SPO++ is a reinforcement learning method designed to improve the efficiency of training language-model agents for reasoning and tool use. It addresses inefficiencies in "group-relative" reinforcement learning, which requires waiting for multiple outputs from the same prompt to be completed before performing an update. By building on Single-stream Policy Optimization (SPO), SPO++ enables asynchronous updates using only one rollout per prompt, while fixing technical mismatches in how data is processed and tracked.
Addressing Measure Mismatch
The researchers—Kai Ruan, Jinghao Lin, Qianshan Wei, Ziqi Zhou, and Zihe Huang—identify a core issue in how previous methods calculate advantages. In standard SPO, trajectory advantages are whitened (normalized) equally, but the actor loss is calculated as a mean across all tokens. Because response lengths vary, the "center" seen by the optimization process is skewed. SPO++ fixes this by standardizing advantages using the same action-token measure consumed by the actor loss. This ensures that every valid action token is weighted appropriately, which the authors identify as the strongest contributor to the method's performance gains.
Event-Time Prompt Tracking
SPO++ introduces a new way to track prompt evidence. Previous methods often relied on the order in which the learner received completed tasks, which can be inconsistent due to system timing. SPO++ instead organizes prompt evidence based on the "policy event"—the specific snapshot of the model that generated the request. By freezing the prompt value at the moment of dispatch and recording outcomes according to the policy event that produced them, the tracker becomes invariant to the order in which results are received.
Performance and Efficiency
The authors evaluated SPO++ on ALFWorld (using Qwen3.5-0.8B and 2B models) and Math-TIR. In these experiments, SPO++ consistently improved online learning efficiency compared to the original SPO. On ALFWorld, SPO++ showed significant gains in normalized reward-curve area, and Math-TIR results also showed consistent improvement. A paired ablation study confirmed that the action-token-measure normalization was the primary driver of these improvements, providing a 10.70 percentage point increase in the learning composite compared to trajectory-normalized variants.
Limitations and Considerations
The researchers note that their findings are based on online learning efficiency experiments using smaller model scales and limited training budgets. While the method effectively reduces dependency on group-based rollouts, it still requires repeatable task identities and offline initialization of prompt values. Additionally, the current implementation uses a "first-completed" collection strategy, which may favor shorter trajectories, and the method remains subject to the challenges of sparse rewards in cold-start scenarios. Future work could explore how these techniques scale to longer reasoning horizons and out-of-distribution tasks.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!