Multiple institutions · 2026-04-28 · notable
Unstructured Pruning Boosts LLM Test-Time Scaling — Pruned Models Can Outperform Their Unpruned Versions
Unstructured weight pruning augments test-time compute scaling in reasoning LLMs, and at times outperforms the unpruned full-weight model. Experiments on s1.1-7B and Qwen3-8B show layer-wise sparsity allocation is critical to preserving reasoning ability.
Targeted weight removal can actually improve reasoning under test-time scaling — flipping the assumption that pruning hurts capability.
What is it?
This paper submitted to arXiv on April 28, 2026, revisits the effectiveness of unstructured pruning for reasoning LLMs under test-time compute scaling. Unlike prior work showing structured pruning degrades reasoning, the authors find selective individual weight removal can enhance test-time scaling performance.
How does it work?
The study contrasts structured pruning (removing whole attention heads or MLP blocks) with unstructured pruning (removing individual weights with high redundancy or negative contribution). Experiments on reasoning-tuned models (s1.1-7B, Qwen3-8B) vary sparsity ratios and layer-wise allocation strategies, measuring performance on reasoning benchmarks under different inference compute budgets.
Why does it matter?
Smaller models are cheaper to serve and scale to longer reasoning chains per token budget. If pruning can produce models that reason better under extended inference budgets, this opens a path to more capable edge and on-device reasoning agents without additional training.