KG-WISE decomposes GNN models and uses LLM-generated query templates for partial loading of relevant components, achieving up to 28x faster inference and 98% lower memory on KGs with up to 42 million nodes while preserving accuracy.
Accelerating large-scale inference with anisotropic vector quantization
7 Pith papers cite this work. Polarity classification is still indexing.
7
Pith papers citing it
representative citing papers
RETRO matches GPT-3 and Jurassic-1 performance on the Pile benchmark using 25 times fewer parameters by conditioning on retrieved chunks from a 2-trillion-token database.
A per-component SimHash fingerprint supplies structural identity for AI agent skills, recovering family membership under paraphrase and refactoring with AUC 0.974 while localizing changes.
XTR training does not improve retrieval effectiveness over ColBERT but enhances IVF engine efficiency by flattening token scores to produce more discriminative centroids.
RACORN-1 adds adaptive search fallback to ACORN-1 to fix recall collapse at low selectivity in filtered vector search, achieving 9-26x speedups over HNSW with recovered recall on 1M-40M datasets.
A distilled LLM generates real-time, natural-language user interest personas—combining summarized interests with novel exploration topics—and this system produced small but significant viewer-value gains in a billion-user A/B test.
Ascend-RaBitQ is the first heterogeneous NPU-CPU optimized IVF-RaBitQ system for billion-scale vector search that decouples coarse ranking on NPU from fine ranking on CPU to leverage optimal hardware per stage.
citing papers explorer
-
An LLM-Guided Query-Aware Inference System for GNN Models on Large Knowledge Graphs
KG-WISE decomposes GNN models and uses LLM-generated query templates for partial loading of relevant components, achieving up to 28x faster inference and 98% lower memory on KGs with up to 42 million nodes while preserving accuracy.
-
Improving language models by retrieving from trillions of tokens
RETRO matches GPT-3 and Jurassic-1 performance on the Pile benchmark using 25 times fewer parameters by conditioning on retrieved chunks from a 2-trillion-token database.
-
The Decomposition Is the Fingerprint: Per-Component Identity for Agent Skills
A per-component SimHash fingerprint supplies structural identity for AI agent skills, recovering family membership under paraphrase and refactoring with AUC 0.974 while localizing changes.
-
A Replicability Study of XTR
XTR training does not improve retrieval effectiveness over ColBERT but enhances IVF engine efficiency by flattening token scores to produce more discriminative centroids.
-
RACORN-1: Adaptive Recall-Preserving Speedup for Low-Selectivity Filtered Vector Search
RACORN-1 adds adaptive search fallback to ACORN-1 to fix recall collapse at low selectivity in filtered vector search, achieving 9-26x speedups over HNSW with recovered recall on 1M-40M datasets.
-
LLM-Based User Personas for Recommendations at Scale
A distilled LLM generates real-time, natural-language user interest personas—combining summarized interests with novel exploration topics—and this system produced small but significant viewer-value gains in a billion-user A/B test.
-
Ascend-RaBitQ: Heterogeneous NPU-CPU Acceleration of Billion-Scale Similarity Search with 1-bit Quantization
Ascend-RaBitQ is the first heterogeneous NPU-CPU optimized IVF-RaBitQ system for billion-scale vector search that decouples coarse ranking on NPU from fine ranking on CPU to leverage optimal hardware per stage.