Story generation learns to plan reader suspense and humor: trajectory fidelity rises 12%–35% while coherence largely holds
Synopsis
Large language models write stories that often lack the rising and falling reader experience that makes narrative engaging. A new framework first learns how reader responses such as suspense or humor evolve in high-quality human-written stories, then uses a small model to predict the response the next segment should elicit and steers an off-the-shelf writer model with that prediction. Across suspense and humor, with GPT-5.5 and Qwen3.5-27B as writers, generated stories improve trajectory-based fidelity to human-written stories by 12%–35% while largely preserving story coherence.
Interpretation
Story continuation gains an explicit control signal: instead of asking only for fluent text, each segment is pushed toward a predicted reader response, so the rise and fall of suspense or humor can be planned. Prior attempts to teach models reader experience implicitly did little: one-shot in-context learning barely improved dynamics fidelity in most settings, with gains below 5% even where they occurred, and direct supervised fine-tuning on story text reached competitive fidelity in some settings but cost up to 30% coherence degradation. Across two reader-response domains, suspense and humor, with GPT-5.5 and Qwen3.5-27B as writer models, the framework improves trajectory-based fidelity to human-written stories by 12%–35% while largely preserving coherence; gains are strongest with GPT-5.5, while Qwen3.5-27B shows smaller gains and greater coherence sensitivity.
The method splits the job in two: predict how readers should respond to the next segment, then have the writer model write toward that target. Prediction is handled by a steering model, a Qwen3.5-27B fine-tuned with LoRA that reads the story prefix, the existing reader-response trajectory and auxiliary information such as planned length and genre, and outputs the next segment's reader response; the writer side is an off-the-shelf LLM with strong instruction following, and candidates are chosen by best-of-N sampling against the predicted response. Reader responses are scored by another off-the-shelf model, GPT-5.4-mini, at temperature 0; the main experiments represent suspense and humor as integers from 0 to 100 plus a one-sentence reason, with humor drawn from 4,012 FunnyBench stand-up clips and 25,712 segments, and suspense from 1,733 stories and 12,834 segments across FlawedFictions, Gutenberg and STORAL.
Perspective
Experiments cover two reader-response domains, suspense and humor, with GPT-5.5 and Qwen3.5-27B as writer models and reader responses represented as integers from 0 to 100; the measure model's validity is checked against ratings from 10 native English speakers rather than a full-scale human preference evaluation of generated stories.
Reader responses are subjective and vary across individuals and cultures, and LLM-based measurement does not perfectly match aggregated human judgments, being less reliable at distinguishing subtle differences; steering covers one dimension at a time, and whether multiple simultaneous steering signals can be followed, or would conflict, remains unclear; error accumulation in long-form generation also remains to be addressed.