The paper studies learned schedules for a continuous penalty relaxation of the MAX-CUT problem by backpropagating through an unrolled classical solver. It compares three architectural variants: a state‑blind schedule, a state‑adaptive bilinear schedule, and an additive ablation. The bilinear schedule generalizes well to larger, out‑of‑distribution instances and outperforms its untrained baseline, whereas the additive model performs similarly to the bilinear model on‑distribution but fails to generalize out‑of‑distribution.
arXiv:2609.13922v1 Announce Type: new Abstract: Minibatch persistency reuses data instead of reading it: rather than drawing a fresh minibatch at every optimizer step, it takes K consecutive steps on...
arXiv:2608.24593v1 Announce Type: new Abstract: Adaptive optimizers retain gradient history in moment variables, allowing a local change in loss weighting to alter later updates. We examine whether t...
arXiv:2609.39634v1 Announce Type: cross Abstract: Common policy improvement methods, including TRPO, PPO, and GRPO, estimate policy improvement under the behavioral policy's state-visitation distribu...
The paper investigates reinforcement learning with verifiable rewards (RLVR), where costly model-generated responses encourage larger rollout batches and a square‑root learning‑rate scaling. Using a two‑clock framework, the authors show that while the square‑root rule preserves first‑order drift per sample, larger batches suffer from curvature losses, leading to a critical batch size beyond which sample efficiency degrades. Experiments on Qwen2.5‑Math‑1.5B confirm the predicted progress gap and failure clock, and the authors propose a simple multi‑step RLVR schedule that starts below the critical batch and switches to a larger batch before failure.
The paper introduces OrbitTrace, a benchmark of 50 physics‑grounded compute‑availability traces from satellite orbits, and investigates whether specialized interruption‑resilient optimizers are needed when training is interrupted by predictable compute gaps. Experiments on CIFAR‑10/ResNet‑18 and GPT‑2/AdamW show that a strong checkpoint‑and‑resume baseline that preserves full optimizer state and indexes learning‑rate schedules in effective time matches uninterrupted training, rendering most availability‑aware methods unnecessary. Only in a narrow regime—large models with non‑persistable optimizer state and frequent short pauses—does reactive adaptation recover a modest portion of the state‑loss penalty, and even this benefit disappears for eclipse‑scale gaps.
Spectral Prioritized Sweeping (SPS) extends traditional Prioritized Sweeping by incorporating graph topology through the resolvent and Laplacian diffusion, creating a smoother priority score that propagates reward changes more effectively in nonstationary reinforcement learning. The method, called Graph Topology Augmentation for Prioritized Sweeping (GTA-PS), uses a mixing of regularized Laplacian inverses and an adaptive scheduler based on the Second Largest Eigenvalue Modulus to adjust the influence of topology during replanning. Experiments on FourRooms and GARNET domains show that GTA-PS improves replanning efficiency compared to standard PS under both exact dynamic programming and Dyna-style planners.
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.
arXiv:2609.23594v1 Announce Type: new Abstract: Orthogonality in a LoRA factor does not by itself specify what the composed update protects: the answer depends on the task-start state, the parameteri...
arXiv:2609.13739v1 Announce Type: cross Abstract: Language-model agents are increasingly deployed through diverse harnesses that differ in system prompts, tool schemas, control loops, and trajectory...
arXiv:2610. 27628v1 Announce Type: new Abstract: Looped language models can run more iterations at test time, but additional iterations need not help them solve longer problems.
arXiv:2609.24678v1 Announce Type: new Abstract: Continual learning with Low-Rank Adapters (LoRA) typically mitigates forgetting by penalizing the overlap between a new update and the accumulated past...
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.