Technology flow (TF) forecasting aims to predict whether a technology will be utilized by another, which is significant for future cross-integration technology innovation. Although existing methods have made substantial progress in intra-domain TF forecasting, inter-domain TF forecasting remains underexplored and faces several fundamental challenges. Specifically, semantic gaps between technologies from different domains hinder effective representation learning; historical inter-domain TFs are highly sparse, limiting the ability to learn reliable structural dependencies; and intra-domain and inter-domain TFs exhibit distinct flow patterns, which are difficult to model within a unified framework. To address these challenges, in this paper, we propose an LLM-enhanced Graph Contrastive Learning ( LGCL ) framework that supports both intra-domain and inter-domain TF forecasting. Specifically, we first leverage LLMs with bias-aware representation adjustment to generate semantically enriched and accurate technology representations to bridge gaps across multiple domains. Then, to mitigate the sparsity of inter-domain TF data, a technology co-occurrence view is incorporated to complement historical TF information. We further develop a non-negative contrastive learning module to integrate the TF and co-occurrence views, enabling effective representation alignment without relying on negative samples. In addition, we design decoupled loss functions to separately supervise intra-domain and inter-domain flows, mitigating the difficulty of jointly modeling their different flow patterns under a unified objective. Finally, experiments on real-world datasets validate the effectiveness of our approach and provide valuable insights into future technology flow trends. Our code is available at https://github.com/huijieliu2023/LGCL .
Long-context inference is costly and can be brittle for large language models (LLMs): self-attention grows quadratically with input length, and retrieval-augmented generation (RAG) may introduce redundant context that obscures decisive evidence. Existing compressors face a representational trade-off. Hard compression keeps inspectable source tokens but can remove document-level cues needed for disambiguation, whereas soft compression preserves compact semantic summaries but may omit fine-grained facts. We formulate compression as allocating a constrained context budget across these complementary roles and propose Hybrid Context Compression (HyCo 2 ). HyCo 2 combines instruction-aware soft tokens from a hybrid adapter with original tokens selected by a lightweight evidence pathway. Ablations support pathway-specific reconstruction and completion pretraining before joint instruction tuning with self-distillation. Across seven knowledge-intensive QA benchmarks and three frozen LLM families, HyCo 2 attains QA accuracy comparable to the strongest extractive compressor while using a shorter compressed context, a smaller resident compressor, and lower measured pipeline cost. Relative to uncompressed RAG, it reduces the average decoder-context length by 88.8% and matches or exceeds accuracy on several dataset–backbone pairs. On LongBench, HyCo 2 obtains 22.03 under a shared 2,000-token source cap, within 0.29 points of LongLLMLingua with its 7B compressor. 1
Legal document retrieval is a core component of intelligent legal systems. In practice, candidate documents often originate from multiple law codes, each with distinct structures, formats, and textual styles. However, most existing retrieval methods assume all documents belong to a single law code, making them ill-suited to handle the structural diversity and textual inconsistencies inherent in multi-law-code (multi-LC) scenarios. This mismatch leads to degraded retrieval performance in practical deployments. To address this, we propose UniGEAR, a generative retrieval framework for multi-LC legal scenarios. UniGEAR encodes candidate documents with unified hierarchical identifiers that capture both structural and semantic information across heterogeneous law codes. Retrieval is performed via a sequence-to-sequence generation process guided by these identifiers. To construct the identifiers, we design a Residual Quantization Variational Autoencoder (RQ-VAE) enhanced with contrastive learning to align document semantics with legal code structures. For queries, large language models generate structured, multi-step chain-of-thought summaries, producing consistent representations for diverse query styles. A constrained decoding mechanism ensures generated identifiers remain valid within the legal hierarchy, improving precision and interpretability. Experiments on multi-LC benchmarks show UniGEAR significantly outperforms state-of-the-art methods in accuracy and efficiency, while maintaining strong performance in single-LC settings, demonstrating its adaptability and practical value.
Sequential recommendation generates personalized item rankings by estimating latent user-preference distributions from historical interactions. Although Diffusion Models (DMs) have gained prominence due to their exceptional capacity to capture complex distributions, current DM-based recommenders exhibit a critical limitation: Distributional Misalignment , where mode-seeking denoising processes disproportionately allocate probability mass to popular items, thereby diverging from users’ genuine preferences. To bridge this gap, we introduce ScoreRec , a novel sequential recommendation framework that reformulates the recommendation task as a continuous-time score-based generative process. Unlike conventional discrete denoising diffusion probabilistic models (DDPMs) that rely on fixed timestep approximations, ScoreRec models the diffusion process via a Stochastic Differential Equation. This formulation allows ScoreRec to estimate the gradient of the log-density with infinite temporal resolution, effectively reallocating probability mass to low-frequency yet personally relevant items and thereby mitigating the mode-seeking bias. Specifically, ScoreRec leverages a stochastic differential equation with a continuous-time noise schedule, ensuring that the noise level is dynamically adjusted to the local embedding density. During inference, an adaptive-step-size solver integrates the corresponding probability-flow ordinary differential equation to deterministically generate recommendations, striking a principled balance between accuracy and efficiency. Comprehensive experiments on three public benchmarks demonstrate that ScoreRec consistently surpasses eleven state-of-the-art baselines in both accuracy and diversity. Our code is publicly available at: https://github.com/AONE-NLP/ScoreRec .
Graph contrastive learning (GCL) is a promising approach for improving collaborative filtering (CF), especially in sparse interaction settings. Despite their effectiveness, most existing GCL-based recommendation methods risk disrupting meaningful collaborative signals by relying on uncontrollable graph augmentations. Besides, these methods fail to explicitly capture and integrate diverse fine-grained preferences because they model user–item interactions at a coarse level. To address these issues, we propose Prompt-Guided Graph Contrastive Learning (PGCL) , a novel framework that introduces learnable prompt nodes to construct controllable augmented graphs and enable fine-grained preference modeling. Specifically, these prompt nodes serve as latent fine-grained preferences and can be dynamically connected to user and item nodes in the graph based on semantic relevance. With this prompt-guided augmented graph, PGCL then incorporates two modules: the prompt-guided and the interaction-enhanced contrastive learning. The former adaptively aligns users and items with their most semantically relevant prompt nodes, enabling the model to capture and distinguish multi-dimensional preference signals. The latter incorporates real user–item interaction pairs as supervision, mitigating potential bias from prompt semantics and reinforcing collaborative signal learning. Together, these designs enable PGCL to model fine-grained preferences via a more semantically guided augmentation process. Extensive experiments across five public datasets demonstrate that PGCL consistently outperforms state-of-the-art baselines, thereby validating its effectiveness for recommendation tasks.
