The paper introduces PPO‑FP16, a half‑precision variant of proximal policy optimization that incorporates three structure‑aware modules—global scalar protection, policy‑change residual protection, and efficient persistent‑update protection—to maintain numerical stability. These modules preserve the original PPO objective and hyperparameters while enabling reliable training on commodity GPUs. Empirical results on six continuous‑control tasks show PPO‑FP16 achieves the best mean return‑AUC rank, speeds up updates by up to 2×, and reduces optimizer‑state memory by up to 75% and peak GPU memory by up to 45%.
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...
Massively parallel simulation changes the data regime in which off-policy reinforcement learning (RL) is trained, challenging stabilizers designed for data-limited replay. Through controlled experimen...
arXiv:2610.01546v1 Announce Type: cross Abstract: Primal-dual hybrid gradient (PDHG) methods solve large-scale linear programs (LPs) using GPU-friendly matrix-vector products and projections, but the...
arXiv:2605. 05481v2 Announce Type: replace Abstract: We revisit a classic "chicken-and-egg" problem in reinforcement learning: to safely improve a policy, the value function must be accurate on the state-visitation distribution of the updated policy.
The paper introduces TACO, a new optimizer for fine‑tuning large language models that drastically reduces optimizer state memory while preserving first‑order gradients. TACO selects the sign of the largest magnitude entry in each column of weight matrices, achieving a 174× reduction in persistent optimizer memory compared to AdamW8bit and a 2.9× decrease in peak training memory on OPT‑13B. This allows full‑parameter fine‑tuning of 30–32B‑parameter models on a single 80 GB GPU across multiple model families and tasks, with comparable accuracy and runtime to existing methods.
arXiv:2608. 01997v1 Announce Type: new Abstract: Single-optimizer training is a poor fit for the distinct phases of deep network optimization: adaptive methods handle noisy early gradients well but overshoot flat minima, while SGD with momentum generalizes better in the late phase but converges slowly early on.
arXiv:2512. 06547v4 Announce Type: replace-cross Abstract: Decoupled PPO has been a successful reinforcement learning (RL) algorithm to deal with the high data staleness under the asynchronous RL setting.
The paper investigates how reusing past samples can improve the sample efficiency of Proximal Policy Optimization (PPO). Two variants, wPPO-U and wPPO-BH, are introduced within a multiple importance weighting framework, each reusing data from recent iterations while preserving core PPO mechanics. The authors derive theoretical policy improvement bounds for both variants and empirically evaluate their impact on continuous control tasks.
The paper introduces Adaptive Log‑Space (AL) quantization, a block‑wise representation that adapts the non‑zero range per block and preserves the exact‑zero invariant for non‑negative optimizer states. AL8 and AL16 are combined with signed‑momentum encodings and state‑specific precision choices, rather than a single policy for all states. Experiments on TinyLlama‑1.1B and GPT‑2 show that AL‑based quantization can match or exceed full‑precision performance while dramatically reducing optimizer‑state storage.
arXiv:2606. 17526v1 Announce Type: new Abstract: Efficient optimization is essential for training large language models.
The paper introduces a GPU-resident, batched Levenberg–Marquardt solver that efficiently optimizes constants in tree-based genetic programming for symbolic regression. By using reverse-mode automatic differentiation to assemble per-tree Jacobians in a single backward sweep, the solver’s per-iteration cost becomes independent of the number of constants per tree, achieving up to 510,000 trees per second on an NVIDIA A100. Integrated into EvoGP, the solver enables end-to-end search that recovers governing equations on 10 of 18 constructed problems, a significant improvement over stock EvoGP.
arXiv:2607. 14952v1 Announce Type: new Abstract: A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at deployment.