arXiv subjects

Changcheng Yuan

Publications and source records attributed to Changcheng Yuan.

4 recordsLinked to original sources

arXiv2026

KGCache: Amortized Subgraph Retrieval for KG Reasoning with LLMs

Large language models can answer knowledge-intensive questions more reliably when they are grounded with knowledge graphs, but systems such as Think-on-Graph and Reasoning-on-Graph repeatedly query the same graph neighborhoods across different questions. In this work, we study this repeated retrieval in Knowledge Graph Question Answering~(KGQA) workloads and propose KGCache, an in-memory cache for one-hop knowledge graph neighborhoods. KGCache is designed to be compatible with both iterative traversal (ToG) and one shot planning (RoG) KGQA paradigms. KGCache is placed between the KGQA engine and the backend serving the KG, so repeated entity requests can be served from cache instead of issuing new KG queries. We evaluate KGCache on WebQSP and CWQ using LRU, LFU, and a trace-aware Oracle policy. Our analysis shows that both datasets contain substantial entity reuse among starting entities and entities reached during traversal. We also explore semantic caching for similar queries, which shows additional hit-rate gains on WebQSP and needs further accuracy testing on CWQ. Entity caching accelerates KG retrieval by up to $1.91\times$, while semantic-context caching achieves up to $1.06\times$ full-system speedup in the evaluated WebQSP configurations, with each hit being up to $3.73\times$ faster.

cs.AI↗

arXiv2026

Fast Sparse Matrix Permutation for Mesh-Based Direct Solvers

We present a fast sparse matrix permutation algorithm tailored to linear systems arising from triangle meshes. Our approach produces nested-dissection-style permutations while significantly reducing permutation runtime overhead. Rather than enforcing strict balance and separator optimality, the algorithm deliberately relaxes these design decisions to favor fast partitioning and efficient elimination-tree construction. Our method decomposes permutation into patch-level local orderings and a compact quotient-graph ordering of separators, preserving the essential structure required by sparse Cholesky factorization while avoiding its most expensive components. We integrate our algorithm into vendor-maintained sparse Cholesky solvers on both CPUs and GPUs. Across a range of graphics applications, including single factorizations and repeated factorizations, our method reduces permutation time and improves the sparse Cholesky solve performance by up to 6.27x. Our code is available at https://github.com/BehroozZare/fast-permute.

cs.GR↗

arXiv2026

A New Class of Geometric Analog Error Correction Codes for Crossbar Based In-Memory Computing

Analog error correction codes have been proposed for analog in-memory computing on resistive crossbars, which can accelerate vector-matrix multiplication for machine learning. Unlike traditional communication or storage channels, this setting involves a mixed noise model with small perturbations and outlier errors. A number of analog codes have been proposed for handling a single outlier, and several constructions have also been developed to address multiple outliers. However, the set of available code families remains limited, covering only a narrow range of code lengths and dimensions. In this paper, we study a recently proposed family of geometric codes capable of handling multiple outliers, and develop a geometric analysis that characterizes their m-height profiles.

cs.IT↗

arXiv2026

On the Height Profile of Analog Error-Correcting Codes

In recent work, it has been shown that maintaining reliability in analog vector--matrix multipliers can be modeled as the following coding problem. Vectors in $\mathbb{R}^k$ are encoded into codewords of a linear $[n,k,d]$ code $C$ over $\mathbb{R}$. For prescribed positive reals $δ< Δ$, additive errors of magnitude at most $δ$ are tolerable and need no handling, yet outlying errors of magnitude greater than $Δ$ are to be located or detected. The trade-off between the ratio $Δ/δ$ and the number of outlying errors that can be handled is determined by the height profile of $C$; as such, the height profile provides a finer description of the error handling capability of $C$, compared to the minimum distance $d$, which only determines the number of correctable errors. This work contains a further study of the notion of the height profile. Several characterizations of the height profile are presented, thereby yielding methods for computing it. The starting point is formulating this computation as an optimization problem that is solved by a set of linear programs. This, in turn, leads to a combinatorial characterization of the height profile as a maximum (or max--min) over a certain finite set of codewords of $C$. Moreover, this characterization is shown to have a simple geometric interpretation when the columns of the generator matrix of $C$ all have the same $L_2$ norm. Through examples of several code families, it is demonstrated how the results herein can be used to compute the height profile explicitly.

cs.IT↗