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.
arXiv:2607. 01083v1 Announce Type: cross Abstract: High-throughput RLHF systems often decouple rollout generation from policy optimization, leading to the use of stale rollouts during learner updates.
arXiv:2606. 04560v1 Announce Type: cross Abstract: Reinforcement learning from verifiable rewards with GRPO is a standard approach for post-training reasoning LLMs.
arXiv:2610.09519v1 Announce Type: new Abstract: Production machine-learning models are derived artifacts of time-bounded training snapshots: a deployed model is a materialized view over a training cu...
arXiv:2606. 08446v1 Announce Type: cross Abstract: Despite being powerful, reinforcement learning with verifiable rewards (RLVR) induces extremely long COT, making it computationally expensive.
The paper investigates why reinforcement learning with verifiable rewards (RLVR) reduces the diversity of solutions in reasoning tasks. By analyzing the Countdown task, the authors show that RLVR contracts the solution space mainly at the entrance—before the first arithmetic operation—causing a 67% drop in solution coverage. They demonstrate that providing an unselected entrance prefix or applying entrance‑targeted interventions can restore or even improve coverage without harming accuracy.
arXiv:2602. 00781v2 Announce Type: replace Abstract: Online reinforcement learning in non-episodic, finite-horizon MDPs remains underexplored and is challenged by the need to estimate returns to a fixed terminal time.
arXiv:2608.29296v1 Announce Type: cross Abstract: Larger batches reduce the variance of stochastic gradients per update and are therefore often expected to accelerate training. Yet whether this stati...
arXiv:2606. 26997v1 Announce Type: cross Abstract: Large language model (LLM) post-training for reasoning increasingly relies on reinforcement learning with verifiable rewards (RLVR), where models learn from ground-truth feedback on mathematical, logical, and scientific tasks.
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.
arXiv:2610. 13901v1 Announce Type: cross Abstract: We study the last iterate of standard asynchronous Q-learning from a single trajectory of a finite Markov decision process.
arXiv:2602.06797v3 Announce Type: replace-cross Abstract: We study optimal learning rate (LR) schedules under the functional scaling law (FSL) framework (Li et al., 2025), which decomposes training d...
Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but staleness is an inevitable byproduct compounded by policy lag, engine delays, and mixture-of-experts routing. From a trust-region perspective, this mismatch is critical: training-inference divergence governs approximation error in finite-horizon bounds, whereas PPO clipping only gates sampled outward updates, acting as a sampled surrogate rather than a full-policy constraint.