Preprint

Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling

arXiv:2609.19499 — For self-consistency, candidate count N is not enough. At fixed N, how samples are batched can change LLM test-time scaling energy and latency by several times on A100 GPUs.

Mobina Kashaniyan, Ali Jannesari

Posted Nov 1, 2026 Conference Paper, Workshop Paper, Preprint, Large Language Models, High-Performance Computing, Energy Efficiency

ATEST: Adaptive Resource-Aware Test-Time Scaling for LLM Inference

ATEST (RATTS) is resource-aware self-consistency for LLM inference. It adapts how many samples to draw to live GPU load. The same four-sample query can cost 4.1× more energy when the GPU is busy. arXiv:2610.25519.

Mobina Kashaniyan, Ali Jannesari

Posted Oct 18, 2026 Preprint, Large Language Models, High-Performance Computing, Energy Efficiency, Test-Time Scaling

Interpretable Adaptive Sampling for LLM Test-Time Scaling

Interpretable adaptive sampling for self-consistency. A fuzzy controller uses prompt complexity and model confidence to choose how many self-consistency samples each prompt receives.

Mobina Kashaniyan, Ali Jannesari

Posted Aug 4, 2026 Preprint, Artificial Intelligence, Large Language Models