arXiv:2606. 16733v1 Announce Type: new Abstract: Policy gradient algorithms for language models optimize the same objective $J(\theta) = \mathbb{E}*{\tau \sim p*\theta(\tau)}[R(\tau)]$, which has exactly two factors: the trajectory probability $p_\theta(\tau)$ and the reward $R(\tau)$.
arXiv:2601. 03895v2 Announce Type: replace-cross Abstract: Group Relative Policy Optimization (GRPO) has emerged as a popular algorithm for reinforcement learning with large language models (LLMs).
arXiv:2607. 18163v1 Announce Type: cross Abstract: PPO and the GRPO baseline studied here use clipped surrogate objectives whose favorable-direction saturation introduces an abrupt change in the scalar objective's derivative.
arXiv:2604. 01499v2 Announce Type: replace Abstract: Evolution Strategies (ES) have emerged as a scalable gradient-free alternative to reinforcement learning based LLM fine-tuning, but it remains unclear whether comparable task performance implies comparable solutions in parameter space.
arXiv:2606. 05434v1 Announce Type: new Abstract: Group Relative Policy Optimisation (GRPO) has emerged as an effective reinforcement-learning algorithm for aligning language models on reasoning tasks, but it treats every token position and every sampled rollout symmetrically.
The paper introduces GRAFT, a Graph-based Faithful sTep-level credit-assignment framework that constructs a trajectory graph from rollout trajectories, recovers node state-values via Bellman iteration, and assigns step-level advantages based on node value differences. It also proposes Graph GAE to further reduce state-value estimation bias. Experiments on multi-turn agentic benchmarks demonstrate consistent improvements over GRPO and other recent agentic RL algorithms.
arXiv:2607. 19313v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) improves reasoning in large language models.
Reinforcement learning from verifiable rewards (RLVR) for mathematical reasoning suffers from a structural blind spot: on "cliff" prompts-those on which every sampled rollout in a group fails-the group-normalized advantage is identically zero, so GRPO produces no gradient on precisely the prompts at the frontier of the model's capability. We introduce LoRA Scaffolded Policy Optimization (LSPO), a sampling-time mechanism that recovers this lost gradient.
arXiv:2609.26355v1 Announce Type: new Abstract: Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted ma...
arXiv:2511. 02577v2 Announce Type: replace Abstract: Proximal Policy Optimization (PPO) is widely regarded as one of the most successful deep reinforcement learning algorithms, known for its robustness and effectiveness across a range of problems.
arXiv:2607. 27787v1 Announce Type: new Abstract: Reinforcement learning from verifiable rewards (RLVR) for mathematical reasoning suffers from a structural blind spot: on "cliff" prompts-those on which every sampled rollout in a group fails-the group-normalized advantage is identically zero, so GRPO produces no gradient on precisely the prompts at the frontier of the model's capability.
The paper investigates why Group Relative Policy Optimization (GRPO) benefits from per‑prompt normalization by examining the local curvature of the sequence‑level policy gradient. It shows that standard deviation normalization acts as an adaptive gradient, yielding a provably faster convergence rate than unnormalized REINFORCE under mild conditions, with the improvement tied to the average within‑prompt reward standard deviation. The authors also propose IS‑GRPO, an importance‑sampling variant that maintains alignment with the full gradient and offers a tighter convergence guarantee, and empirically validate these theoretical insights on GSM8K and MATH datasets at 1.5B and 7B model scales.
Iterative GRPO is a batch‑online policy iteration framework that enables multi‑turn reinforcement learning for conversational agents without requiring an interactive user simulator. It alternates between learning a turn‑level Q‑function from logged returns (policy evaluation) and applying single‑turn GRPO against this Q‑function (policy improvement), thereby scoring candidate responses by their expected downstream return. The method is validated on six multi‑turn negotiation environments, demonstrating its practicality for real‑world deployment patterns.