# Novel Feedback v4 research audit

This audit turns the v4 request into concrete benchmark and encoder-design constraints. The direction remains personalized user-feedback-driven generation: feedback changes dynamic reader/user state, state changes future direct generation, and every artifact keeps `uses_reranking=false`.

## Scope guard

- No protected Qidian reader-response scraping: Qidian/WebNovel comments and replies require a Data Transfer Agreement. V4 keeps a DTA-ready adapter and uses public/metadata/synthetic/pilot schemas only until authorized access exists.
- No personalized reranking: the v4 agent writes one state-conditioned plan/chapter excerpt. Reward models can judge and diagnose, but the generation policy is not `generate K -> rerank`.
- Metrics must explain causal future effects, state-diff quality, story/canon consistency, and simulator fidelity rather than only showing aggregate return.

## Personalized reward / preference learning

- **LaMP: When Large Language Models Meet Personalization** — https://arxiv.org/abs/2304.11406
  Implication: Use static-profile, full-history and retrieval personalization baselines; report lift over generic generation.
- **PRISM Alignment Dataset** — https://arxiv.org/search/?query=PRISM+Alignment+Dataset+individualized+human+feedback&searchtype=all
  Implication: Model plural user-specific preference, not a single universal reward.
- **P-GenRM / Personalized Generative Reward Models** — https://arxiv.org/search/?query=P-GenRM+Personalized+Generative+Reward+Model&searchtype=all
  Implication: Condition reward/critic on user history and dynamic state; evaluate user-swap separation.

## User simulator and dynamic state

- **Generative Agents: Interactive Simulacra of Human Behavior** — https://arxiv.org/abs/2304.03442
  Implication: Separate memory, reflection/planning, and behavior emission; simulator should update latent state.
- **ProPerSim / proactive personalized user-assistant simulation** — https://arxiv.org/search/?query=ProPerSim+personalized+user+simulator+LLM&searchtype=all
  Implication: Use stateful user-assistant rollouts rather than static persona prompting.
- **PrefDisco / preference discovery** — https://arxiv.org/search/?query=PrefDisco+preference+discovery+LLM&searchtype=all
  Implication: Add active preference discovery and value-of-information metrics.
- **Swap-guided preference learning** — https://arxiv.org/search/?query=%22swap-guided%22+preference+learning+personalized&searchtype=all
  Implication: Use same-text/different-user and same-user/different-content hard negatives.

## Novel consistency and long-horizon coherence

- **ROCStories / Story Cloze Test** — https://aclanthology.org/N16-1098/
  Implication: Short causal story-ending checks are a minimal local coherence control.
- **WritingPrompts / Hierarchical Neural Story Generation** — https://arxiv.org/abs/1805.04833
  Implication: Open-ended reference metrics are weak; use human/LLM pairwise and axis-specific rubrics.
- **PlotMachines: Outline-Conditioned Generation with Dynamic Plot State Tracking** — https://arxiv.org/abs/2004.14967
  Implication: Maintain explicit plot/story state and test plan adherence.
- **STORIUM: Dataset and Evaluation Platform for Machine-in-the-Loop Story Generation** — https://aclanthology.org/2020.emnlp-main.525/
  Implication: Evaluate continuations against structured character/location/plot constraints.
- **Re3: Generating Longer Stories with Recursive Reprompting and Revision** — https://arxiv.org/abs/2210.06774
  Implication: Long-form generation needs recursive revision and long-horizon coherence checks.
- **DOC: Improving Long Story Coherence With Detailed Outline Control** — https://arxiv.org/abs/2212.10077
  Implication: Score detailed-outline adherence separately from prose quality.
- **OpenMEVA story-generation metric benchmark** — https://github.com/thu-coai/OpenMEVA
  Implication: Automatic metrics need validation against human preference; avoid BLEU-only scoring.
- **PG-19 long-form language modeling corpus** — https://arxiv.org/abs/1911.05507
  Implication: Public-domain long novels support long-context smoke tests, though not preference labels.
- **NarrativeQA** — https://aclanthology.org/Q18-1023/
  Implication: QA over prior story state can probe whether generated chapters preserve canon.
- **QuALITY long-document QA** — https://aclanthology.org/2022.naacl-main.391/
  Implication: Long-context multiple-choice probes can be converted into chapter-continuity checks.
- **LitBank and BookNLP** — https://github.com/dbamman/litbank
  Implication: Character/entity extraction enables automated character consistency metrics.
- **FactCC / QAGS / FEQA / SummaC factual consistency family** — https://aclanthology.org/2020.emnlp-main.750/
  Implication: Adapt claim extraction + retrieval + NLI/QA to plot/canon contradictions.

## Encoder training design implications

V4 should train an encoder-state simulator with four supervised heads and two intervention probes:

1. **Behavior head**: continue/drop, dwell time, fast-swipe and comment propensity from `(profile, state_before, history, candidate text/axes)`.
2. **State-transition head**: next-state or state-delta for trust, fatigue, satisfaction, confidence and narrative-axis affinity.
3. **Comment/ordinal head**: type/intensity of feedback, allowing Likert labels instead of binary preference only.
4. **Counterfactual head/probe**: profile/content/fatigue swaps must change behavior in the expected direction.
5. **Prototype memory**: learn a small set of reader-prototype vectors for cold-start and user-swap separation.
6. **Story consistency features**: feed canon-risk, outline-adherence and entity/plot-state features into the simulator so engagement cannot be optimized by contradicting the novel.

The resulting training target is not “does the user like the current chapter?” only; it is “how does this interaction update reader state and future generation constraints?”