PrunePath introduces a budget-adaptive structured sparsification framework that efficiently prunes Feed-forward networks in large language models, achieving hardware-friendly sparsity and measurable s…

The paper introduces SecRL-Prune, a structured reinforcement learning framework that effectively prunes CodeLLMs while preserving their critical ability to generate adversarial, functionality-preservi…

MAESTRO is a structured pruning framework designed for MoE language models that models autoregressive expert activation trajectories as Ergodic Markov chains, yielding a globally aware importance heur…

This paper introduces Polynomial-Sensitivity-Aware Pruning (PSAP), a reliability-aware structured pruning method for neural networks under homomorphic encryption, reducing catastrophic layer vulnerabi…

This paper introduces structured cut sets, a novel preprocessing technique for Maximum k-Cut, and extends existing techniques from Maximum Cut. The rules are optimality-preserving and yield significan…

The paper proposes PALS, a method for adjusting per-layer sparsity based on activation magnitudes in transformer models, achieving better performance than uniform one-shot pruning methods.

This paper investigates the use of spectral filtering for continuous subgraph matching over dynamic graphs and presents three key findings.

The paper proposes Locality-Aware Redundancy Pruning (LoRP), a training-free method that prunes LLM layers by exploiting localized inter-layer redundancy, leading to improved efficiency while maintain…

The paper introduces PRAXIS, a novel algorithm that efficiently approximates the computation of 'Rashomon sets' for decision trees, significantly reducing memory and runtime complexity.

The paper proposes an aggressive, parameter-efficient method to prune non-essential experts from Mixture-of-Experts (MoE) LLMs, significantly compressing the model while maintaining high machine trans…

DominoTree, a new best-first draft tree method for LLM inference, achieves up to 6.6x speedup over autoregressive decoding and the highest mean accept length, using a GPU-native CUDA-graph builder.

The paper presents an adaptive scaling algorithm with a competitive ratio of 1.373 for incremental submodular maximization under increasing cardinality constraint, improving upon the previous best res…

This paper presents an optimized algorithm for computing cut sets of a path set in graph theory and introduces a vectorized computational framework for property calculations.

A new solver-free parallel spectral sparsification algorithm for weighted graphs is presented, relying on low-diameter decompositions and independent sampling, eliminating dependence on target approxi…

The paper introduces CaDDTree, a cost-aware method that optimizes token throughput by jointly selecting the tree structure and node budget for speculative decoding, outperforming existing methods like…

The paper proposes SubFit, a novel compression technique that achieves superior LLM compression by replacing non-contiguous, submodule-level components (Attention and FeedForward) with lightweight res…