Large language models (LLMs) augmented with retrieval exhibit robust performance and extensive versatility by incorporating external contexts. However, the input length grows linearly in the number of retrieved documents, causing a dramatic increase in latency. In this paper, we propose a novel paradigm named Sparse RAG, which seeks to cut computation costs through sparsity. Specifically, Sparse RAG encodes retrieved documents in parallel, which eliminates latency introduced by long-range attention of retrieved documents. Then, LLMs selectively decode the output by only attending to highly relevant caches auto-regressively, which are chosen via prompting LLMs with special control tokens. It is notable that Sparse RAG combines the assessment of each individual document and the generation of the response into a single process. The designed sparse mechanism in a RAG system can facilitate the reduction of the number of documents loaded during decoding for accelerating the inference of the RAG system. Additionally, filtering out undesirable contexts enhances the model’s focus on relevant context, inherently improving its generation quality. Evaluation results on four datasets show that Sparse RAG can be used to strike an optimal balance between generation quality and computational efficiency, demonstrating its generalizability across tasks.

Large Language Models (LLMs) have demonstrated impressive capabilities in creative tasks such as storytelling and E-mail generation. However, as LLMs are primarily trained on final text results rather than intermediate revisions, it might be challenging for them to perform text rewriting tasks. Most studies in the rewriting tasks focus on a particular transformation type within the boundaries of single sentences. In this work, we develop new strategies for instruction tuning and reinforcement learning to better align LLMs for cross-sentence rewriting tasks using diverse wording and structures expressed through natural languages including 1) generating rewriting instruction data from Wiki edits and public corpus through instruction generation and chain-of-thought prompting; 2) collecting comparison data for reward model training through a new ranking function. To facilitate this research, we introduce OpenRewriteEval, a novel benchmark covers a wide variety of rewriting types expressed through natural language instructions. Our results show significant improvements over a variety of baselines.

Reinforcement learning (RL) can align language models with non-differentiable reward signals, such as human preferences.However, a major challenge arises from the sparsity of these reward signals -typically, there is only a single reward for an entire output.This sparsity of rewards can lead to inefficient and unstable learning.To address this challenge, our paper introduces an novel framework that utilizes the critique capability of Large Language Models (LLMs) to produce intermediate-step rewards during RL training.Our approach pairs a policy model with a critic language model that provides feedback on each part of the policy's output.This feedback is then translated into token or span-level rewards that can be used to guide the RL training process.We investigate this approach under two different settings: one where the policy model is smaller and is paired with a more powerful critic model, and another where a single language model fulfills both roles.We assess our approach on three text generation tasks: sentiment control, language model detoxification, and summarization.Experimental results show that incorporating artificial intrinsic rewards significantly improve both sample efficiency and the overall performance of the policy model, supported by both automatic and human evaluation.The code is available under Google Research github * .

Store site recommendation is one of the essential business services in smart cities for brick-and-mortar enterprises. In recent years, the proliferation of multisource data in cities has fostered unprecedented opportunities to the data-driven store site recommendation, which aims at leveraging large-scale user-generated data to analyze and mine users’ preferences for identifying the optimal location for a new store. However, most works in store site recommendation pay more attention to a single data source which lacks some significant data (e.g., consumption data and user profile data). In this paper, we aim to study the store site recommendation in a fine-grained manner. Specifically, we predict the consumption level of different users at the store based on multisource data, which can not only help the store placement but also benefit analyzing customer behavior in the store at different time periods. To solve this problem, we design a novel model based on the deep neural network, named DeepStore, which learns low- and high-order feature interactions explicitly and implicitly from dense and sparse features simultaneously. In particular, DeepStore incorporates three modules: 1) the cross network; 2) the deep network; and 3) the linear component. In addition, to learn the latent feature representation from multisource data, we propose two embedding methods for different types of data: 1) the filed embedding and 2) attention-based spatial embedding. Extensive experiments are conducted on a real-world dataset including store data, user data, and point-of-interest data, the results demonstrate that DeepStore outperforms the state-of-the-art models.

Skeletonization offers a compact representation of an object while preserving important topological and geometrical features. Literature on skeletonization of binary objects is quite mature. However, challenges involved with skeletonization of fuzzy objects are mostly unanswered. This paper presents a new theory and algorithm of skeletonization for fuzzy objects, evaluates its performance, and demonstrates its applications. A formulation of fuzzy grassfire propagation is introduced; its relationships with fuzzy distance functions, level sets, and geodesics are discussed; and several new theoretical results are presented in the continuous space. A notion of collision-impact of fire-fronts at skeletal points is introduced, and its role in filtering noisy skeletal points is demonstrated. A fuzzy object skeletonization algorithm is developed using new notions of surface- and curve-skeletal voxels, digital collision-impact, filtering of noisy skeletal voxels, and continuity of skeletal surfaces. A skeletal noise pruning algorithm is presented using branch-level significance. Accuracy and robustness of the new algorithm are examined on computer-generated phantoms and micro- and conventional CT imaging of trabecular bone specimens. An application of fuzzy object skeletonization to compute structure-width at a low image resolution is demonstrated, and its ability to predict bone strength is examined. Finally, the performance of the new fuzzy object skeletonization algorithm is compared with two binary object skeletonization methods.