According to arxiv.org, researchers have conducted the first systematic study of how large language model (LLM) training performance responds to GPU power reductions. The paper, published September 12, 2026, introduces the Power Flexibility Index (PFI), a metric that quantifies performance costs of power reductions and provides “a control primitive for SLA-aware power flexibility.”
The study collected data from 131 LLM training runs on H200 GPUs (plus 24 validation runs and 34 matched H100 runs), examining both dense and mixture-of-experts models across pretraining and fine-tuning tasks with up to 32 GPUs. According to the research, LLM training jobs exhibit “substantial but variable power elasticity,” and the team identified telemetry signals that can predict PFI at runtime.
Under a 30% power reduction, PFI-aware power allocation recovered “~1.5k tokens/s per job, 63% of the performance gap between an equal-weight allocation and an oracle with perfect information,” according to the findings.
In separate research also published September 12, 2026, arxiv.org reports on a study of knowledge attribution in LLMs. The paper introduces AttriWiki, a self-supervised pipeline for identifying whether model outputs originate from internal knowledge or provided context. According to the study, simple linear probes achieved up to 0.96 Macro-F₁ scores on models including Llama-3.1-8B, Mistral-7B, and Qwen-7B, with the research noting that “attribution mismatches raise error rates by up to 70%.”