aiqbench analysis
This paper introduces Iterative GRPO, a method that enables principled multi-turn reinforcement learning for LLM agents by leveraging single-turn RLHF techniques without requiring interactive user simulators. By alternating between policy evaluation using learned Q-functions and policy improvement via standard GRPO, the approach effectively handles the "batch online" deployment pattern common in production environments. The technique demonstrates improved efficiency on complex negotiation tasks by scoring responses based on expected downstream returns rather than relying on full trajectory rollouts.
Iterative GRPO: Batch-Online Policy Iteration for Multi-Turn RL via Single-Turn RLHF
arXiv:2511.21638v2 Announce Type: replace Abstract: Practical LLM agents often operate over multi-turn conversations where success is determined only after the full interaction ends. Most multi-turn RL methods train via on-policy rollouts, but unlike in single-turn RLHF, the policy cannot produce a trajectory alone, since an external environment must respond after each agent turn. For conversational agents, this environment is a user, but real users are generally unavailable inside the training loop and simulated users are difficult to build faithfully. Separately, real-world deployment is rarely fully online or fully offline. The common production pattern is called "batch online," where the current policy is deployed to collect a batch of interaction data, then retrained on that batch and redeployed. We show that this batch-online setting creates an opportunity for applying the classical approximate policy iteration algorithm. Our central observation is that running standard token-level, single-turn GRPO with a learned turn-level $Q$-function as the reward model is a policy improvement step for the multi-turn problem. Building on this, we present Iterative GRPO, which alternates between fitting $Q^\pi$ from logged Monte Carlo returns via standard reward modeling (policy evaluation) and running single-turn GRPO against $Q^\pi$ (policy improvement). During policy improvement, the learned $Q^\pi$ scores candidate responses by their expected downstream return, avoiding the need for simulator rollouts of the remaining conversation. Overall, this allows us to do principled multi-turn RL using only single-turn RLHF methods without an interactive environment or user simulator inside the training loop. We demonstrate the effectiveness of Iterative GRPO on six multi-turn negotiation environments.
Abstract:Practical LLM agents often operate over multi-turn conversations where success is determined only after the full interaction ends. Most multi-turn RL methods train via on-policy rollouts, but unlike in single-turn RLHF, the policy cannot produce a trajectory alone, since an external environment must respond after each agent turn. For conversational agents, this environment is a user, but real users are generally unavailable inside the training loop and simulated users are difficult to build faithfully. Separately, real-world deployment is rarely fully online or fully offline. The common production pattern is called "batch online," where the current policy is deployed to collect a batch of interaction data, then retrained on that batch and redeployed. We show that this batch-online setting creates an opportunity for applying the classical approximate policy iteration algorithm. Our central observation is that running standard token-level, single-turn GRPO with a learned turn-level $Q$-function as the reward model is a policy improvement step for the multi-turn problem. Building on this, we present Iterative GRPO, which alternates between fitting $Q^\pi$ from logged Monte Carlo returns via standard reward modeling (policy evaluation) and running single-turn GRPO against $Q^\pi$ (policy improvement). During policy improvement, the learned $Q^\pi$ scores candidate responses by their expected downstream return, avoiding the need for simulator rollouts of the remaining conversation. Overall, this allows us to do principled multi-turn RL using only single-turn RLHF methods without an interactive environment or user simulator inside the training loop. We demonstrate the effectiveness of Iterative GRPO on six multi-turn negotiation environments.
Machine Learning (cs.LG)
arXiv:2511.21638 [cs.LG]
(or arXiv:2511.21638v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2511.21638 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 341 words
Related coverage
Introducing North Mini Code: Cohere’s First Model For Developers
Hugging Face · Jun 9, 2026
Granite 4.2 LLMs: How They're Built
Hugging Face · Aug 25, 2026
AgentOps: Operationalize agentic AI at scale with Amazon Bedrock AgentCore
AWS · Jun 1, 2026
[AINews] not much happened today
Latent Space · Sep 18, 2026