Jiachuan Wang

Jiachuan Wang, Shimin Di, Lei Chen et al.

Recently, emergence has received widespread attention from the research community along with the success of large-scale models. Different from the literature, we hypothesize a key factor that promotes the performance during the increase of scale: the reduction of monosemantic neurons that can only form one-to-one correlations with specific features. Monosemantic neurons tend to be sparser and have negative impacts on the performance in large models. Inspired by this insight, we propose an intuitive idea to identify monosemantic neurons and inhibit them. However, achieving this goal is a non-trivial task as there is no unified quantitative evaluation metric and simply banning monosemantic neurons does not promote polysemanticity in neural networks. Therefore, we first propose a new metric to measure the monosemanticity of neurons with the guarantee of efficiency for online computation, then introduce a theoretically supported method to suppress monosemantic neurons and proactively promote the ratios of polysemantic neurons in training neural networks. We validate our conjecture that monosemanticity brings about performance change at different model scales on a variety of neural networks and benchmark datasets in different areas, including language, image, and physics simulation tasks. Further experiments validate our analysis and theory regarding the inhibition of monosemanticity.

Qingfa Xiao, Jiachuan Wang, Haoyang Li et al.

Recent advances in large language models (LLMs) have showcased exceptional performance in long-context tasks, while facing significant inference efficiency challenges with limited GPU memory. Existing solutions first proposed the sliding-window approach to accumulate a set of historical \textbf{key-value} (KV) pairs for reuse, then further improvements selectively retain its subsets at each step. However, due to the sparse attention distribution across a long context, it is hard to identify and recall relevant KV pairs, as the attention is distracted by massive candidate pairs. Additionally, we found it promising to select representative tokens as probe-Query in each sliding window to effectively represent the entire context, which is an approach overlooked by existing methods. Thus, we propose \textbf{ActQKV}, a training-free, \textbf{Act}ivation-aware approach that dynamically determines probe-\textbf{Q}uery and leverages it to retrieve the relevant \textbf{KV} pairs for inference. Specifically, ActQKV monitors a token-level indicator, Activation Bias, within each context window, enabling the proper construction of probe-Query for retrieval at pre-filling stage. To accurately recall the relevant KV pairs and minimize the irrelevant ones, we design a dynamic KV cut-off mechanism guided by information density across layers at the decoding stage. Experiments on the Long-Bench and $\infty$ Benchmarks demonstrate its state-of-the-art performance with competitive inference quality and resource efficiency.

Jialiang Wang, Hanmo Liu, Shimin Di et al.

Database systems have recently advocated for embedding machine learning (ML) capabilities, offering declarative model queries over large, managed model repositories, thereby circumventing the huge computational overhead of traditional ML-based algorithms in automated neural network model selection. Pioneering database studies aim to organize existing benchmark repositories as model bases (MB), querying them for the model records with the highest performance estimation metrics for given tasks. However, this static model selection practice overlooks the fine-grained, evolving relational dependencies between diverse task queries and model architecture variations, resulting in suboptimal matches and failing to further refine the model effectively. To fill the model refinement gap in database research, we propose M-DESIGN, a curated model knowledge base (MKB) pipeline for mastering neural network refinement by adaptively weaving prior insights about model architecture modification. First, we propose a knowledge weaving engine that reframes model refinement as an adaptive query problem over task metadata. Given a user's task query, M-DESIGN quickly matches and iteratively refines candidate models by leveraging a graph-relational knowledge schema that explicitly encodes data properties, architecture variations, and pairwise performance deltas as joinable relations. This schema supports fine-grained relational analytics over architecture tweaks and drives a predictive query planner that can detect and adapt to out-of-distribution (OOD) tasks. We instantiate M-DESIGN for graph analytics tasks, where our model knowledge base enriches existing benchmarks with structured metadata covering 3 graph tasks and 22 graph datasets, contributing data records of 67,760 graph models. Empirical results demonstrate that M-DESIGN delivers the optimal model in 26 of 33 data-task pairs within limited budgets.

Yushi Sun, Jiachuan Wang, Peng Cheng et al.

Annotation through crowdsourcing draws incremental attention, which relies on an effective selection scheme given a pool of workers. Existing methods propose to select workers based on their performance on tasks with ground truth, while two important points are missed. 1) The historical performances of workers in other tasks. In real-world scenarios, workers need to solve a new task whose correlation with previous tasks is not well-known before the training, which is called cross-domain. 2) The dynamic worker performance as workers will learn from the ground truth. In this paper, we consider both factors in designing an allocation scheme named cross-domain-aware worker selection with training approach. Our approach proposes two estimation modules to both statistically analyze the cross-domain correlation and simulate the learning gain of workers dynamically. A framework with a theoretical analysis of the worker elimination process is given. To validate the effectiveness of our methods, we collect two novel real-world datasets and generate synthetic datasets. The experiment results show that our method outperforms the baselines on both real-world and synthetic datasets.

Maocheng Li, Jiachuan Wang, Libin Zheng et al.

In this paper, we study the privacy-preserving task assignment in spatial crowdsourcing, where the locations of both workers and tasks, prior to their release to the server, are perturbed with Geo-Indistinguishability (a differential privacy notion for location-based systems). Different from the previously studied online setting, where each task is assigned immediately upon arrival, we target the batch-based setting, where the server maximizes the number of successfully assigned tasks after a batch of tasks arrive. To achieve this goal, we propose the k-Switch solution, which first divides the workers into small groups based on the perturbed distance between workers/tasks, and then utilizes Homomorphic Encryption (HE) based secure computation to enhance the task assignment. Furthermore, we expedite HE-based computation by limiting the size of the small groups under k. Extensive experiments demonstrate that, in terms of the number of successfully assigned tasks, the k-Switch solution improves batch-based baselines by 5.9X and the existing online solution by 1.74X, with no privacy leak.