New EMNLP paper delivers 'downscaling laws': 10% lower loss costs 4-8x carbon, pruning 43% keeps 70% performance
Tanmoy_Chak · x · 2026-10-11
Following their ICML 2025 position paper arguing the field needs downscaling laws, the team delivers one at EMNLP 2026 after pruning ten LLMs. Key findings: a 10% lower loss costs 4-8x the carbon emissions, while pruning 43% of a model keeps 70% of its performance with zero retraining. Their thesis: scaling laws tell you what a bigger model buys, but scaling down is the predictable, cheap path.
More from Research
- Parkinson's patients embraced DBS without knowing how it works — a lesson for AI-designed drugs — ShahabBakht · 2026-10-12
- DeepMind weather model nailed Hurricane Isaias track 11 days out — joao_gante · 2026-10-12
- Models trained on their own output lose rare knowledge first, drift into nonsense — TejasKumar_ · 2026-10-12
- Stanford Paper Uses CNN on Raw Prices to Predict S&P 500, Challenging Factor Tradition — mdancho84 · 2026-10-12
- Paradromics BCI lets clinical participant communicate via imagined speech — PeterDiamandis · 2026-10-12
- Did OpenAI limit solutions to protect academia? Debating decentralized, GitHub-speed research — IgorCarron · 2026-10-12