 
     @staticmethod
     def _best_match_order(tsq):