TableTune is a system that automatically measures and corrects draft-length tables for speculative decoding in large language model serving. It profiles decoding performance across batch sizes and draft lengths, generates an optimized table in minutes, and a controller plugin continuously updates the table based on live traffic. Experiments on H100 GPUs show that TableTune consistently achieves at least 94% of the best fixed draft-length throughput, outperforms other speculative control methods, and can recover up to 96% of performance lost by stale tables, delivering significant speedups in real deployments.
By Abhishek Malvankar, Paul Castro, Nick Mitchell, Lionel Villard
ScaleRLT introduces a hybrid learner log‑probability method that reconciles differences between dense and deferred rollout computations in looped language models like Ouro. By using a stop‑gradient replay of the rollout schedule, the approach achieves bit‑for‑bit reproducibility across FP32 and bf16 real‑weight cells while maintaining the same gradient accuracy as dense implementations. The technique boosts rollout throughput by up to 1.56× on a single GPU and significantly speeds up multi‑node GRPO updates, leading to substantial accuracy gains on GSM8K in short training runs.
By Dezhen Lu, Zixiang Ni, Runguo Li
The paper introduces Temperon, a training strategy that uses plain SGD for the first 43% of the epoch budget and then hands off to a SAM‑wrapped Muon refiner for the remaining training. On datasets such as CIFAR‑10/100, SVHN, and Tiny ImageNet, Temperon achieves the same or better accuracy as full‑time SAM while reaching key performance targets faster and at lower cost. Ablation studies show that the Muon refiner contributes the majority of the performance gain, while the initial SGD explorer and its restarts add negligible benefit.
By Stamatis Mastromichalakis
The paper presents an end‑to‑end MXFP8 reinforcement learning implementation on Blackwell GPUs, evaluating its benefits and limitations with Qwen3-8B and Qwen3-30B-A3B models. It finds that matching FP8 precision between sampler and learner improves consistency, that speed gains appear only for large micro‑batches, and that several weight‑synchronization correctness conditions can fail silently. The study also introduces seven explicit checks to detect such failures during training.
By Wengee Zhang
arXiv:2607. 20792v1 Announce Type: new Abstract: Memoir combines per-sample fast memory, shared slow parameters, variable-depth latent recurrence, and a future-latent energy objective.
By Jaber Jaber, Osama Jaber
SCLATE is a new execution substrate that allows continual‑learning benchmarks and agents to share a single event scheduler via adapters, enabling tasks, session events, and memory consolidation to run on a compressed, real‑time timeline. It also functions as a rollout engine that records every model call’s tokens and log probabilities without modifying the agent’s harness or memory. Using SCLATE, the authors ported seven benchmarks, compared ten harness‑memory configurations across ten models, and demonstrated that post‑training Qwen3.5‑4B can effectively leverage both harness and memory, improving performance on multiple metrics.
By Youngmok Jung, Sirajul Salekin, Henry Tran, Javier Movellan, Zhao Huang, Manjot Bilkhu
arXiv:2606. 19989v1 Announce Type: cross Abstract: Modern LLM training breaks a core assumption behind offline batch samplers: the true training cost of a sample is only observable after preprocessing, augmentation, templating, tokenization, and multimodal visual-token expansion.
By Dian Li, Zekun Wang, Yaoru Wang, Jiahong Yan
The paper reports fine‑grained, deterministic instability in Dynamic Tensor Rematerialization (DTR), an online eviction policy for memory‑constrained DNN training. On an LSTM trace, tiny changes in memory budget (0.10% of peak) switch the system between fast and slow execution regimes with up to 7.3× overhead differences, driven by repeated re‑eviction of the same storages. On a ResNet‑32 trace, a deterministic feasibility inversion is observed: the run is feasible at a 0.101 budget ratio, infeasible (OOM) between 0.102–0.106, and feasible again from 0.107, caused by a fully pinned recursive rematerialization frontier exceeding the budget after all evictable tensors are removed. The authors attribute the LSTM instability to the joint size‑staleness scoring term and argue that these represent two distinct budget‑sensitive pathologies rather than a single mechanism.
By Mahesh Reddy Pagadala
The paper introduces Galahad, a memory layer that stores a transformer language model’s key‑value state for blocks of text, allowing subsequent requests to reuse previously computed attention rather than recomputing it. On seven real‑world datasets, 98.7% of prompt tokens were already read, and with Galahad the model could attend to an entire 97,000‑token corpus, achieving 98–100% recall on a 100‑fact test while reducing inference time and energy consumption dramatically. The approach was validated across 30 models and all runtimes, demonstrating that stateful inference can replace stateless serving without loss of accuracy.
By Sietse Schelpe
arXiv:2607. 19058v1 Announce Type: cross Abstract: Optimizer state is the largest single line item in the memory budget of mixture-of-experts (MoE) training: on a 6.
By Nuemaan Malik
arXiv:2607. 13124v1 Announce Type: cross Abstract: Structured pruning is a hardware-friendly way to compress LLMs, but it is mostly validated on multiple-choice recognition tasks, while the same compressed checkpoints can collapse on the free-form generation that deployment actually requires.
By Qingyu Zhang, Qianhao Yuan, Hongyu Lin, Yaojie Lu, Xianpei Han, Le Sun, Xiang Li, Ming Xu, Jiarui Li, Xiuyin Zhao
arXiv:2609.36953v1 Announce Type: new
Abstract: Production RL for language models lets the sampler fall behind the learner and repairs the resulting mismatch with a truncated importance weight. We as...
By Taiheng Pan