AIHugging Face
Heuristic editor, no API keyVerdict: RoutineSparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference
The memory-bound nature of the decoding stage of large language model (LLM) inference incurs significant latency.
Key numbers
- 1.48x end-to-end wall-clock decoding speedup
Caveats
- Preprint; not yet peer reviewed.
VerdictCompetent work. Briefs at most.
Abstract
The memory-bound nature of the decoding stage of large language model (LLM) inference incurs significant latency. Layer-wise training-free network pruning approaches guided by the Hessian have been a prominent solution to this problem, as pruning reduces the number of nonzero parameters read from memory during decoding. Nevertheless, typical methods in this line compute the Hessian using pre-collected natural sequences, whereas the model is fed self-generated tokens during decoding, creating a distribution shift between the two sequences. The Hessian calculated on the natural sequence is different from that calculated on the generated sequence. We observe that this discrepancy causes the activation distribution during generation to deviate from that used for pruning, further hurting the pruned model performance. Moreover, most existing LLM pruning methods that bring actual speedup primarily target the sparse matrix-matrix (SpMM) multiplication, providing limited support for the sparse matrix-vector (SpMV) operations, which dominate decoding. To solve these problems, we introduce SparseDecoding, a principled decoding-aware pruning framework tailored for accurate and efficient LLM decoding. Specifically, at the algorithmic axis, SparseDecoding constructs calibration matrices from layer-wise activations collected during the dense-model autoregressive generation, excluding prefill, thereby aligning the pruning objective with the decoding activations. At the system axis, we develop an optimized N:M sparse matrix-vector kernel with bitmask indexing and fixed-step traversal. Substantial empirical results on representative LLMs (Llama-3.1-8B, Llama-3.3-70B, Qwen3-14B / 32B) demonstrate that our method consistently outperforms standard fixed-text calibration on the long-form generation benchmarks while achieving up to 1.48x end-to-end wall-clock decoding speedup on A100 GPUs.
The editor's rubric
| Dimension | Level | Weight | What that level means |
|---|---|---|---|
| Leverage | ███░░ 3 | 24% | A method or resource many groups across the field will adopt within a year. |
| Magnitude | ███░░ 3 | 18% | Large gain: roughly 2x, or a clear new state of the art on a hard, unsaturated problem. |
| Evidence | ██░░░ 2 | 14% | Limited: single setting, weak baselines, or an observational association presented as causal. |
| Novelty | ███░░ 3 | 16% | A genuinely new approach to an open problem. |
| Trajectory | ██░░░ 2 | 18% | Some room to improve with obvious engineering. |
| Stakes | ██░░░ 2 | 10% | Benefits a professional community (practitioners, clinicians, engineers). |
Editor’s rationale
Heuristic triage from title and abstract text only, not a reading of the paper. Cues found: method (we propose); gains (x-fold, outperforms); novelty (discovery); scale (efficient); stakes (general AI).
How the score was computed
- Merit
- 4.2 / 10
- Adjusted merit
- 4.1 / 10
- Attention
- 61%
- Freshness
- 32%
- Citations0 (reference 15, via semantic-scholar, Oct 11, 2026, 13:49 UTC)
- Influential citations0 (reference 3, via semantic-scholar, Oct 11, 2026, 13:49 UTC)
- Hugging Face upvotes28 (reference 25, via hf-daily, Oct 11, 2026, 13:49 UTC)