Research map / Large language models / Instruction and fine-tuning
Towards Resource-Efficient LLMs: End-to-End Energy Accounting of Distillation Pipelines
Katherine Lambert, Sasha Luccioni
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace baseline SFT with logit KD from 32B teacher | OLMo-2 7B student · TULU-3 | Total energy (kWh): 19.50 -> 28.40 | Hurt |
| replace baseline SFT with logit KD | OLMo-2 7B student · TULU-3 | Quality Q: 0.90 -> 0.78 | Hurt |
| replace baseline SFT with synthetic SFT | OLMo-2 13B student · TULU-3 | Quality Q: 0.99 -> 0.85 | Hurt |
| replace baseline SFT with KD or synthetic SFT | OLMo-2 1B student · TULU-3 | Total energy (kWh): 7.00 -> 16.90 | Hurt |
| replace baseline SFT with KD or synthetic SFT | OLMo-2 1B student · TULU-3 | Quality Q: 0.69 -> 0.70 | Improved |
| scale student from 1B to 13B (ablation) | Baseline SFT · TULU-3 | Total energy (kWh): 7.00 -> 34.60 | Mixed |
| scale student from 1B to 13B (ablation) | Baseline SFT · TULU-3 | Quality Q: 0.69 -> 0.99 | Improved |
| reuse cached teacher artifacts across N students (ablation) | KD / synthetic SFT · Amortization | Energy per student: break-even N approx 10 (KD 1B) | Improved |
| use distillation (KD) for student training stage | OLMo-2 13B student · TULU-3 | Student training energy (kWh): 33.15 -> 30.05 | Improved |
| increase KD temperature T (ablation) | KD 7B student · TULU-3 sweep | Quality | No clear effect |
| increase KD loss weight alpha (ablation) | KD students · TULU-3 sweep | Quality | Improved |
| increase max_new_tokens for synthetic generation (ablation) | Synthetic SFT 7B · TULU-3 sweep | Quality | Improved |
Closest papers by meaning
- Demystifying Cost-Efficiency in LLM Serving over Heterogeneous GPUs (ICML 2025 · similarity 0.86)
- ConServe: Fine-Grained GPU Harvesting for LLM Online and Offline Co-Serving (ICML 2026 · similarity 0.84)
- DéjàVu: KV-cache Streaming for Fast, Fault-tolerant Generative LLM Serving (ICML 2024 · similarity 0.83)
- Distillation Scaling Laws (ICML 2025 · similarity 0.83)
- PipeOffload: Improving Scalability of Pipeline Parallelism with Memory Optimization (ICML 2025 · similarity 0.83)
- The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
