Controlled Text Generation
Give AI a writing plan and review what it produces.
Build a writing plan from reference material, then give AI the character details and earlier story events each scene needs. Review its draft against the plan and writing requirements.
TextMind
AI composition becomes a story
What is controlled generation
Tell the AI writing the next sentence where the whole story is going.
Controlled generation gives the writing AI its goals, references, structure and current context at every step. Instead of requesting a long text at once, it plans what to write in which order and checks each result before moving on.
- Writing materialReferences and story rules
- Scene planWhat this scene should do
- AI draftsPrior events and current context
- Review and reviseCheck the draft against the plan
Why it matters
A plan for direction
Break the overall story into episodes and scenes, separating the present task from later developments.
Context for continuity
Give the AI the established characters and facts so the current scene stays consistent with what came before.
Criteria for revision
Check structure, length and expression, then turn review findings into focused revisions.
The value of a writing process
The same generator. A different writing process.
Re³ compared direct continuation with a system combining planning, context and rewriting.
GPT-3-175B (davinci) · 100 English premises · 2,000–2,500 words · 1,024-token context · three evaluators per pair.
Coherence rose from 45.7% to 60.0%, a 14.3-point increase.
Share meeting the human-rated coherence criterion; this is not a pairwise win rate.
View the comparison data
| Configuration | Story coherence (%) |
|---|---|
| ROLLING · direct continuation | 45.7% |
| Re³ · planning, context and rewriting | 60.0% |
Yang et al. · Re³ · EMNLP 2022 · Table 1
Re³ uses additional calls and a trained reranker. The gain cannot be attributed to review alone.
External research, not a TextMind measurement or a comparison of current models.
TextMind supplies the writing AI with structure and context.
BLUEURBAN studies how reference records become usable writing material, from overall direction down to scenes and sentences. TextMind hands an external LLM the plan and current context, has it write, then checks the result against the generation requirements.
Research behind the technology
Research and application at BLUEURBAN
TextMind: prepare what AI needs to write, then review its draft.
For the web-novel service Taste Edition, TextMind organizes references into an overall storyline, episode plans, scene outlines and material for sentences. It gives these materials and established story details to an external LLM, then uses review findings to guide revisions.
A scene must respect what its characters know.
The protagonist does not yet know what the key is for.
AI
- Draft the scene using its plan and the preceding story.
- Revise the passage in response to review findings.
TextMind
- Supply current story rules and established facts as context.
- Review the draft for conflicts with the plan and story rules.
Illustrative review and revision
He already knew the secret of the key.
He turned the key over, trying to guess its purpose.
- Turn references into writing material
- Select and organize plot, structure and expression details from the reference database. Here, “distillation” means summarizing and condensing reference material, rather than training model weights.
- Assemble the context for each scene
- For each scene, give AI the overall storyline, characters already introduced, facts revealed in earlier episodes and the story rules currently in effect.
- Revise the problems found in review
- Apply structure and expression checks suited to the generation path, with a bounded number of retries. Review-guided repairs keep a record of the original failure and check the revised result again.
Frequently asked questions
Q1Does this build a new language model?
No. An external LLM writes the prose. TextMind prepares references, plans and context, hands them to the LLM and reviews what comes back.
Q2Is reference distillation model training?
It prepares reference records as writing structure and material; it does not distill model weights.
Q3Does the chart measure TextMind’s improvement?
It reports the Re³ study. TextMind needs evaluation with its own references, model and writing conditions.
Papers and resources
Start with what AI should write and what it must respect.
Review your references and writing process to identify where planning, context and review belong.
Discuss the writing workflow