Output-layer gradients cut LLM data selection costs by 10x
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Researchers have identified a computational shortcut for post-training data selection in large language models. Rather than computing full-parameter gradients to rank training samples by relevance, the LESSER method uses only output-layer gradients, which require cheaper forward passes instead of backward passes. This reduces feature-extraction costs by 9.7x for supervised fine-tuning and 3x for other tasks. The finding matters because data selection directly influences model quality, and this efficiency gain makes gradient-based ranking practical at scale, lowering the barrier for teams optimizing training pipelines without massive compute budgets.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper doesn't just claim efficiency gains; it shows that output-layer gradients preserve ranking quality despite being orders of magnitude cheaper to compute. The critical detail is that this works because relevance signals concentrate in the final layer, not because the method is approximate or lossy in ways that degrade downstream model quality.
This connects directly to the SFT data-curation work from October 1st, which demonstrated that supervised finetuning quality depends heavily on training data distribution. LESSER solves a practical bottleneck that work identified: if data selection is the lever for capability gains, teams need a way to rank samples without prohibitive compute costs. The efficiency gain also echoes the zero-order training paper from late September, which tackled memory constraints in fine-tuning by rethinking what computation is actually necessary. Both papers share the insight that standard approaches waste resources on steps that don't drive the outcome.
If LESSER's 9.7x speedup on SFT holds when applied to the same datasets used in the October 1st SFT paper, that confirms the method generalizes beyond toy benchmarks. Watch whether practitioners report adoption in the next 6 months; if the method stays confined to arXiv while teams continue using full-gradient ranking, it suggests the practical barriers to adoption (integration, validation overhead) exceed the compute savings.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
Supervised finetuning outperforms expectations with strategic data sampling
Researchers challenge the conventional hierarchy between supervised finetuning and reinforcement learning in posttraining, demonstrating that SFT can achieve stronger generalization than previously assumed when data distribution is optimized appropriately. Rather than redesigning loss functions, the work focuses on curating training data to capture on-policy learning benefits while retaining access to offline expert trajectories. This finding…
MentionsLESSER
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “LESSER: Post-Training Data Selection with Output-Layer Gradients”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.