The paper investigates run‑to‑run variance in semantic query engines that use large language models (LLMs) to evaluate rows. By re‑executing benchmarks (SemBench, LRO‑Bench, UDA‑Bench) and logging 1.2 million model calls, the authors find that between 0.07% and 10.2% of rows change output at temperature 0, with some tasks showing up to 24% change. They show that plan amplification can cause a large fraction of answers to differ, that many benchmark comparisons are unresolved, and that voting and calibrated cascades reduce but do not eliminate variance, leading to new reporting and design guidelines.
By Oktie Hassanzadeh, Dharmashankar Subramanian
arXiv:2602. 10387v2 Announce Type: replace-cross Abstract: Traditional query optimization relies on cost-based optimizers that estimate execution cost (e.
By Mehmet Hamza Erol, Xiangpeng Hao, Federico Bianchi, Ciro Greco, Jacopo Tagliabue, James Zou
The paper investigates whether training Mixture-of-Experts (MoE) routers can improve memory‑bandwidth locality on consumer GPUs. Using a new zero‑surgery telemetry tool, the authors measure that a large Qwen3‑235B model is bottlenecked by disk‑based expert access, and that an LRU cache can serve a majority of requests. They pre‑register experiments training 137 M‑parameter MoE models with locality‑aware losses, finding that while cache misses can drop up to 60 % (99 % static‑pin hit rate), every configuration fails to meet a strict 1 % perplexity threshold, indicating a tight coupling between cache efficiency and model quality.
By Shriniwas Ramesh Suram
The paper presents a Pareto atlas of LLM inference optimizations, mapping cost, quality, and latency trade‑offs for Qwen2.5‑7B‑Instruct on L4, A100, and H100 GPUs. Using 54 measured configurations and a calibrated simulator, it identifies 18 of 36 setups on the Pareto frontier, showing that combined methods outperform single ones. Quality tests reveal that AWQ 4bit and FP8 weights offer significant latency reductions while largely preserving accuracy, but naive FP8 KV caching fails to answer any questions correctly.
By Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly
LESS (Lightweight Evolutionary Supernet Search) is a data‑driven NAS method that uses a brief hard‑path warm‑up and CMA‑ES to evaluate candidate architectures as decoded hard genotypes after six supernet updates. On NAS‑Bench‑201, LESS attains 93.189 % CIFAR‑10 accuracy in just 409.1 seconds, nearly matching FairNAS while using only about 1/24 of its search time. The approach also transfers well to CIFAR‑100, ImageNet16‑120, and the larger DARTS space, achieving high accuracies with searches completed in roughly 43.5 minutes on a single GPU.
By Aviral Gandhi, Jinglue Xu, Jialong Li, Hitoshi Iba
arXiv:2608. 07911v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard.
By Yu Zhang
PageWeaver introduces a KV‑guided query union strategy for sparse attention, assembling query groups based on selected‑page affinity while preserving each query’s support and output ownership. The design employs a bounded GPU search to generate query IDs and an ID‑aware two‑CTA kernel that consumes them without materializing reordered Q tensors or cross‑page partial outputs. Experiments on the H200 Union8 with FP8 KV show a 1.70× speedup over FlashInfer, with additional latency reductions of 3.26–7.66% on selected captures and modest throughput gains for whole‑model prefill.
By Zhiyuan Li, Zihan Li, Zefang Yuan, Lei Wang, Hao Wang
arXiv:2609.01244v1 Announce Type: new
Abstract: Every supervised fine-tuning run forces the same chain of decisions, such as learning rate, batch size, LoRA or full fine-tuning, how many epochs, whic...
By Charles O'Neill, Mudith Jayasekara, Harry Partridge
Serving a 235B-parameter Mixture-of-Experts (MoE) model on a single 8 GB GPU is bottlenecked not by compute but by memory bandwidth: decode must stream each token's active experts from whichever tier holds them, and on consumer hardware most experts sit on an SSD far slower than RAM. We quantify this bandwidth wall on Qwen3-235B (Q4_K_M, 134 GB): measured decode is 0.
Every supervised fine-tuning run forces the same chain of decisions, such as learning rate, batch size, LoRA or full fine-tuning, how many epochs, which optimiser, and what data to feed the model. Eac...
arXiv:2608.07911v4 Announce Type: replace
Abstract: Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard. This makes expert c...
By Yu Zhang
arXiv:2609.05637v2 Announce Type: replace
Abstract: A popular way to improve Retrieval-Augmented Generation (RAG) is to rewrite the user's question into several variants and search with all of them....
By Sara Shanian, Xiaoqin Yi, Pavlo Ruban, Kurt MacDonald