liyuxuan-academic/behavior_grounding

Behaviorally Grounded User Profiles from the Wild

Open-ended, anonymized user profiles distilled from authentic social-media behavior, released with the paper "Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning."

Persona-driven methods for personalizing LLMs typically rely on rigid synthetic personas built from a small set of categorical attributes (age, gender, nationality). These flatten individual variation and lean on stereotypes. This dataset instead provides open-ended profiles extracted from real behavioral traces: short, coherent textual bios synthesized from users' historical social-media posts. Alongside them we release a synthetic baseline generated by prompting an LLM, so the two can be compared directly.

How the data was created

Dataset structure

Two configurations, each a CSV with the same schema:

Column	Type	Description
user_id	string	Pseudonymous random UUID; no real handle
user_profile	string	A short free-text bio describing the user's interests and traits

Config	File	Rows	Description
open_ended	open_ended_profiles.csv	824	Behaviorally grounded profiles extracted from real Bluesky post histories
synthetic_baseline	synthetic_baseline_profiles.csv	842	Purely synthetic profiles from LLM prompting

Open-ended profiles average ~116 words; synthetic baseline profiles ~210 words.

Example (open-ended)

{
  "user_id": "5078790f-63fe-4d2a-a116-60ec37c17263",
  "user_profile": "The person is a nature-loving individual who enjoys flowers, music, and a wide variety of foods, with a particular fondness for chicken."
}

Results

Downstream results for the Qwen3 models, comparing No Profile (base model), the Synthetic baseline, and our Open-Ended behaviorally grounded profiles. RecBench columns (Netflix, Books, News) report F1; URS columns (Leisure, Creativity, Advice, Avg.) report the 1–10 LLM-judge score. Higher is better; bold = best per column within each model. (Full results with additional models are in the paper and the accompanying code repository.)

Model	Variant	Netflix (F1)	Books (F1)	News (F1)	Leisure	Creativity	Advice	Avg.
Qwen3-8B	No Profile	0.421	0.515	0.318	5.48	5.31	5.99	5.59
	Synthetic	0.420	0.625	0.319	6.34	6.72	7.08	6.72
	Open-Ended	0.450	0.649	0.322	6.76	7.40	7.65	7.27
Qwen3-14B	No Profile	0.419	0.308	0.303	7.49	7.90	8.06	7.82
	Synthetic	0.416	0.538	0.327	7.06	7.54	7.91	7.50
	Open-Ended	0.459	0.632	0.321	7.29	8.10	8.27	7.88
Qwen3-32B	No Profile	0.403	0.569	0.308	6.79	7.09	6.90	6.93
	Synthetic	0.427	0.580	0.317	7.20	7.87	8.06	7.71
	Open-Ended	0.455	0.658	0.315	7.35	8.06	8.23	7.88

Profile diversity

Birth-location distribution of the baseline synthetic profiles (left) and our open-ended behaviorally grounded profiles (right). Synthetic personas collapse toward a narrow set of nationalities, while the open-ended profiles maintain a long-tailed, representative distribution (top-12 countries shown; see the paper for the full comparison and categorical entropy analysis).

Source data & licensing

Profiles are derived from the public "2 Million Bluesky Posts" corpus, released under Apache 2.0. Collection followed the platform's Terms of Service and API guidelines. This derived dataset is released under Apache 2.0.

Citation

@inproceedings{behaviorally_grounded_profiles,
  title     = {Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning},
  author    = {PLACEHOLDER},
  booktitle = {PLACEHOLDER},
  year      = {2026}
}

Please also cite the source corpus (Alpin Dale, "2 Million Bluesky Posts", 2024).