The paper investigates the energy costs of multilingual large language model (LLM) inference, revealing significant disparities across languages. Using the ML.Energy framework, the authors find that energy consumption per output token can differ by up to 8.3×, and total energy for a fixed request set can vary up to 179×, with English being the cheapest and Pashto the most expensive. The study attributes these differences to higher per-token costs for complex or rare scripts and longer outputs for low‑resource languages, and notes that high‑energy languages also tend to have lower task accuracy.
arXiv:2609.39001v1 Announce Type: cross Abstract: LLM services are billed per token and context windows are measured in tokens, yet the number of tokens needed for the same content varies across lang...
The paper introduces AssameseLM-E2B, a 2‑billion‑parameter language model tailored for the low‑resource Assamese language, and its instruction‑tuned variant AssameseLM-E2B‑it. Built on the Gemma‑4 architecture, it uses an expanded byte‑level BPE vocabulary that cuts tokenization overhead by nearly 500 % compared with multilingual baselines, and employs a parameter‑efficient Continual Pre‑Training pipeline with Hybrid LoRA followed by supervised fine‑tuning on a balanced, multi‑domain dataset. Evaluation shows strong performance with a Bits‑per‑Byte score of 0.7488 on FLORES‑200, a 96.7 % RAG score, and 100 % refusal accuracy on adversarial prompts, highlighting the effectiveness of parameter‑efficient adaptation for locally deployable models under constrained resources.
arXiv:2608. 15964v1 Announce Type: cross Abstract: Language-specific competency (LSC) is the phenomenon of a language model performing better or worse depending on the language of the prompt.
Large Language Models (LLMs) are increasingly used to optimize code for performance, but this study shows that such optimizations often increase energy consumption rather than reduce it. Across 11 open‑source models and 1,000 benchmark problems, inference energy ranged from 0.79 to 8.68 MJ, and 42 of 44 configurations produced code that used 4% to 886% more energy than human‑written reference solutions. Mixture‑of‑Experts architectures and token‑limit constraints can reduce inference energy, yet performance‑optimized code remains energy‑heavier, indicating that speed and sustainability are not always aligned goals.
arXiv:2410. 07809v2 Announce Type: replace-cross Abstract: Multilingual instruction tuning (MIT) is challenged by the curse of multilinguality, data scarcity, and high computational cost.
arXiv:2610.20247v1 Announce Type: cross Abstract: Multilingual models share capacity across languages, so adding languages can hurt performance on existing ones. Mixture-of-experts models expand this...
arXiv:2605. 30580v2 Announce Type: replace-cross Abstract: Speculative decoding is a popular technique for large language model (LLM) inference, enabling faster generation by drafting multiple tokens with a smaller draft model.
The paper investigates how prompt design influences energy consumption in on-device large language models (LLMs). It examines two prompt properties—cognitive load and phrasing pattern—across various datasets, models, and devices, using phase-level profiling to separate prefill and decode energy. Findings show that cognitive load mainly affects energy per token, while phrasing pattern influences energy mainly through token usage, and that prompt design reshapes the energy-quality trade‑off differently for each model.
arXiv:2609.23490v1 Announce Type: new Abstract: Large language model (LLM) agents increasingly execute multi-step workflows through tool use and interaction with users and environments. However, curr...
arXiv:2604.13286v2 Announce Type: replace Abstract: Despite the widespread multilingual deployment of large language models, post-training pipelines remain predominantly English-centric, contributing...
The paper investigates how token length affects compute time and energy usage in Large Language Models, focusing on both prefill (input) and decode (output) phases. It finds that decode energy scales linearly with output length, while prefill energy shows both linear and quadratic scaling with input length, highlighting the importance of large input contexts in agentic workflows. The study also demonstrates that compute time can reliably proxy energy consumption when direct measurements are unavailable, providing a framework for more sustainable AI deployment.
arXiv:2601. 06649v2 Announce Type: replace-cross Abstract: Research in machine learning has questioned whether increases in training token counts reliably produce proportional performance gains in large language models.