Source-linked AI summary
Vela: User Behavior as Supervision Sparks LLM Intelligence
Yang Xiao, Yusong Sun, Haoyi Wu, Wenyang Hui, Weiqiang Hang, Luqing Wang, Zejian Wang, Zirui Tu, Jingsheng Mao, Shaobo Hu, Yang Luo, Jiale Guo, Jin Zhang, Yudan Liu, Zhaokai Luo, Ren Huan, Yao Hu, Chengyue Jiang, Wenjie Li, Chuan Mu
TL;DR
Current training sources may not capture how individual users’ behavior evolves, despite the need to infer preferences, intent, and likely future actions from long histories. Vela converts real behavior records into training tasks and trajectories, and the resulting models improve user-related and general benchmark performance.
Problem
Common training sources may not capture how individual users’ behavior evolves, while user-facing models can struggle with real needs, long histories, and personalization.
Source: §1 · p. 2Method
Vela converts anonymized behavior records from 100,000 users into 100,000 tasks across 91 templates, filtering them to retain approximately 75,000 tasks with selected training trajectories.
Source: §1 · p. 2Results
11.5 points on VelaBench, up to 11.2 points on user-related OOD benchmarks, and 4 average points on general OOD benchmarks over the base model.
Source: §1 · p. 2Takeaways & Limitations
The findings support real-world user behavior as a scalable source of supervision for user-centered capabilities and broader LLM intelligence.
Source: §6 · p. 9
Abstract
from arXiv · showhide
Large language models are increasingly deployed in applications that act on users' long-term behavior histories, from personal assistants to recommendation and advertising systems. These applications demand the user dimension of LLM intelligence: inferring a person's preferences, intent, and likely future actions from weeks to years of behavior data. Yet current models are trained on static text or synthetic data, neither of which captures how user behavior evolves over time. Real-world behavior records, already available at scale on content platforms and in personal-assistant agents, offer a natural supervisory signal. We propose user behavior as supervision and introduce Vela, a framework that converts real user behavior data into trainable tasks and trajectories. From the anonymized behavior records of 100,000 users, Vela generates 100,000 tasks across 91 real-world task templates; difficulty-based filtering and trajectory acceptance criteria retain approximately 75,000 tasks with selected training trajectories. Fine-tuning Qwen3.6-35B-A3B on these trajectories improves performance over the base model by up to 11.5 points on user-related tasks and by an average of 4 points on general benchmarks. Training on more user behavior data further improves user-related performance, and reinforcement learning after fine-tuning outperforms reinforcement learning from the base model by up to 7.7 points. These gains suggest that real-world user behavior is a scalable source of supervision for both user-centered capabilities and the broader intelligence of LLMs.
1 Introduction
Current training sources may miss how individual users’ behavior evolves, limiting the user dimension of LLM intelligence. Vela turns observed behavior into tasks and trajectories, with reported gains on user-related and general benchmarks.
1 Introduction
- Web pages, textbooks, code, and synthetic data may not capture how individual behavior evolves, while user-facing models can struggle with long histories and personalization.The paper frames the user dimension as inferring an individual’s preferences, intent, and likely future actions from behavior data.Source: §1 · p. 2
- Prior behavior-supervision work targets specific applications; Vela addresses open questions about task construction and effects on user-related and general capabilities.The paper asks how to convert raw behavior into tasks and trajectories, whether it improves user-related intelligence, and whether it strengthens or harms general capabilities.Source: §1 · p. 2
- 11.5 points on VelaBench, up to 11.2 points on user-related OOD benchmarks, and 4 average points on general OOD benchmarks over the base model.Training on more behavior data further improves user-related performance; reinforcement learning after fine-tuning scores 7.7 points higher on both VelaBench and ProfileSummary than reinforcement learning from the base model.Source: §1 · p. 2
- Vela converts anonymized records from 100,000 users into 100,000 tasks across 91 templates, retaining approximately 75,000 tasks with selected trajectories.The tasks form Vela-Instruct and Vela-RFT training corpora.Source: §1 · p. 2, Figure 2 · p. 3
2 Vela: User Behavior as Supervision
Vela turns anonymized user behavior histories into tasks that combine prediction and rubric-based questions, then filters tasks and training trajectories. Its evaluation setup spans in-domain, user-related out-of-domain, and general out-of-domain benchmarks.
2.1 Real-world user behavior data
- The source corpus contains anonymized histories from 100,000 users, spanning weeks to years and organized into six behavior categories.Direct identifiers and records containing personally identifiable information are removed or excluded before task generation and training.Source: §2.1 · p. 3
2.2 User behavior as supervision
- Each task combines questions about one user, answered from a visible history that ends before the hidden future.The model produces an answer for each question; each answer is scored and the task score is their mean.Source: §2.2 · pp. 3–4
- Vela defines 91 templates across six behavior categories and four question types, with Predictive questions grounded in future behavior and Rubric questions evaluated against task-specific criteria.Objective and Open-prediction are Predictive types; Open and Attribution are Rubric types.Source: §2.2 · p. 4, Figure 3 · p. 4
- Predictive answers are evaluated against hidden-future ground truth, while Rubric answers receive scores on task-specific criteria.Objective predictions use exact candidate matching; Open-prediction responses use set-F1 after semantic matching.Source: §2.2 · p. 4
- Task synthesis segments histories, uses roughly the first two-thirds as visible context, and creates 100,000 candidate tasks while avoiding paired Predictive and Attribution questions about the same future event.Open questions receive history-based rubrics; Attribution questions use an observed future behavior to ask for an explanation grounded in earlier evidence.Source: §2.2 · p. 4
2.3 Difficulty-based filtering and trajectory generation
- Difficulty filtering retains tasks scoring between 0.2 and 0.8, and trajectory acceptance leaves about 75,000 Vela-Instruct tasks with one selected trajectory each.A trajectory must score above the task's mean score during difficulty filtering; Vela-RFT instead contains 9,375 training samples.Source: §2.3 · p. 5
2.4 Evaluation benchmarks
- Evaluation includes 982 held-out VelaBench tasks, four user-related out-of-domain benchmarks, and five general out-of-domain benchmarks.The user-related benchmarks cover purchase intent, profile summarization, persona understanding, and ad recommendation; general benchmarks cover mathematics, science, instruction following, and context learning.Source: §2.4 · pp. 5–6
3 Experiments
Fine-tuning on Vela behavior trajectories improves user-related and general benchmark performance, and larger training sets further improve scores across benchmark groups. The experiments also compare models trained at different data scales and establish the benchmark setup.
3.1 Setup
- The authors fine-tune Qwen3.6-35B-A3B on Vela-Instruct and evaluate it on three benchmark groups; they also train on five random subsets containing 4.7K, 9.4K, 18.8K, 37.5K, and 75K examples.Unless otherwise stated, Vela-35B-A3B refers to the model trained at the 75K scale.Source: §3.1 · pp. 6–7
3.2 User behavior as supervision improves user-related intelligence
- Vela-35B-A3B scores 63.0 on VelaBench, 11.5 points above Qwen3.6-35B-A3B, and gains 6.2 points on average across four user-related OOD benchmarks.Its largest user-related OOD gain is 11.2 points on AdRecommendation.Source: §3.2 · p. 7
3.3 User behavior as supervision benefits general LLM intelligence
- Vela-35B-A3B outperforms Qwen3.6-35B-A3B on all five general OOD benchmarks, with an average gain of 4.0 points.The reported results extend to general capabilities; Vela-35B-A3B-RFT improves user-related performance while maintaining the base model’s average general OOD score.Source: §3.3 · p. 7
3.4 Scaling user behavior data improves LLM intelligence
- From 4.7K to 75K examples, VelaBench rises from 51.5 to 63.0, user-related OOD average from 43.5 to 47.5, and general OOD average from 55.8 to 58.2.VelaBench continues improving from 37.5K to 75K examples; all tested scales outperform the base model on general OOD performance.Source: §3.4 · pp. 7–8, Figure 5 · p. 7
4 Analysis
Starting reinforcement learning from a Vela fine-tuned checkpoint yields higher scores than starting from the base model on both tested benchmarks. Difficulty-based filtering also improves user-related OOD performance while preserving the reported in-domain score.
4.1 User behavior as supervision provides a stronger foundation for downstream tasks
- With the same GSPO configuration, Vela-35B-A3B initialization scores 66.2 versus 58.5 on VelaBench and 37.6 versus 29.9 on ProfileSummary, 7.7 points higher on both.The Vela checkpoint is the 18.8K-example SFT model; RL from the base remains below the SFT checkpoint before RL on both benchmarks.Source: §4.1 · p. 8
4.2 Difficulty-based filtering improves user-related intelligence
- At 37.5K training examples, difficulty filtering raises average performance by 1.1 points across four user-related OOD benchmarks while VelaBench remains 62.4.The comparison holds trajectory generation, acceptance criteria, fine-tuning settings, and final training-set size fixed.Source: §4.2 · p. 8, Table 2 · p. 9
5 Related Work
Prior work evaluates user memory, preference following, evolving profiles, and behavior prediction, and explores user representations, recommendation, and simulation. Vela uses observed future behavior as predictive ground truth across a broad task collection and studies data scale and downstream effects.
Evaluation of user understanding and behavior
- User-focused evaluations cover long conversational memory, explicit and implicit preferences, evolving profiles, memory updates, multimodal personal references, and long-horizon behavior prediction.Examples include LongMemEval [32], PrefEval, PersonaMem [10], Memora, ATM-Bench, OPeRA [31], FineRob, and OmniBehavior [4].Source: §5 · p. 9
User representations, recommendation, and behavior simulation
Vela's supervision source and scope
- Vela grounds tasks and trajectories in observed real user behavior, using observed future behavior as predictive ground truth and evaluating across task breadth, data scales, user-related performance, general capabilities, and subsequent reinforcement learning.The related-work discussion notes risks of losing original-distribution information with model-generated data.Source: §5 · p. 9
6 Conclusion
Vela uses observed user behavior to train LLMs, with experiments showing gains in user-related and general capabilities, further improvements with more data, and stronger downstream reinforcement learning after fine-tuning.
6 Conclusion
- Experiments show gains in both user-related and general LLM capabilities.The paper also reports further improvements with data scaling and a stronger foundation for downstream tasks through reinforcement learning.Source: §6 · p. 9
- User behavior as supervision derives training signals from observed behavior, while Vela converts behavior histories into trainable tasks and trajectories.Source: §6 · p. 9
- The findings establish real-world user behavior as a scalable source of supervision and motivate extending the approach to more tasks and scenarios.Source: §6 · p. 9
Full Author List
This material lists the paper’s authors, describes AI assistance and privacy safeguards, and outlines planned releases and the paper’s methods and examples.
Full Author List
- The full author list names Yang Xiao, Yusong Sun, and the other paper authors.Source: Full Author List · p. 9
AI use statement
- Generative AI assisted code development, literature retrieval, result interpretation, drafting, and revision; authors reviewed outputs and retain responsibility for the paper’s methods, results, and claims.Source: AI use statement · p. 10
Ethics statement
- The study’s behavior-data use was platform-authorized and privacy- and ethics-reviewed, with identifiers removed and records containing personally identifiable information excluded.The stated purpose of these measures is to reduce privacy risks.Source: Ethics statement · p. 10
Reproducibility statement
- The authors plan to release task templates, generation and scoring prompts, and training and evaluation code.The paper also maps its sections to processing, task construction, filtering, trajectory selection, scoring, fine-tuning, reinforcement learning, and filtering ablations; appendices provide illustrative records and task templates.Source: Reproducibility statement · p. 10
A.1 Record Examples
The examples show a behavior history followed by a hidden future purchase, while representative templates distinguish prediction, objective selection, open analysis, and attribution tasks.
A.1 Record Examples
- The synthetic example records searches, views, posts, follows, and a dislike before a hidden future purchase of a 20 L daypack.The cutoff is Day 5; the purchase appears at Day 7.Source: §A.1 · p. 15
B Representative Task Templates
- The templates cover six behavior categories and four question types, with questions based on visible history and predictive tasks targeting hidden-future behavior.Objective questions provide candidates; open questions analyze visible history; attribution questions explain disclosed future behavior using earlier evidence.Source: B Representative Task Templates · p. 16, Table 3 · p. 16
- Representative templates distinguish following from viewing or liking, and require evidence rather than unsupported explanations of behavior.Examples ask for future follows, explicit dislikes, or explanations of follows and hidden content; attribution responses must identify evidence gaps and uncertainty rather than treat association as proof of causation.Source: B Representative Task Templates · p. 17
D Benchmark Examples
The benchmark examples list evaluation sets and metrics alongside sample purchase, profile, category-prediction, and advertising-rewrite tasks.
D Benchmark Examples
- The listed benchmarks use recall@20 for PurchaseIntent, macro-F1 across subsets for PersonaBench, and LLM-based scores for ProfileSummary and AdRecommendation.The passage lists 500 PurchaseIntent users and 1,137 labels, 230 ProfileSummary examples, 3,690 PersonaBench questions across 32 subsets, and 996 AdRecommendation examples.Source: D Benchmark Examples · p. 21
- The purchase examples ask for likely future products or product categories, with one category example identifying food and fresh groceries as the ground truth.Source: D Benchmark Examples · p. 21
- Other examples request a structured profile with interest summaries and reasoning, or commercially valuable query rewrites for advertising retrieval.The profile example includes rubric references, and the rewrite example identifies food as the category.Source: D Benchmark Examples · p. 21
E RL Configuration
The RL experiments compare Qwen3.6-35B-A3B and Vela-35B-A3B initializations across VelaBench and ProfileSummary using shared training data and configuration. ProfileSummary rewards are scored with task-specific rubrics.
E RL Configuration
- Vela-35B-A3B starts from an SFT checkpoint trained on 18.8K Vela-Instruct examples, and both initializations use identical RL data and configuration within each domain.The experiments use 2,000 VelaBench training tasks or 2,330 ProfileSummary examples; RL training and evaluation have no overlapping samples or users.Source: E RL Configuration · p. 22
- All four experiments use Group Sequence Policy Optimization with 16 responses per prompt and no learned critic.Megatron-LM handles training and SGLang generates rollouts; 32 of 48 GPUs are allocated to policy training and 16 to rollout inference.Source: E RL Configuration · p. 22, Table 4 · p. 22
- For ProfileSummary, task-specific rubrics score generated user profiles, core interests, and inferred interests.Source: E RL Configuration · p. 22