Hypergraph-based methods extend structured retrieval-augmented generation (RAG) by organizing external knowledge as multi-entity relational structures, helping large language models (LLMs) generate better-grounded answers. However, existing structured RAG systems typically assume static corpus construction, requiring the retrieval structure to be rebuilt over the expanded corpus when new documents arrive. To address this limitation, we introduce HyperRAG-IHM, a hypergraph-based RAG framework that performs incremental hyperedge maintenance to incorporate newly arriving knowledge without reconstructing the historical retrieval structure. HyperRAG-IHM links relational evidence extracted from new documents to existing hypergraph knowledge, avoiding reconstruction of the entire retrieval structure. It uses LSH to localize candidate historical hyperedges, then applies multi-dimensional comparison and a three-action decision policy to integrate new relational evidence. Experiments show that HyperRAG-IHM achieves answer quality comparable to or better than full reconstruction under 5D scoring and 8D pairwise evaluation, and remains robust across ablation, structural behavior, cross-dataset, and cross-answer-model analyses.

While Chain-of-Thought (CoT) reasoning significantly enhances the performance of Multimodal Large Language Models (MLLMs), its autoregressive nature incurs prohibitive latency constraints. Current efforts to mitigate this via token compression often fail by blindly applying text-centric metrics to multimodal contexts. We identify a critical failure mode termed Visual Amnesia, where linguistically redundant tokens are erroneously pruned, leading to hallucinations. To address this, we introduce V-Skip that reformulates token pruning as a Visual-Anchored Information Bottleneck (VA-IB) optimization problem. V-Skip employs a dual-path gating mechanism that weighs token importance through both linguistic surprisal and cross-modal attention flow, effectively rescuing visually salient anchors. Extensive experiments on Qwen2-VL and Llama-3.2 families demonstrate that V-Skip achieves a speedup with negligible accuracy loss. Specifically, it preserves fine-grained visual details, outperforming other baselines over 30% on the DocVQA.