Document-level relation extraction (DocRE) requires integrating scattered evidence to uncover challenging implicit relations. While recent LLM-based approaches leverage rich prior knowledge for relation prediction, they suffer from two limitations. Pipeline-based approaches cause irrecoverable candidate pruning, and the absence of explicit logical constraints leaves spurious over-prediction unsuppressed. To address this, we propose Gated Logical-rule INTernalization (GLINT), a novel DocRE framework that integrates an LLM with end-to-end inductive rule reasoning within a fully differentiable architecture. GLINT preserves per-pair signals across all entity pairs through a Fused Fact Tensor with Gumbel-softmax selection, and refines LLM predictions via an inductive rule reasoner with per-atom learnable negation gates. Extensive experiments on three benchmark datasets show that GLINT outperforms existing LLM-based baselines.

We propose a novel topic-guided coherence modeling (TGCM) for sentence ordering. Our attention based pointer decoder directly utilize sentence vectors in a permutation-invariant manner, without being compressed into a single fixed-length vector as the paragraph representation. Thus, TGCM can improve global dependencies among sentences and preserve relatively informative paragraph-level semantics. Moreover, to predict the next sentence, we capture topic-enhanced sentence-pair interactions between the current predicted sentence and each next-sentence candidate. With the coherent topical context matching, we promote local dependencies that help identify the tight semantic connections for sentence ordering. The experimental results show that TGCM outperforms state-of-the-art models from various perspectives.