With external tools, large language models (LLMs) can tackle complex tasks through sustained environmental interaction. LLM-based agents are therefore increasingly used to tackle complex long-horizon...
arXiv:2608. 08255v1 Announce Type: new Abstract: Agentic reinforcement learning (RL) often suffers from delayed and sparse rewards in real-world environments.
Long-horizon large language model (LLM) agents are typically optimized with sparse terminal outcomes, making fine-grained credit assignment across multi-step interactions difficult. Existing approache...
arXiv:2608. 11967v1 Announce Type: cross Abstract: Large language model agents increasingly rely on long-horizon reasoning to solve complex tasks involving planning, tool use, and memory.
Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning proposes TreeOPSD, a critic‑free framework that treats each interaction turn as a tree node, branches alternative continuations, and recursively backs up terminal rewards to generate implicit turn‑level credit. The method maps this credit to a bounded teacher distribution based on the likelihood ratio to a frozen behavior policy, and minimizes the divergence between teacher and student to recover policy‑gradient directions while limiting updates across mini‑batches. Empirical results show TreeOPSD enhances agentic performance on long‑horizon benchmarks, notably improving Qwen3‑30B‑A3B‑Thinking‑2507’s accuracy by 4.97 points over GRPO in the Deep Research setting.
arXiv:2607. 27973v1 Announce Type: new Abstract: Recently, Reinforcement Learning (RL) has emerged as a crucial paradigm for the post-training of Large Language Model (LLM) agents.
TreeHCA introduces a method that couples hindsight-guided exploration with hindsight credit assignment for multi-turn large language model agents. By using changes in reference-answer likelihood to revisit intermediate histories and sample alternative continuations, it builds a tree of reasoning and tool interactions, and applies an inverse-hindsight importance correction to propagate terminal outcomes into turn-level advantages. Experiments on seven question-answering benchmarks with three Qwen backbones show that TreeHCA outperforms all baselines in aggregate exact match.
EmbodiedMind introduces a three-stage training paradigm for embodied foundation models that tackles inefficient sample use, task imbalance, and credit assignment in long-horizon planning. The stages—Rejection Sampling-based Fine‑Tuning, Iterative Rejection GRPO, and Trie‑GRPO—filter low‑informative data, balance task difficulty, and use action prefix trees for step‑level advantage estimation. This approach yields a state‑of‑the‑art average performance of 70.02% across 18 benchmarks, notably improving long‑horizon task planning accuracy.
The paper introduces Effect-Marginal Policy Optimization (EMPO), a method that addresses credit-update misalignment in multi-turn reinforcement learning for large language model agents. EMPO aligns policy updates with execution-level equivalence by using an effect-level surrogate that couples updates across different textual responses that produce the same effect. Experiments on ALFWorld, WebShop, and Sokoban demonstrate that EMPO improves both task performance and learning efficiency compared to strong baselines.
TrajEvolve introduces a trajectory evolution paradigm with hindsight credit assignment to improve reinforcement learning for large language models on long-horizon tasks. It refines low-quality steps in historical trajectories to expose the policy to high-quality compositions that are rarely sampled, and reformulates step-level continuous credit assignment as a binary classification problem to identify redundant steps. Experiments on two datasets show that this approach outperforms existing methods.
arXiv:2606. 10507v1 Announce Type: new Abstract: While Large Language Models (LLMs) have demonstrated strong capabilities as autonomous agents across a wide range of tasks, their performance often degrades in multi-turn long-horizon agentic tasks.
Group-Graph Policy Optimization (G2PO) is a novel group-based reinforcement learning algorithm designed for long-horizon agentic tasks. It transforms linear interaction trajectories into a global state‑transition graph, aggregates identical observations across trajectories to reduce sampling variance, and redefines actions as transitions between state nodes. By employing an edge‑centric advantage estimation and globally standardizing TD errors, G2PO identifies and prioritizes critical transitions, achieving up to 22.2% higher success rates than state‑of‑the‑art baselines on benchmarks such as WebShop, ALFWorld, and AppWorld.