Before the story is written, event types are decodable: four open-weight models look ahead at least four events, but only the next one beats text-only prediction

Before a story is written out, the model already carries information about what kind of event comes next, and that information reaches beyond a single event. Earlier work on hidden states recovered information about subsequent tokens or response-level attributes such as response length, reasoning steps, or character choices, rather than the events that change characters, relationships, or the story world. Four 24–32B open-weight instruction-tuned models each generated 500 stories, yielding 23,355 annotated events; probing hidden states at event ends gives next-event Macro-AP of .701 for OLMo, .643 for Qwen, .754 for Gemma, and .656 for Mistral, with gains over constant training prevalence of .229, .213, .282, and .190, all significant after within-model Holm correction.

Decodability does not mean hidden states predict better than reading the same text; the advantage is concentrated on the immediately next event. Text-only predictors use the same character cutoff prefix as the activation and the same story splits, targets, horizons, and probe-fitting procedure. Against the larger of full-prefix and chunk-averaged embedding Macro-AP, activation gains at the next event are .090 for OLMo, .041 for Qwen, .081 for Gemma, and .081 for Mistral, with all eight pairwise comparisons significant after Holm correction in the shared 120-test family; no horizon satisfies positive, significant gains against both primary text predictors.

In OLMo, sustained intervention along a direction derived from the next-event SOCIAL probe changes the proportion of SOCIAL events in generated stories in both intended directions. Gemma shows comparably strong next-event SOCIAL decodability but no comparable bidirectional control under the same protocol, with both directions reducing SOCIAL scores and widespread degeneration at higher strengths. With 100 seed-matched stories per condition at each strength, positive steering at strength 16 raises the mean per-story SOCIAL proportion by 17.13 percentage points (95% CI [12.45, 21.88]) and negative steering lowers it by 24.36 points; at strength 32 the changes are +26.07 and −32.70 points, and all four contrasts are significant after within-strength Holm correction.

The study covers stories generated by four 24–32B open-weight instruction-tuned models under a single minimal English prompt, Write a long story, and one sampling temperature; events and labels come from a single proprietary LLM annotator, validated against human judgments only on OLMo-generated stories. It is aimed at researchers studying internal states and controllability in long-form generation; extending the findings to other prompts, genres, model scales, or human-authored narratives would require separate validation.

Event-type decodability generally declines with distance, but only 5–12 stories remain at each model's largest eligible horizon, so the curves do not identify a sharp boundary beyond which predictive information disappears. The text comparison concerns the specified representations and linear readouts and does not establish that the observed text lacks equivalent predictive information. Steering was tested only in OLMo, targeting SOCIAL events at one selected layer throughout generation; higher strengths bring generation-quality costs, and effects extend to INTERNAL and WORLD-STATE, so this is not selective control over individual future events.