Building LLMs that behave well socially, not merely correctly, requires Building LLMs that behave well socially, not merely correctly, requires more than producing locally helpful responses. A socially competent agent must infer users' unstated goals, respect their preferences, and adapt as the conversation unfolds. These behaviors are inherently multi-turn and social, making them hard to optimize: real interaction data is scarce, and user preferences are typically latent rather than directly observable. To address these challenges, we build on a persona-driven social simulation environment (consisting of a persona library, LLM-based user simulators, and a user-satisfaction scoring system ranging from [0, 1]), to introduce preference-batched GRPO (PB-GRPO), a post-training algorithm that learns socially adaptive policies from conversation-level feedback. Compared to vanilla GRPO, PB-GRPO computes advantages using a normalization estimated across a bucket of users with similar preferences, stabilizing training across a diverse social population. Empirical evidence shows that PB-GRPO improves models' social behavior over strong reinforcement learning baselines in our simulated environment.
Figures & tables
Figure 1: Overview of persona-driven simulation and PB-GRPO training. Conversation-level rewards are normalized within preference persona buckets for policy optimization.
Training Method
Max Reward
Success Rate
REINFORCE++
0.5149
0.3850
Vanilla GRPO
0.8625
0.5463
PB-GRPO
0.9675
0.6375
Table 1: Social adaptation metrics. Max Reward : highest preference satisfaction score achieved. Success Rate : fraction of dialogues where the agent successfully adapted to hidden preferences ( r>0.70 ).
Figure 2: Training dynamics of the three training runs. Left: mean training reward (rollout return) over training steps. Right: test-time success rate on held-out personas. PB-GRPO improves faster and more stably than vanilla GRPO and REINFORCE++. Training dynamics with extended training steps can be found in Appendix B
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Extended mean training-reward trajectories over 400 optimization steps. Light traces show raw per-step mean rollout returns; solid curves show their time-weighted exponential moving averages (TW-EMAs). PB-GRPO continues to improve over the extended training horizon and maintains a higher reward trend than vanilla GRPO and REINFORCE++.