Relative Kinetic Utility: Calibrating Cross-Layer Credit for Global Structured LLM Pruning

Name: Relative Kinetic Utility: Calibrating Cross-Layer Credit for Global Structured LLM Pruning
Uploaded: 2026-10-03T06:05:30.784388+00:00
Duration: 358 s
Description: Relative Kinetic Utility (RKU) treats structural pruning as a continuous integral over a model's depth manifold using Alternating Gradient Flow (AGF). It overcomes the 'magnitude trap' of traditional methods by normalizing kinetic utility using Fisher Information Matrix trace. This approach preserves complex reasoning performance, especially at high sparsity levels, and improves generalization.

Relative Kinetic Utility (RKU) treats structural pruning as a continuous integral over a model's depth manifold using Alternating Gradient Flow (AGF). It overcomes the 'magnitude trap' of traditional methods by normalizing kinetic utility using Fisher Information Matrix trace. This approach preserves complex reasoning performance, especially at high sparsity levels, and improves generalization.

Open the full episode