Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning
Authors: Yuxuan Li, Victor Zhong, Ehsan Kamalloo
Publication Date: August 11, 2026 (arXiv:2609.00014 [cs.CL])
Abstract
Persona-driven techniques are increasingly used to adapt large language models (LLMs) to diverse contexts. However, existing methods largely rely on rigid, synthetic personas that flatten individual variation, depend on stereotypes, and fail to capture the nuanced signals that drive real human preferences. To address this, we introduce profile behavioral grounding, a framework that extracts open-ended, high-fidelity user profiles directly from authentic, anonymized social media posts.
We evaluate these profiles across two complementary paradigms: train-time personalization via supervised fine-tuning (SFT) and non-parametric, test-time multi-perspective reasoning. Across benchmarks for complex recommendation and open-ended queries, behaviorally grounded profiles consistently improve base model performance and outperform synthetic profile baselines. These profiles not only enhance parametric alignment but also enable richer, multifaceted reasoning.
Our findings establish open-ended, behavior-derived profiles as a highly diverse and effective foundation for the next generation of personalized language systemsβespecially relevant as 2026 sees growing demand for adaptive AI that respects individual differences. Our codebase is publicly available at github.com/ServiceNow/behavior-grounding.
Key Contributions
- Profile Behavioral Grounding: A novel framework for constructing user profiles directly from real-world social media data, preserving behavioral nuance rather than relying on synthetic archetypes.
- Two-Paradigm Evaluation: Demonstrates the effectiveness of these profiles in both supervised fine-tuning (train-time) and multi-perspective reasoning (test-time) settings.
- Strong Empirical Gains: Outperforms synthetic-profile baselines on complex recommendation and open-ended query benchmarks, with improvements across multiple LLM architectures.
Methodology
- Data Collection & Anonymization: Aggregates social media posts with strict privacy safeguards to ensure anonymous yet authentic behavioral signals.
- Profile Extraction: Uses LLM-based pipelines to distill open-ended user profiles from post histories, capturing preferences, communication styles, and reasoning patterns.
- Personalization Paradigms:
- Train-time SFT: Fine-tune base LLMs on user-conditioned instruction data to embed profile knowledge into model parameters.
- Test-time multi-perspective reasoning: Non-parametric retrieval of relevant profiles at inference, enabling dynamic, context-aware responses.
Results & Impact
Across diverse benchmarks, behaviorally grounded profiles consistently improved accuracy and user satisfaction metrics over both non-personalized baselines and synthetic persona methods. The approach proves particularly effective for tasks requiring subjective judgment, such as content recommendation and open-ended Q&A.
Conclusion & Outlook
As personalization becomes a core expectation for LLMs in 2026 and beyond, this work highlights the value of grounding user models in real behavioral data. Future work will explore cross-domain generalization, privacy-preserving profile representations, and integration with reinforcement learning from human feedback (RLHF) pipelines.
Keywords: personalization, user profiling, behavioral grounding, large language models, supervised fine-tuning, multi-perspective reasoning, social media, synthetic personas
via ArXiv CL+LG