Theory of Mind (ToM) is widely regarded as central to effective persuasion, yet existing evaluations often fail to capture the infer–apply loop that arises in real-world dialogue. We introduce Theory-of-Mind-Guided Elaboration-Likelihood Persuasion (ToMELP), a benchmark that jointly conditions on the audience persona p and the Elaboration Likelihood Model (ELM) route r ∈ {central, peripheral} within persuasive conversations. The benchmark tests whether large language models can perform ToM inference over multi-turn interactions and leverage these inferences for controllable persuasive generation. ToMELP provides a structured interface with evidence annotations, enabling automated evaluation of persuasive effectiveness, route alignment/deviation, evidence quality under the central route, and robustness to perturbations.

Long-form narrative generation requires maintaining multi-dimensional consistency over extensive contexts, covering long-term constraints in character settings, event causality, spatial rules, and worldviews. Existing methods compress these heterogeneous elements into homogeneous representations, creating a topological mismatch that lowers retrieval signal-to-noise ratios and induces long-range logical collapse. We propose Topo-Narrator, a framework built on two complementary decouplings. Heterogeneous Topological Memory (HTM) organizes narrative state into a Character Graph, a Plot Chain, and an Environment Tree, providing topology-specific inductive biases. Plan-Reason-Write (PRW) multi-agent collaboration separates high-level reasoning, memory retrieval, and stylistic realization across dedicated agents. Experiments on MoPS and WritingPrompts show that Topo-Narrator consistently generates coherent narratives averaging over 13,000 words and substantially outperforms strong baselines on HANNA quality and fine-grained consistency metrics, validated by three frontier LLM judges and human evaluation.