Learning Graph Rationales to Compress Long Chains of Thought in Multimodal Reasoning

WhisperRec: Latent Reasoning for Efficient Foundation Recommendation Models

Large language models (LLMs) have demonstrated strong reasoning capabilities, motivating their adoption as backbones for foundation recommendation models (FRMs). Existing approaches typically enhance recommendation with explicit Chain-of-Thought (CoT) under the Think-then-Answer paradigm. However, generating lengthy rationales introduces substantial inference overhead, while fixed CoT templates struggle to model diverse, dynamic, and context-dependent user interests. We propose WhisperRec, an efficient latent reasoning framework for FRMs. WhisperRec compresses teacher-generated CoT into learnable latent reasoning tokens, enabling a Latent-Reason-then-Answer paradigm that performs reasoning in latent space without producing verbose rationales. This design retains decision-relevant reasoning information while avoiding the latency bottleneck of autoregressive rationale generation. Specifically, it first introduces Multi-View Adaptive CoT (MV-ACoT) to construct diverse, high-quality supervision from complementary perspectives on user interests. MV-ACoT also adapts reasoning complexity to each instance, applying lightweight analysis to clear cases and targeted multi-factor reasoning to challenging ones. Building on a pre-trained FRM, WhisperRec then employs a three-stage Latent Reasoning Alignment procedure to progressively internalize teacher CoT into latent representations. Finally, curriculum-based post-training activates latent-token reasoning for downstream recommendation while preserving standard recommendation capability. Experiments on an industrial-scale Kuaishou dataset and the public Kuaishou LLM-Rec benchmark show that WhisperRec consistently outperforms explicit-CoT methods and conventional baselines. Compared with explicit CoT Think and No-Think variants, WhisperRec improves SID@64 by 17.44% and 9.33%, respectively, and achieves over 10x higher online inference throughput.

MuSe: Multi-Stage Graph Reasoning via Vision-Language Models

Graph-related tasks are traditionally addressed with Graph Neural Networks (GNNs) or graph transformers, but their task-specific training limits generalization. Large Language Models (LLMs) offer stronger generalization, yet encoding graphs as one-dimensional text struggles to capture multi-hop dependencies and two-dimensional topology. Vision-Language Models (VLMs) provide an alternative by visualizing graphs, but rendering large graphs in a single image causes clutter, occlusion, and distraction, hindering reasoning. We propose MuSe, a novel multi-stage graph reasoning framework based on VLMs. Instead of processing entire graphs at once, MuSe incrementally samples and visualizes task-relevant subgraphs, enabling progressive reasoning. The framework employs a two-stage training paradigm: supervised fine-tuning to acquire local sampling and reasoning skills, followed by reinforcement learning with GRPO to refine the sampling strategy and control dialog length. To support evaluation, we introduce LGVLQA, a new multimodal dataset with larger and more complex graph structures, addressing the scalability limitations of existing benchmarks. Experiments show that MuSe consistently out-performs leading LLM and VLM baselines, demonstrating improved structural understanding and reasoning ability. Our code and data are available at this url.

HyGRL: Adaptive Hybrid Graph Reasoning for Multi-Entity Questions

Multi-entity compositional questions pose significant challenges to existing retrieval-augmented language models. Conventional methods fall into a dilemma: standard RAG lacks dynamic reasoning, traditional Graph-RAG is limited by structural sparsity, and LLM-constructed Graph-RAG incurs prohibitive costs. We propose \textbf{\fwa}, a unified framework that embeds unstructured text into structured knowledge graphs, creating a heterogeneous network for flexible evidence retrieval. Reasoning is formulated as adaptive structure induction, learned via a robust two-stage process: (1) imitation learning distills heuristic expert signals, and (2) reinforcement learning refines the policy using LLM-driven preference rewards. Experiments demonstrate that {\fwa} effectively merges textual richness with structural knowledge, outperforming SOTA baselines in answer accuracy and reasoning fidelity while maintaining extremely low token costs and near real-time inference((code available at https://github.com/wjywjy123/HyGRL) .

A cost-effective approach for knowledge graph reasoning path retrieval and enhanced large language model reliability

Large Language Models (LLMs) often face challenges in performing reliable multi-hop reasoning due to issues such as incomplete evidence chains and hallucinations. Incorporating knowledge graphs (KGs) can mitigate these problems, but existing approaches either suffer from suboptimal accuracy or are computationally expensive. To address these issues, we propose Reasoning Path Retrieval for RAG (RPR-RAG), a novel KG-based retrieval framework that incrementally builds a subgraph from the knowledge graph, extracts explicit reasoning paths, and provides them as structured external evidence to downstream LLMs. The experimental results on WebQuestionsSP (WebQSP) and Complex WebQuestions (CWQ) indicate that RPR-RAG achieves competitive Hit and F1 in multi-hop reasoning tasks, while maintaining runtime, LLM call frequency, and token usage at reasonable levels. Moreover, without additional task-specific training, RPR-RAG also shows strong zero-shot performance on MetaQA. RPR-RAG is built on a lightweight embedding model which can be trained and executed on a single consumer-grade GPU ( e.g ., RTX 3060, 6 GB). Ablation studies reveal that the path validity evaluation and stopping criterion play important roles in retrieval quality and efficiency. RPR-RAG is compatible with a range of backbone LLMs, from smaller 7B models to larger models such as GPT-5, providing a practical and interpretable framework for KG-grounded reasoning tasks. The source code is available at https://doi.org/10.5281/zenodo.19334059 .

Hybrid Knowledge Graph and Large Language Model Architectures for Predictive Analytics

Artificial Intelligence has significantly advanced predictive analytics, but traditional machine learning and deep learning models often struggle to integrate structured knowledge and provide explainable reasoning. Hybrid Knowledge Graph–Large Language Model (KG–LLM) architectures address these limitations by combining the structured semantic representation of Knowledge Graphs with the contextual reasoning capabilities of LLMs. This paper reviews hybrid KG–LLM frameworks for predictive analytics, highlighting graph embeddings, Retrieval-Augmented Generation (RAG), transformer-based reasoning, and contextual embedding fusion to improve prediction accuracy, interpretability, and robustness. The framework supports applications including healthcare, finance, cybersecurity, manufacturing, and customer analytics. Performance is evaluated using standard metrics such as Accuracy, Precision, Recall, F1-Score, AUC, and MAE, demonstrating superior results over standalone approaches. The paper also discusses scalability, computational challenges, explainable AI, federated learning, multimodal knowledge graphs, and future directions for trustworthy AI-driven predictive analytics.

Streamlining Long-Chain Reasoning via Differentiable Hierarchical Fusion