arXiv:2512. 23075v5 Announce Type: replace-cross Abstract: Policy gradient methods for Large Language Models optimize a policy $\pi_\theta$ via a surrogate objective computed from samples of a rollout policy $\pi_{\text{roll}}$.
arXiv:2602. 04879v3 Announce Type: replace-cross Abstract: Reinforcement learning (RL) has become a cornerstone for fine-tuning Large Language Models (LLMs), with Proximal Policy Optimization (PPO) serving as the de facto standard algorithm.
arXiv:2607. 10848v1 Announce Type: new Abstract: Reinforcement learning for large language models (LLMs) typically relies on trust-region masks to stabilize off-policy updates.
arXiv:2606. 09821v1 Announce Type: new Abstract: Reinforcement learning (RL) has become a key component of post-training large language models (LLMs).
arXiv:2607. 22186v2 Announce Type: replace Abstract: Asynchronous reinforcement learning (RL) accelerates large language model (LLM) post-training by overlapping rollout generation with policy optimization, but the resulting stale, off-policy data can destabilize optimization and ultimately cause policy collapse.
Reinforcement learning for large language models (LLMs) typically relies on trust-region masks to stabilize off-policy updates. The dominant PPO-style approach uses the sampled-token importance ratio for two criteria: a proximity criterion, which asks whether the policy has moved too far from the behavior policy, and a direction criterion, which asks whether the update pushes it farther away.
arXiv:2605. 25582v2 Announce Type: replace Abstract: Reinforcement learning for large language models faces a fundamental trade-off between sample efficiency and asymptotic performance: strictly on-policy methods discard trajectories after a single update, while off-policy reuse introduces distribution mismatch that existing trust-region techniques mitigate primarily by enforcing conservative optimization, often leaving rich training signals underutilized.
The paper presents a convergence framework for deep $V$‑learning over a finite horizon $H$, deriving explicit bounds on policy loss by decomposing the Bellman update error into six residuals. It shows how $L^s$ concentrability controls expected $L^1$ loss, quantifies the impact of shared sampling across horizon levels, and provides optimal and near‑optimal sample allocations for statistical error rates. The work also establishes sharp action‑gap bounds under a margin condition, transfers optimal‑gap results to frozen‑iterate gaps, and offers consistency guarantees for generative‑reset approximate‑ERM procedures with exact action scores.
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.
arXiv:2609.07666v1 Announce Type: new Abstract: Full-parameter fine-tuning of large language models has substantial memory costs because backpropagation stores activations and gradients. Zeroth-order...
The paper introduces Actor‑Critic with Action Chunking (AC2), a method that assigns credit to short action chunks instead of entire trajectories, enabling policy updates without waiting for terminal rewards. AC2 employs local readiness, reference solutions, and 10k‑token chunks to make critic‑based credit assignment reliable. Experiments on Qwen3‑4B with FineProofs‑RL show AC2 surpasses GRPO’s peak validation score while using 2.5× fewer decoding FLOPs and fewer training steps.
arXiv:2607. 18722v1 Announce Type: new Abstract: 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.
EPIG-Tree proposes a compute‑optimal branching strategy for gradient‑efficient reinforcement learning, arguing that branches should be placed where they most reduce policy‑gradient uncertainty per unit of compute. By deriving allocation laws from a law‑of‑total‑variance decomposition, the method introduces an EPIG‑Tree score that guides branch placement using already computed rollouts, estimating occupancy‑ and score‑weighted value uncertainty. Empirical results show EPIG‑Tree reduces gradient MSE in cloned‑state control, improves frozen‑LLM gradient calibration, and outperforms flat GRPO and entropy branching in both single‑turn math and multi‑turn Wordle tasks.