Behaviorally Grounded User Profiles from the Wild

Code for the paper "Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning."

We introduce profile behavioral grounding: a framework that extracts open-ended, high-fidelity user profiles directly from authentic, anonymized social-media activity, and uses them to personalize language models. Unlike rigid synthetic personas built from a few categorical attributes, these behavior-derived profiles capture the nuanced, idiosyncratic signals that drive real preferences. We study them under two paradigms:

• Train-time personalization — profile-conditioned synthetic data for supervised fine-tuning (SFT) on recommendation and open-ended QA.
• Test-time multi-perspective reasoning — a non-parametric approach that samples semantically relevant profiles as candidate "experts" and aggregates their viewpoints into a single multifaceted answer.

Repository layout

conf/                     Hydra configuration
  config.yaml             top-level defaults
  model/                  model configs (Qwen3 8B/14B/32B, Gemma3 4B, Olmo3 7B, GPT-OSS 20B/120B, Qwen3 embedding)
  inference/              inference backends (vLLM, Transformers, OpenRouter)
  task/
    persona/              profile extraction (single/multi-tweet, dedupe/summarize)
    data/                 profile-conditioned SFT data synthesis + synthetic-baseline generation
    finetuning/           supervised fine-tuning
    evaluate/             URS scoring, finetuned-persona evaluation, persona prediction, profile QA
    test_time/            multi-perspective sampling, aggregation, scoring
pipeline/                 implementation for each task category (entry points referenced by the configs)
main.py                   Hydra entry point
env.yaml                  conda environment

Setup

conda env create -f env.yaml
conda activate behavior_grounding

Requires Python 3.11 and CUDA-capable GPUs for training/inference.

Environment variables

Configs reference these paths via environment variables (no absolute paths are hardcoded):

Variable	Meaning
PWD	Project root (set automatically to the current directory)
MODEL_ROOT	Directory containing local model weights
DATA_ROOT	Directory containing external datasets (see below)

export MODEL_ROOT=/path/to/models
export DATA_ROOT=/path/to/datasets

External datasets

These are not redistributed here; obtain them from their original sources and place them under $DATA_ROOT:

• 2 Million Bluesky Posts (Dale, 2024, Apache 2.0) — raw posts for profile extraction.
• RecBench — pairwise recommendation evaluation.
• URS Bench — multi-intent user-query benchmark.

The extracted user profiles used throughout are released as a dataset (see Data); place the CSVs under persistent_data/user_profile/ or point the relevant configs at them.

Usage

The project uses Hydra. Run any task with:

python main.py task=<CATEGORY>/<TASK> model=<MODEL> [overrides...]

Examples:

# 1. Extract open-ended profiles from social-media posts
python main.py task=persona/social_media_twitter_persona model=gptoss120b
python main.py task=persona/social_media_twitter_dedupe   model=gptoss120b

# 2. Synthesize profile-conditioned SFT data
python main.py task=data/recommendation_recbench_interest model=gptoss120b
python main.py task=data/urs_personalised_data            model=gptoss120b

# 3. Fine-tune a model on the synthetic data
python main.py task=finetuning/sft_v2 model=qwen3_8b

# 4. Evaluate
python main.py task=evaluate/urs_persona            model=qwen3_8b
python main.py task=evaluate/twitter_persona_predict model=qwen3_8b

# 5. Test-time multi-perspective reasoning
python main.py task=test_time/test_time_urs_related_persona model=qwen3_8b
python main.py task=test_time/test_time_urs_summary         model=gptoss120b

Override any config value on the command line, e.g. python main.py task=finetuning/sft_v2 model=qwen3_8b training.learning_rate=2e-5.

Outputs default to outputs/<package>/<task>/<model>/. Experiment tracking via Weights & Biases is configured in conf/config.yaml.

Data

The anonymized user profiles are released as a companion dataset:

• open_ended_profiles.csv — 824 behaviorally grounded profiles extracted from real Bluesky histories.
• synthetic_baseline_profiles.csv — 842 purely synthetic baseline profiles.

Both are pseudonymized and scrubbed of handles, contact details, and URLs. See the dataset card for schema, provenance, and ethical considerations.

License

Released under the Apache License 2.0 (see LICENSE), consistent with the source Bluesky corpus.

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}
}