Generating long-form linear fiction from open-ended themes remains a major challenge for large language models, which frequently fail to guarantee global structure and narrative diversity when using premise-based or linear outlining approaches. We present BiT-MCTS, a theme-driven framework that operationalizes a "climax-first, bidirectional expansion" strategy motivated by Freytag's Pyramid. Given a theme, our method extracts a core dramatic conflict and generates an explicit climax, then employs a bidirectional Monte Carlo Tree Search (MCTS) to expand the plot backward (rising action, exposition) and forward (falling action, resolution) to produce a structured outline. A final generation stage realizes a complete narrative from the refined outline. We construct a Chinese theme corpus for evaluation and conduct extensive experiments across three contemporary LLM backbones. Results show that BiT-MCTS improves narrative coherence, plot structure, and thematic depth relative to strong baselines, while enabling substantially longer, more coherent stories according to automatic metrics and human judgments.
Figures & tables
Figure 1: Comparison of fiction outline generation methods: Sequential outline generation leads to overly formatic narratives, while bidirectional MCTS can generate diverse and creative outlines.
Figure 2: An overview of the four-stage fiction generation pipeline, which proceeds: (1) Conflict and Climax Generation establishes the core conflict and the core climax. (2) Bidirectional MCTS Exploration searches forward and backward from the climax for coherent and creative plot outlines. (3) Outline Refinement refines the rough outline generated by Bidirectional MCTS and segmented. (4) Segmented Fiction Generation expands the refined outline into the final, fluent narrative fiction. The pipeline transforms an abstract theme into a complete, structured long-form fiction. Note that the example English texts are translated from Chinese texts for better understanding.
Backbone LLM
Method
NC
CR
ER
PS
CD
SD
GR
FL
DI
OQ
Avg
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
DeepSeek-V3
Vanilla
0.0
0.0
0.0
0.0
0.0
0.0
7.5
2.5
0.0
0.0
1.0
StoryWriter
32.5
17.5
20.0
5.0
17.5
42.5
25.0
25.0
40.0
42.5
26.8
Narrative Studio
42.5
57.5
20.0
42.5
20.0
15.0
7.5
10.0
25.0
30.0
27.0
BiT-MCTS
25.0
25.0
60.0
52.5
62.5
42.5
60.0
62.5
35.0
27.5
45.3
GPT-5-Mini
Vanilla
5.0
12.5
5.0
2.5
2.5
10.0
10.0
7.5
2.5
2.5
6.0
Table 1: Win rates of different methods across ten dimensions when evaluated by LLM judge. NC, CR, ER, PS, CD, SD, GR, FL, DI, OQ represent narrative complexity, creativity, emotional resonance, plot structure, character development, setting description, grammaticality, fluency, diversity, and overall quality, respectively.
Backbone LLM
Average Fiction Length (tokens)
StoryWriter
Narrative Studio
BiT-MCTS
deepseek-v3
5904.35
4239.20
8059.55
gpt-5-mini
15008.10
3530.27
58657.00
gemini-2.5-flash
8438.75
1283.13
25374.10
Table 2: Average fiction length (in tokens) generated by different methods across different backbone LLMs.
Method
NC
CR
ER
PS
CD
SD
GR
FL
DI
OQ
TH
Avg
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
StoryWriter
29.17
30.83
32.49
17.49
19.16
26.66
25.38
22.04
27.88
14.99
33.33
26.23
Narrative Studio
38.33
37.50
21.67
35.84
30.00
24.17
34.81
32.28
29.81
35.84
15.00
29.95
BiT-MCTS
32.50
31.67
45.84
46.67
50.84
49.17
39.81
45.68
42.31
49.17
51.67
43.82
Table 3: Win rates of different methods across 11 dimensions when evaluated by human judges. NC, CR, ER, PS, CD, SD, GR, FL, DI, OQ, TH represent narrative complexity, creativity, emotional resonance, plot structure, character development, setting description, grammaticality, fluency, diversity, overall quality, and theme, respectively.
Configuration
NC
CR
ER
PS
CD
SD
GR
FL
DI
OQ
Avg
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
Beam Search
61.25
61.25
66.25
61.25
66.25
63.75
63.75
68.75
71.25
71.25
65.5
Best of N
56.25
68.75
66.25
56.25
73.75
53.75
73.75
73.75
96.25
76.25
69.5
- MCTS(direct)
100
100
90
100
87.5
100
35
30
92.5
100
93.5
- Refinement
100
100
100
100
100
87.5
82.5
80
90
100
94
- Bidirectional
100
100
100
100
100
100
100
100
100
100
100
Table 4: Ablation study on search strategy variants and key components of BiT-MCTS( loss rate against the complete method)
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
LLM Judge
Approach
NC
CR
ER
PS
CD
SD
GR
FL
DI
OQ
Avg
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
DeepSeek-V3
Vanilla
87.5
97.5
62.5
70.0
97.5
92.5
45.0
45.0
92.5
90.0
78.0
StoryWriter
5.0
0.0
25.0
17.5
0.0
2.5
30.0
30.0
0.0
2.5
11.3
Narrative Studio
5.0
0.0
7.5
12.5
0.0
2.5
22.5
22.5
0.0
2.5
7.3
BiT-MCTS
2.5
2.5
5.0
0.0
2.5
2.5
2.5
2.5
7.5
5.0
3.3
GPT-5-Mini
Vanilla
75.0
75.0
50.0
45.0
75.0
75.0
25.0
25.0
70.0
62.5
57.2
Appendix
Table 5: Loss rates (probability of being judged as the worst) of different approaches across ten dimensions when evaluated by LLMs. NC, CR, ER, PS, CD, SD, GR, FL, DI, OQ represent narrative complexity, creativity, emotional resonance, plot structure, character development, setting description, grammaticality, fluency, diversity, and overall quality, respectively.
LLM Judge
Against
NC
CR
ER
PS
CD
SD
GR
FL
DI
OQ
Avg
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
DeepSeek-V3
Vanilla
100.0
95.0
92.5
100.0
95.0
100.0
75.0
75.0
100.0
100.0
93.3
StoryWriter
55.0
55.0
75.0
45.0
70.0
55.0
82.5
82.5
65.0
65.0
65.0
Narrative Studio
35.0
25.0
65.0
50.0
65.0
70.0
95.0
95.0
85.0
65.0
65.0
GPT-5-Mini
Vanilla
100.0
80.0
90.0
90.0
90.0
70.0
70.0
70.0
100.0
100.0
86.0
StoryWriter
100.0
95.0
70.0
95.0
75.0
70.0
60.0
60.0
90.0
95.0
81.0
Appendix
Table 6: Win rates of our method (BiT-MCTS) against three baselines in pairwise comparisons across ten dimensions. NC, CR, ER, PS, CD, SD, GR, FL, DI, OQ represent narrative complexity, creativity, emotional resonance, plot structure, character development, setting description, grammaticality, fluency, diversity, and overall quality, respectively.
Model
Against
NC
CR
ER
PS
CD
SD
GR
FL
DI
OQ
Avg
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
(%)
DeepSeek-V3
StoryWriter
62.5
60.0
70.0
65.0
62.5
70.0
45.0
47.5
70.0
75.0
62.8
Narrative Studio
45.0
27.5
60.0
50.0
65.0
75.0
95.0
95.0
67.5
60.0
64.0
Appendix
Table 7: Win rates of BiT-MCTS against StoryWriter and Narrative Studio in pairwise comparisons across ten dimensions (DeepSeek-V3, controlled length 3000–5000 words).
Figure 3: Win rates of BiT-MCTS when generating stories of different lengths in pairwise comparisons across ten dimensions (DeepSeek-V3, short average length: 4669 words vs. long average length: 8059 words).