V8 Workflow

Audio-novel closed loop: interaction events → Adaptation Brief → voice/images/video → Simulator / Judge.

Workflow 保留 v7 的 belief-guided personalization，但输出从纯文本扩展到有声小说：Frozen Story Generator、TTS Voice Model、Image Generator、Video Composer 和 NetEase-style Player 都被纳入评价闭环。

Data construction

500 user simulator records

每条记录是一个有声小说播放回合：story/audio context + comments, likes, favorites, fast-forward 等统一 interaction events + hidden_user_state_after + tts constraints。

Data construction

500 planner records

每条 planner 样本把事件聚合成 interaction_summary，再输出 Adaptation Brief、candidate_budget、test_time_scaling_trace 和 DPO/GRPO training target。

Budgeted candidate expansion

test-time scaling

Budgeted candidate expansion: B=1/2/4/8/16/32，扩展 story/audio/image/video directives，使用 frozen Simulator / Judge 评分。No-Train 和 Train-Time Controller 共用同一评测。

TTS-NCFU = E[J_audio(G(S_t ⊕ ΔS_model), q) - J_audio(G(S_t), q)] / (E[J_audio(G(S_t ⊕ ΔS_oracle), q) - J_audio(G(S_t), q)] + epsilon)

Planner Utility@B = E[max_{c in Candidates(B)} ScoreAggregator(c; UserSim, RubricJudge)]

Audio Engagement Return = completion + like + favorite + relisten + comment_fit - fast_forward - voice_mismatch

Budgeted candidate expansion = sample B briefs, plans, voice directives, image prompts, then select argmax scorer return