SCOPE: Cheapest Models for Compound AI Systems
Pick a cheaper LLM for every agent in a compound AI system without dropping below a quality bar.
// Blogs
Short, plain-language introductions to each research work — what problem it tackles, the key idea, and why it matters. Every post links back to the paper, code, slides, and poster on the publications list.
GEM-Bench: Measuring Ads Inside AI Answers
The first benchmark for the satisfaction–engagement trade-off when ads are injected into generated answers.
KET-RAG: Cutting Graph-RAG Indexing Cost
Keep Graph-RAG retrieval quality while cutting the LLM cost of building the knowledge graph.
Customer Personas from Purchase Histories
Turn raw e-commerce purchase logs into natural-language personas that downstream models can use.
Influence Maximization under Matroid Constraints
One framework for the whole family of influence maximization problems whose seed rules form a matroid.
When Diffusion Meets Invitation Mechanisms
Diffusion when content spreads by invitation rather than passive sharing, modelled on real game campaigns.
Capacity Constrained Influence Maximization
Influence maximization when each seed can only serve a bounded audience.
Fast Approximation of Spanning Centrality
Provable estimates of how structurally critical each edge is, on graphs with hundreds of millions of edges.
PPRviz: Graph Layout Powered by PageRank
Lay out a graph so on-screen distance reflects personalized PageRank proximity — interactively, at scale.
Friendship Closeness via Social Identity Theory
Measure tie strength from social identity theory rather than raw interaction counts.
ACMin: Clustering Massive Attributed Graphs
Cluster attributed graphs by attributed multi-hop conductance, scaling to tens of millions of nodes.
RCELF: Residual-Based Influence Maximization
A residual-based speedup of the classic CELF greedy algorithm that keeps its approximation guarantee.