LLM-Pruner: On the Structural Pruning of Large Language Models

Compressing an LLM usually means task-specific pruning, which destroys general capability, and recovery normally needs the original (enormous) training corpus.

TASK-AGNOSTIC STRUCTURAL pruning under two constraints (stay general, need little of the original data): use gradient information from a small calibration set to identify non-critical COUPLED structures - groups of parameters that must be removed together - rather than independent weights, so the remaining model stays a valid architecture. Recovery is done with LoRA, not full fine-tuning, which is what makes the 50K-example budget sufficient.

Paper's own claim (abstract): performance of the pruned models recovered via LoRA 'in merely 3 HOURS, requiring only 50K data', validating on LLaMA, Vicuna and ChatGLM with the compressed models retaining 'satisfactory' zero-shot classification and generation. COMPRESSION RATIO: structured pruning (whole groups/structures removed), so the ratio depends on the pruning budget chosen - no single ratio is quoted in the abstract. QUALITY COST: the meaningful cost is not perplexity but the LoRA recovery budget (3 h / 50K examples), which is the number to compare against.

arXiv ID 2305.11627 verified by Semantic Scholar /paper/search/match, which returned the title as 'LLM-Pruner: On the Structural Pruning of Large Language Models' - note this differs slightly from the common 'A Structural Pruning Approach' form, and the Semantic Scholar record is what is cited here. Venue NeurIPS 2023 from Semantic Scholar; year = venue year 2023. v1 date 2023-05-18 NOT independently confirmed against the /abs/ page during this pass (arXiv was rate-limiting) - flagged rather than guessed. arXiv ID and title verified by fetching https://arxiv.org/abs/2305.11627 and reading the citation_title meta tag on 2026-10-03. Venue and year from the Semantic Scholar graph API batch endpoint (paper/arXiv:2305.11627) unless the record's notes say otherwise. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: neurips has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.