The paper investigates whether idle inference resources can help cut the high cost of scarce GPU usage during training. Using a simulated compute ledger that bills fleet work at a fraction of a GPU forward pass, the authors propose an algorithm that predicts gradients with a low‑precision, inference‑style reverse‑mode program and then refines these predictions with a few exact gradients via a control variate, turning approximation error into variance rather than bias. Experiments on a 124‑million‑parameter language model and across models ranging from 10 M to 774 M parameters show that the method can reduce simulated ledger cost when fleet work is cheap, though it also exhibits both successful transfers and failures, and does not evaluate inference‑only hardware or full optimizer‑by‑batch‑size sweeps.
arXiv:2606. 08854v1 Announce Type: cross Abstract: Standard Reinforcement Learning with Verifiable Rewards (RLVR) training allocates a fixed rollout budget to every query, without regard for what each query's difficulty means for the current policy.
arXiv:2606. 16059v1 Announce Type: cross Abstract: For thirty years, quantitative finance has paid a costly two-language tax: models researched in Python are rewritten in C++ for production, often introducing numerical discrepancies.
arXiv:2607. 29398v1 Announce Type: new Abstract: Diffusion models have revolutionized generative tasks but incur high latency due to iterative denoising.
The paper introduces Open-1B, a language model trained under a new fully auditable regime that ensures every training operation is reproducible on heterogeneous commodity hardware with bitwise certainty. By enforcing a fixed order on sources of nondeterminism—GPU reductions, data batch ordering, and inter/intra-node communication—the authors enable auditors to replay and verify individual training steps on a single machine. The release includes the full pretraining dataset, all intermediate checkpoints, the training codebase, and an audit harness for step-by-step verification.
arXiv:2608. 07524v1 Announce Type: new Abstract: Training deep learning models on variable long sequences poses significant computational challenges.
arXiv:2605. 21312v2 Announce Type: replace-cross Abstract: Modern LLM serving is no longer homogeneous or monolithic.
TIDE (Temporal Incremental Draft Engine) is a serving‑engine‑native framework that integrates online draft adaptation into high‑performance LLM inference. By reusing intermediate hidden states from the target model as training signals, TIDE avoids extra target model computation and serving‑time overhead, activating speculation and draft training only when beneficial. On heterogeneous GPU clusters, TIDE achieves up to 1.66× higher throughput than no‑speculation baselines, reduces training time by up to 3.02×, cuts storage needs by 24×, and improves system throughput by up to 1.22×.
arXiv:2604. 09731v2 Announce Type: replace-cross Abstract: Tree-based speculative decoding accelerates autoregressive generation by verifying a branching tree of draft tokens in a single target-model forward pass.
arXiv:2608. 08020v1 Announce Type: new Abstract: Test-time compute scaling is a primary driver of performance in large reasoning models (LRMs), but extreme inefficiency bounds current approaches, shifting the critical question from \emph{how much} compute to spend, to \emph{where} to allocate it.
arXiv:2608. 08961v1 Announce Type: new Abstract: AI training's rising resource intensity is straining electricity supplies and carbon budgets, motivating systematic study of memory-efficient training on constrained hardware.
arXiv:2608.24479v1 Announce Type: new Abstract: Massively parallel simulation changes the data regime in which off-policy reinforcement learning (RL) is trained, challenging stabilizers designed for...
arXiv:2606. 18431v1 Announce Type: new Abstract: LLM serving exhibits extreme length variability, making size-based scheduling difficult in practice.