Suspense Atlascomputational literary science

Generative-AI experiment · RQ5 · a separate experiment

Do generated stories follow the suspense trajectory they were asked for?

Twenty short stories were generated by an instruction-tuned language model under five explicit trajectory instructions (low, rising, early peak, late climax, repeated peaks), four premises each, then measured by the identical pipeline. These stories are never mixed into the human corpus.

01

Requested vs measured

For each trajectory instruction, the four generated stories’ measured curves (z-scored, smoothed) are drawn over the idealized target shape (dashed). Agreement is the Spearman ρ between the measured curve and the target; for the “low” instruction, which has no shape, the raw within-story standard deviation is reported instead (lower = flatter).

low · n=4 · mean raw SD 0.77 · mean peak at 70%

0%25%50%75%100%-101narrative progressz

a lighthouse keeper on… ρ=0.89, peak 74% · two siblings clear out… ρ=0.71, peak 64% · a night-shift nurse on… ρ=0.60, peak 67% · a botanist tracking an… ρ=0.86, peak 74%

rising · n=4 · mean ρ with target 0.55 · mean peak at 71%

0%25%50%75%100%-101narrative progressz

a lighthouse keeper on… ρ=0.61, peak 59% · two siblings clear out… ρ=0.52, peak 74% · a night-shift nurse on… ρ=0.62, peak 74% · a botanist tracking an… ρ=0.46, peak 74%

early peak · n=4 · mean ρ with target -0.25 · mean peak at 71%

0%25%50%75%100%-10123narrative progressz

a lighthouse keeper on… ρ=-0.34, peak 100% · two siblings clear out… ρ=-0.15, peak 56% · a night-shift nurse on… ρ=-0.17, peak 74% · a botanist tracking an… ρ=-0.34, peak 54%

late climax · n=4 · mean ρ with target 0.44 · mean peak at 85%

0%25%50%75%100%-2-10123narrative progressz

a lighthouse keeper on… ρ=0.74, peak 92% · two siblings clear out… ρ=0.33, peak 82% · a night-shift nurse on… ρ=0.44, peak 90% · a botanist tracking an… ρ=0.27, peak 74%

repeated peaks · n=4 · mean ρ with target -0.04 · mean peak at 65%

0%25%50%75%100%-101narrative progressz

a lighthouse keeper on… ρ=0.06, peak 64% · two siblings clear out… ρ=-0.33, peak 77% · a night-shift nurse on… ρ=0.27, peak 82% · a botanist tracking an… ρ=-0.14, peak 36%

02

Generated vs human-written dynamics

Two comparisons, on the primary measure. Left: how much the raw rating varies within a story — human stories vs generated stories (a flat generated story has low variance regardless of instruction). Right: the human corpus’ genre curves for reference.

1.07human within-story SD · LLM suspense (Qwen3-8B)
0.58human within-story SD · Composite lexical proxy
0.93generated within-story SD · LLM suspense (Qwen3-8B)
0.53generated within-story SD · Composite lexical proxy
Human-corpus quantiles of shape descriptors (primary measure) for reference; generated per-story values are in section 01.
descriptorq0.1q0.25q0.5q0.75q0.9
peak_position0.410.640.830.971.00
peak_height1.041.191.381.641.95
trend_slope0.270.641.341.712.00
n_peaks1.102.003.003.004.00
late_minus_early-0.040.420.861.181.33
final_level-0.310.050.460.981.41
roughness0.200.240.270.300.33
0%25%50%75%100%-1-0.500.51narrative progressz (within story), smoothed
all storiesadventure (n=11)comic (n=11)detective (n=10)ghost (n=13)horror (n=18)literary (n=15)speculative (n=14)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

03

Caveats specific to this experiment

  • The generator (Yuu no Sekai) and the annotator are the same model family; a shared notion of “suspense” could inflate agreement. The lexical proxy is the independent check.
  • Twenty stories is a pilot-scale sample; per-trajectory means rest on four stories.
  • Generated stories are ~1,400 words; the human corpus spans 1.2k–43k. Comparisons are on normalized progress and within-story variability only.