PaperlayerView exact revision on arXiv

Source-linked AI summary

Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning

Yuxuan Li, Victor Zhong, Ehsan Kamalloo

arXiv:2609.00014v1cs.CLcs.AI

TL;DR

Existing synthetic personas can flatten user variation and miss behavioral signals relevant to personalization. The paper extracts open-ended profiles from authentic social activity and evaluates them through SFT and test-time multi-perspective reasoning, finding improved personalization and structural diversity closer to human-annotated data. The authors position these profiles as a high-fidelity intermediate representation while noting temporal and platform limitations.

  • Problem

    Synthetic categorical personas can flatten individual variation, rely on stereotypes, and miss nuanced behavioral signals that drive human preferences.

  • Method

    The framework extracts open-ended profiles from authentic social media activity and applies them to SFT personalization and test-time multi-perspective reasoning.

  • Results

    Across train-time SFT and test-time multi-perspective reasoning, behavior-grounded profiles improve personalization and show structural diversity closer to human-annotated survey data.

  • Takeaways & Limitations

    Behavior-grounded profiles provide a high-fidelity intermediate representation for personalized language systems.

  • Takeaways & Limitations

    Profiles are timeagnostic and derived solely from Bluesky, limiting temporal coverage and potentially over-representing one platform’s user group.

Abstract

from arXiv · show

Persona-driven techniques increasingly adapt large language models (LLMs) to diverse contexts. However, existing methods predominantly rely on rigid, synthetic personas that flatten individual variation, rely on stereotypes, and miss the nuanced signals driving actual human preferences. We introduce profile behavioral grounding, a framework for extracting open-ended, high-fidelity user profiles directly from authentic, anonymized social media posts. We evaluate these profiles across two paradigms: train-time personalization via supervised finetuning (SFT) and non-parametric test-time multi-perspective reasoning. Across complex recommendation and open-ended query benchmarks, behaviorally grounded profiles consistently improve base models and outperform synthetic profile baselines, driving stronger parametric alignment and enabling 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 systems. Our code base is available at https://github.com/ServiceNow/behavior-grounding.

1 Introduction

The paper argues that rigid synthetic personas miss the behavioral nuance needed for personalization and introduces open-ended profiles grounded in authentic activity. It evaluates these profiles for train-time SFT and test-time multi-perspective reasoning, finding broader, more effective personalization.

  • Motivation: Synthetic personas flatten individual variation into categorical attributes, rely on stereotypes, and miss idiosyncratic behavioral signals behind preferences.The paper contrasts these limitations with messy, distal traces of authentic behavior.
  • Behavioral grounding: Behavioral grounding extracts coherent, open-ended profiles from authentic social media activity rather than hallucinated synthetic personas.The framework cleans and summarizes observed behavioral sequences into reusable textual representations.
  • Evaluation paradigms: The profiles support both profile-conditioned SFT for downstream personalization and test-time multi-perspective reasoning without weight updates.The two paradigms target parametric alignment and multifaceted responses, respectively.
  • Findings: Across recommendation and open-ended query benchmarks, behavioral grounding improves downstream performance in both paradigms.The reported gains span recommendation domains, open-ended queries, and base models.
  • Findings: The analysis attributes these gains to broader profile diversity and lower concentration than synthetic profile baselines.These structural differences are associated with richer and more nuanced model outputs.

2 Related Work

Prior personalization work uses synthetic personas, user histories, and parameter or memory adaptations, while test-time methods use arbitrary roles or model variation for diversity. This paper instead distills distal social traces into reusable profiles and uses them as behaviorally grounded perspectives.

  • Synthetic personas: Synthetic personas support data augmentation and cultural diversity, but studies caution that they can introduce artifacts, stereotypes, and user misrepresentation.The related work contrasts these risks with behavioral grounding from authentic activity.
  • User-context adaptation: Personalization systems adapt to explicit attributes, interfaces, conversations, or interaction histories through retrieval, memory, or parameter-efficient methods.These approaches primarily optimize adaptation to known users and in-domain histories.
  • Positioning: This work tests whether distal social traces can become reusable open-ended profiles while holding downstream adaptation mechanisms fixed.That design isolates the contribution of the user representation itself.
  • Test-time diversity: Multi-persona prompting and Mixture-of-Agents methods broaden reasoning through arbitrary roles, model variation, or temperature scaling.The paper uses behaviorally grounded profiles as distinct experts for perspective sampling instead.

3 Methodology: The Open-Ended Profiling Framework

The methodology extracts and cleans open-ended profiles from unstructured social activity, then uses them for SFT personalization and test-time perspective aggregation. The two pathways differ in whether profiles update model behavior parametrically or guide inference without weight updates.

  • Framework: The framework extracts rich behavioral signals from social media for parametric personalization and non-parametric multi-perspective reasoning.These are presented as two distinct capabilities in the overall pipeline.
  • Profile synthesis: The profiling pipeline processes user posts, removes redundant or contradictory attributes, and summarizes the remaining descriptors into a cohesive profile.Figure 2 depicts this extraction, cleaning, and summarization sequence.
  • Profile synthesis: The study uses a Bluesky dataset containing 2 million posts collected from 2023 to 2024 to build the profile bank.The authors note that extraction has higher initial cost but yields reusable profiles.
  • Personalized alignment via SFT: For SFT, a teacher LLM estimates likely responses from profile, history, and item inputs, then converts greedy-decoded targets into instruction-tuning pairs.The formulation represents the target with a response sequence Y conditioned on U, H, and I.
  • Multi-perspective reasoning: Test-time reasoning samples relevant profiles, generates viewpoint-specific responses, and aggregates them into one comprehensive answer without parametric updates.This setting seeks broad coverage across plausible viewpoints rather than prediction of the inquirer's own preferences.

4 Experiments and Analysis

The experiments test behaviorally grounded profiles in train-time SFT and test-time multi-perspective reasoning across recommendation and open-ended query benchmarks. Open-ended profiles generally outperform synthetic and no-profile baselines, while analyses link gains to profile diversity and identify boundary cases and failure modes.

  • Experimental design: The evaluation asks whether behavior-derived profiles improve SFT personalization, outperform task-relevant synthetic profiles, and support test-time perspective scaling.It uses RecBench recommendation tasks and URS Bench queries evaluated with F1 and LLM-judge scores aligned with human ratings.
  • Train-time personalization: Open-Ended improves over Synthetic in 13 of 15 RecBench cells and is best on URS average for all five evaluated models.Examples include Gemma3 4B Books F1 increasing from 0.309 to 0.643 and Qwen3 14B from 0.308 to 0.632.
  • Limitations and failure modes: Performance is not uniform: categorical compression can lose information, niche-domain sparsity can cause distribution shift, and personalization can introduce irrelevant framing or safety refusals.The reported boundary cases include PRISM preference prediction, RecBench Lastfm, direct factual queries, and multi-perspective aggregation.
  • Test-time multi-perspective reasoning: Open-Ended generally outperforms No Profile and Synthetic in test-time reasoning, yielding up to +0.50 for Olmo3 7B and +0.52 for Qwen3 32B.The trend holds across model families and sizes, while Synthetic improves over No Profile by less.
  • Test-time multi-perspective reasoning: High-temperature sampling performs worse than profile-grounded reasoning, indicating that grounded behavioral perspectives provide more useful diversity than decoding noise alone.Random profiles also outperform Synthetic but remain below related Open-Ended profiles except for Qwen3 8B.
  • Behavioral grounding analysis: Open-ended profiles are more diverse and representative than synthetic profiles, with broader distributions and higher categorical entropy.PCA and distribution analyses support lower concentration and broader profile coverage as explanations for richer outputs.

5 Conclusion

The paper introduces behavioral grounding to extract open-ended profiles from authentic social media activity. These profiles improve personalization across train-time SFT and test-time multi-perspective reasoning, while retaining challenges including persona leakage, over-personalization, and domain sparsity.

  • Behavioral grounding extracts open-ended user profiles from authentic social media activity rather than synthetic personas.
  • Across train-time SFT and test-time multi-perspective reasoning, behavior-grounded profiles improve personalization and exhibit structural diversity closer to human-annotated survey data.
  • Persona leakage, over-personalization, and domain sparsity remain challenges for behavior-grounded personalized language systems.

Limitations

The study identifies two limitations: profiles are time-agnostic and rely solely on Bluesky data. These choices may omit changing preferences and introduce platform-specific representation bias.

  • Time-agnostic profile extraction does not account for changing user preferences and may miss temporal dynamics important for continual real-world deployment.
  • Using only Bluesky data may over-represent that platform’s user group and introduce additional bias in downstream tasks.
  • Future work will study temporally evolving profiles and extend analysis to more diverse data sources.

Ethical Considerations

The study uses public Bluesky data under the platform’s Terms of Service and API guidelines, while aggregating behavioral information to protect privacy. Final profiles retain neither direct quotes nor personally identifiable information.

  • The study collected 2 million public Bluesky posts while adhering to the platform’s Terms of Service and API use guidelines.
  • The extraction pipeline aggregates granular spatiotemporal details into broad behavioral patterns to mitigate mosaic-effect re-identification.
  • Final user profiles retain no direct quotes or Personally Identifiable Information.

A PCA Analysis Plots on User Profiles and SFT data

PCA analyses show that open-ended user profiles cover more areas than the more homogeneous synthetic baseline. Profile-conditioned SFT data are also generally more diverse while remaining distinguishable from ground-truth data.

  • User Profiles: Open-ended user profiles cover significantly more PCA areas, whereas the synthetic baseline is more homogeneous.
  • SFT Data: In almost all domains except Lastfm, open-ended profile-conditioned SFT data show greater diversity while remaining distinguishable from ground-truth data.
  • SFT Data: The SFT-data diversity pattern is consistent with the reported MAUVE scores.

B Experimental Setup

The experiments use Bluesky-derived user profiles, supervised finetuning, and test-time perspective sampling under model- and context-window constraints.

  • Data and profiles: The study uses 826 user profiles extracted from Bluesky under Apache license 2.0.Each profile incorporates up to 100 posts to fit within the LLM context window.
  • Train-time personalization: Finetuning uses full-parameter SFT where tractable, with QLoRA reserved for Qwen3 32B because full tuning exceeded the compute budget.Training runs use one epoch over 90K SFT examples; Qwen3 32B uses rank 64, scaling factor 16, and dropout 0.1.
  • Train-time personalization: The finetuning setup uses a 5e-5 learning rate, 0.1 warmup ratio, and 0.01 weight decay.The effective batch size is 8 through per-device batch size 1 and gradient accumulation 8.
  • Evaluation split: The SFT data setup reserves 92 held-out user profiles for evaluation without overlap with synthesis data.The passage reports 100K examples for SFT training.
  • Test-time reasoning: Test-time experiments sample five perspectives from five user profiles because the aggregator model has limited context-window space.Inference uses each model’s recommended settings, with repetition and frequency penalties to limit over-lengthy responses.

C Negative Results on Lastfm and PRISM

PRISM shows parity between behavior-grounded and synthetic profiles after categorical compression, while Lastfm exposes domain, distribution, and label mismatches that can degrade finetuning.

  • PRISM Preference Prediction: Behavior-grounded profiles do not meaningfully outperform synthetic profiles on PRISM, despite finetuning improving performance over base models.The authors attribute this parity to structural mismatch between open-ended profiles and PRISM’s categorical task format.
  • PRISM Preference Prediction: 87.2% of synthesized SFT preference pairs received identical labels under open-ended and synthetic profiles after profile compression.The additional LLM extraction into discrete categories is reported to cause significant information loss.
  • RecBench Lastfm: Finetuning degrades F1 across both profile types on RecBench Lastfm.The reported failure is linked to three compounding distributional shifts between synthesized SFT data and downstream evaluation data.
  • RecBench Lastfm: Up to 30.8% of extracted profiles contain no music-related keywords, while only 8.31% mention artist names or track titles.This sparsity makes it difficult to synthesize highly relevant music-recommendation trajectories.
  • RecBench Lastfm: Artist overlap is 56.4% in positive Lastfm candidates, compared with 24.2% in open-ended SFT data and 0% for the synthetic baseline.The synthesized popular-artist distribution also diverges from the ground-truth Lastfm distribution.
  • RecBench Lastfm: Balanced synthesized labels produce a higher positive prior than the Lastfm evaluation set, which contains 90% negative examples.The resulting models systematically over-predict “Yes,” heavily penalizing F1.
  • Profile fidelity: The profile extraction pipeline achieves an 88.2% atomic-claim grounding rate, with 93.4% of verified claim–post links passing substring validation.A human audit estimates 87.5% support, consistently indicating faithful profiles.
  • Evaluation validation: The URS Bench LLM judge agrees with human-majority preferences in 90.14% of cases, with Cohen’s κ = 0.8014.An independent Qwen3-235B judge reproduces model rankings with Spearman ρ = 0.945 and per-response Pearson r = 0.81.

I Test-Time Multi-Perspective Reasoning: High Temperature Does Not Replace Grounded Perspectives

High-temperature sampling without profiles does not match grounded multi-perspective reasoning: aggregation helps, but remains below profile-grounded scores.

  • Ablation design: The ablation tests whether multi-perspective gains come from grounded profile diversity or ungrounded decoding noise.Candidate responses are sampled at temperature T = 1.1 without profile conditioning on URS Bench.
  • Results: High-temperature sampling alone performs substantially worse than profile-grounded multi-perspective reasoning.Aggregating high-temperature candidates improves over raw samples but remains below the profile-grounded scores.
  • Interpretation: The comparison indicates that grounded behavioral perspectives provide a stronger source of useful diversity than decoding noise alone.The table reports URS Bench average scores for high-temperature sampling without profile conditioning.

J Example Prompts

The appendix illustrates prompts for extracting, cleaning, summarizing, and applying user profiles, alongside examples for synthetic-profile generation and profile-conditioned tasks.

  • J.1 Extracting Behaviorally Grounded User Profiles: The descriptor-extraction prompt asks an LLM to identify useful information about a user’s background and preferences from each post.It returns summarized information or “N/A” with a justification when a post is uninformative or unclear.
  • Prompt examples: Examples classify a photography-related post as “Photography Lover” and a simple reply as uninformative.A public-health post is also treated as useful information about the user.
  • Profile summarization: The profile-summarization prompt removes vague, redundant, or contradictory descriptions before generating a cohesive profile.The examples show extracted attributes being consolidated into a longer user-profile representation.
  • Prompt for Fake User Profile Generation: The fake-profile baseline prompt asks the model to generate a profile from example profiles rather than extracted behavioral descriptions.This appendix prompt explicitly labels the requested output as a fake user profile.
  • J.2 Prompt to Synthesize SFT Data: The SFT-data prompt asks a recommender to predict whether a user would be interested in an item and return YES, NO, or MAYBE with justification.The profile-conditioned QA prompt first checks whether a user could plausibly ask the question before answering it.
  • Prompt for Profile-Grounded Question Answering: The profile-grounded QA examples show both rejection and personalization: a German resident is unlikely to ask about Costco pricing, while a Japan-interested traveler receives Osaka recommendations.The examples use MatchUserProfile labels to decide whether personalization is appropriate.
  • Prompt examples: The appendix also demonstrates a personalized Osaka answer for a romantic trip and a Costco price-adjustment question from a German-speaking user.These examples connect user-profile attributes to the scope and content of the generated answer.
Loading 2609.00014v1…