aiqbench analysis
SARATHI introduces a novel framework for structured LLM pruning that decouples neuron selection from weight reconstruction, effectively eliminating domain bias and memory bottlenecks inherent in coupled methods. By leveraging data-free probing techniques and adaptive observation strategies, the approach maintains competitive perplexity and zero-shot accuracy across diverse model architectures at high sparsity levels. This advancement represents a significant step toward more efficient, hardware-agnostic inference without relying on extensive calibration datasets.
SARATHI: Decoupled Data-Free Probing and Adaptive Reconstruction for Structured LLM Pruning
arXiv:2610.15514v1 Announce Type: cross Abstract: Structured pruning reduces the inference cost of large language models (LLMs) by permanently removing neurons and weight rows, yielding dense, hardware-agnostic speedup. Existing post-training methods couple neuron selection and weight reconstruction through a shared calibration dataset, introducing domain bias and out-of-memory failures at high sparsity. We propose SARATHI, a framework that decouples these two stages. A data-free \emph{NMF Residual Subspace Probe} scores neurons using intrinsic weight geometry, without accessing any calibration data, and carries a formal connection to Robust PCA. \emph{Adaptive OBS} then reconstructs the pruned weights with a localized calibration batch, operating in $\mathcal{O}(d^2)$ peak memory per layer and eliminating the global memory bottleneck of coupled methods. A \emph{Bias Shift Compensation} step corrects systematic activation drift in post-LayerNorm architectures that would otherwise cause perplexity collapse at high sparsity. Experiments across six models from two architecture families (gated and non-gated FFN) at 15--40\% sparsity show that SARATHI achieves competitive perplexity and matches or exceeds existing structured pruning baselines on zero-shot accuracy.
Abstract:Structured pruning reduces the inference cost of large language models (LLMs) by permanently removing neurons and weight rows, yielding dense, hardware-agnostic speedup. Existing post-training methods couple neuron selection and weight reconstruction through a shared calibration dataset, introducing domain bias and out-of-memory failures at high sparsity. We propose SARATHI, a framework that decouples these two stages. A data-free \emph{NMF Residual Subspace Probe} scores neurons using intrinsic weight geometry, without accessing any calibration data, and carries a formal connection to Robust PCA. \emph{Adaptive OBS} then reconstructs the pruned weights with a localized calibration batch, operating in $\mathcal{O}(d^2)$ peak memory per layer and eliminating the global memory bottleneck of coupled methods. A \emph{Bias Shift Compensation} step corrects systematic activation drift in post-LayerNorm architectures that would otherwise cause perplexity collapse at high sparsity. Experiments across six models from two architecture families (gated and non-gated FFN) at 15--40\% sparsity show that SARATHI achieves competitive perplexity and matches or exceeds existing structured pruning baselines on zero-shot accuracy.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2610.15514 [cs.LG]
(or arXiv:2610.15514v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.15514 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 262 words
Related coverage
[AINews] not much happened today
Latent Space · Sep 18, 2026
News Archive 2026
Quantum Computing Report · Jul 25, 2026
Decoupled Structure-Feature Alignment via Alternating Optimization for Graph Learning
arXiv · Sep 1, 2026
News
Quantum Computing Report · Jul 25, 2026