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.
arXiv:2607. 14171v1 Announce Type: new Abstract: Reinforcement learning has emerged as the dominant paradigm for training large language model (LLM) agents that interact with executable sandboxes.
arXiv:2609.40035v1 Announce Type: new Abstract: The policy-gradient theorem gives the exact gradient under the current policy, but finite on-policy samples may miss rare high-return trajectories. We...
Contrastive Branch Policy Optimization (CBPO) is a reinforcement learning method that separates the allocation of a fixed rollout budget from the translation of branch outcomes into token-level credit. It uses generation entropy to screen branch positions, path- and node-level decay to distribute the budget, and Contrastive Branch Value (CBV) to estimate local decision sensitivity without changing reward signs. CBPO partitions trajectories into non-overlapping credit segments, preventing duplicated gradients and enabling fine-grained credit assignment using only outcome rewards.
arXiv:2606. 25451v1 Announce Type: new Abstract: Estimating token-level advantages in reinforcement learning (RL) for language models remains challenging because scaling up episodic experience collection is expensive.
arXiv:2608.21501v1 Announce Type: new Abstract: Credit assignment in large-language-model reinforcement learning (LLM RL) can be separated into three objects: evidence about success, a transport oper...
The paper introduces belief‑shift branching, a method for placing forks in tree‑structured reinforcement learning rollouts by identifying points where a model’s answer belief changes most. Unlike traditional structural or entropy‑based approaches, belief‑shift uses a probe, logit‑lens depth profile, or learned activation direction to locate pivots in the value curve, incurring minimal computational overhead. Experiments across multiple models and benchmarks show that belief‑shift forking consistently outperforms baseline methods, yielding significant gains in mathematics and code tasks.
The paper introduces belief‑shift branching, a method for placing forks in tree‑structured reinforcement learning rollouts by detecting where a model’s answer belief changes most sharply. Unlike traditional fixed‑length or entropy‑based forking, this approach uses a lightweight probe or learned activation direction to identify pivots in the value curve, reducing unnecessary sampling. Experiments show that belief‑shift forking consistently outperforms baseline methods across multiple models and benchmarks, yielding significant gains in mathematics and code tasks.
arXiv:2606. 09961v1 Announce Type: cross Abstract: Training large language models (LLMs) as autonomous agents via reinforcement learning (RL) has enabled frontier models to achieve superhuman performance in long-horizon tasks.
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:2606. 29476v1 Announce Type: cross Abstract: Self-distilled agentic reinforcement learning augments trajectory-level reward with a token-level distillation loss, using as its teacher the same policy conditioned on privileged context.
BCPPO is a new variant of Proximal Policy Optimization that uses Bachelier-inspired cost‑prediction networks to generate a smooth penalty based on disagreement among critics. The method keeps temporal‑difference learning unchanged, applies a saturation‑aware controller to manage cost penalties, and deploys only the policy network. Across extensive experiments, BCPPO outperforms comparators in achieving higher mean returns while maintaining lower or comparable CVaR in all tested tasks.
The paper investigates the problem of sharing a single critic across multiple parallel environments in reinforcement learning. It shows that when environments assign different expected returns to the same state, a shared critic must reconcile conflicting value targets, which can distort advantage estimates and misguide policy updates. The authors propose a simple fix—providing the critic with the environment index—demonstrating through bandit models and experiments on CartPole, MuJoCo, BipedalWalker, and 16 Procgen games that this conditional critic stabilizes learning and boosts returns, achieving a 40.8% improvement in aggregate normalized return on unseen levels.