This work characterize the effect of a residual-stream perturbation through both its activation covariance and the local sensitivity of the model output, and derives a tractable spectral upper bound that reduces subspace selection to an eigendecomposition of a sensitivity-weighted covariance matrix, retaining the efficiency and structural simplicity of rotation-based pruning methods.
Abstract
Residual stream pruning methods reduce inference cost by shrinking the model's hidden dimension, but existing approaches typically choose these dimensions by minimizing activation reconstruction error. This criterion implicitly treats all perturbation directions as equally important, ignoring the sensitivity of downstream layers. We introduce a sensitivity-aware approach to residual-stream pruning that directly accounts for this direction-dependent sensitivity. Using a second-order approximation to the output KL divergence, we characterize the effect of a residual-stream perturbation through both its activation covariance and the local sensitivity of the model output. The resulting subspace selection objective couples these two quantities, but is difficult to optimize directly. We derive a tractable spectral upper bound that reduces subspace selection to an eigendecomposition of a sensitivity-weighted covariance matrix, retaining the efficiency and structural simplicity of rotation-based pruning methods. Across several instruction-tuned language model families, our method consistently reduces calibration KL divergence relative to activation-only pruning and improves perplexity and downstream task performance over a range of compression levels. Our results show that preserving activation energy alone is insufficient for residual-stream pruning, and that explicitly accounting for how perturbations propagate to the model output provides a more effective criterion for selecting dimensions to remove.
This work forms the pruning objective as a cardinality-constrained binary quadratic program that explicitly models cross-unit dependencies in the reconstruction error, and develops a greedy interaction algorithm based on dependency-aware marginal costs to optimize unit selection.
Si-Cheng Xu, Hao Shi, Wei Zhang et al.· 0 citations
COEC (Calibrated Orthogonal-Equivalence Compensation), a training-free compensation framework that applies alternating left and right orthogonal rotations to the retained weight, improves perplexity on every model and zero-shot accuracy in most settings over existing compensation methods.
Peiqi Yu, Nam Ling, Wei Wang et al.· 1 citation· ⚡1
Debias-SparseGPT is introduced, a post-training pruning method incorporating representational debiasing using a second-order term defined over demographically contrasting inputs that consistently reduces pruning-induced bias compared to SparseGPT while preserving model perplexity and zero-shot accuracy.
Irina Proskurina, Guillaume Metzler, Antoine Gourru et al.· 0 citations
This work describes the optimal fixed-dimensional linear representation and proposes eigenbasis-corrected one-bit gradient projection (EOGP) to approximate it at scale and predicts retraining outcomes more accurately than the evaluated compression baselines while using one-sixteenth of their per-example storage.
Jaeseung Heo, J. Rosser, Dongwoo Kim· arXiv.org· 0 citations
Output-layer gradients suffice for effective data selection, yet require only the cheaper forward pass, and a drop-in wrapper for selection methods that reduces the feature-extraction FLOP cost by $9.7\times$ for SFT and $3.0\times$ for RL benchmarks, while tracking full-gradient performance on downstream tasks.
Lyuxin David Zhang, Eric Wong, Surbhi Goel et al.· 1 citation
Extending LILA to dynamically allocate sparsity budgets via KS-scores yields state-of-the-art generative preservation at moderate compression, while uncovering fundamental single-layer architectural bottlenecks at higher compression regimes.
Sankar Behera, Dhruv Singh, Anshika Agnihotri et al.· 1 citation
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduNov 17, 2026
By studying the computational foundations of language learning, Associate Professor Jacob Andreas builds intelligent systems that can communicate effectively with humans.
Microsoft Research Blog· microsoft.comNov 16, 2026
AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduNov 13, 2026