Vela: User Behavior as Supervision Sparks LLM Intelligence

AI summary

Overview

Research area: Natural language processing — LLM post-training, user modeling, and personalization. Technical level: Intermediate (assumes familiarity with supervised fine-tuning, reinforcement learning, and benchmark-based evaluation). Scope: The paper proposes "user behavior as supervision," a training paradigm that turns anonymized real-world platform behavior histories into supervised tasks, and introduces Vela, a framework that builds roughly 75,000 such tasks from 100,000 users and fine-tunes Qwen3.6-35B-A3B on them.

What This Paper Is About

LLMs are trained mainly on static text, code, and synthetic data, none of which captures how an individual person's preferences and actions evolve over weeks or years. The paper argues this leaves a "user dimension" of intelligence underdeveloped: the ability to infer a specific person's preferences, intent, and likely next actions from their own behavior history. The goal is to show that a user's actual recorded actions can serve as the supervision signal, and that training on it improves both user-related and general LLM capability.

Key Contributions

1. A paradigm and formal task definition. The authors define user behavior as supervision, where each user's chronologically ordered event history is split at a cutoff into a visible history and a hidden future; a task bundles several questions about that user, each answer scored in [0, 1], and the task score is the mean over its questions.

2. The Vela framework and its task corpus. Vela converts anonymized records from 100,000 users on a social media platform into 100,000 tasks drawn from 91 question templates. Templates span six behavior categories (Views, Searches, Purchases, Posts, Follows, Dislikes) and four question types in two families: Predictive (Objective, Open-prediction) scored against the hidden future, and Rubric (Open, Attribution) scored against task-specific criteria. Difficulty-based filtering and a trajectory acceptance rule jointly retain approximately 75,000 tasks, forming two corpora: Vela-Instruct and Vela-RFT (9,375 samples).

3. Two fine-tuned models and a three-part evaluation suite. Fine-tuning Qwen3.6-35B-A3B on Vela-Instruct and Vela-RFT produces Vela-35B-A3B and Vela-35B-A3B-RFT. Evaluation uses VelaBench (982 held-out tasks from the same 91 templates, with held-out users excluded from training), four user-related out-of-distribution benchmarks (PurchaseIntent, ProfileSummary, PersonaBench, AdRecommendation), and five general benchmarks (AIME 2026, GPQA-Diamond, IFBench, CL-bench, CL-bench Life).

4. Evidence that the gains are not confined to user tasks. The paper reports data-scaling behavior from 4.7K to 75K examples, a reinforcement-learning comparison, and a difficulty-filtering ablation, showing improvements on general benchmarks as well as user-related ones.

Main Findings

• In-domain gain on VelaBench: Vela-35B-A3B scores 63.0 versus 51.5 for the base model, a gain of 11.5 points.
• User-related OOD gain: the average across the four user-related OOD benchmarks rises from 41.3 to 47.5 (+6.2). Per-benchmark: AdRecommendation 51.1 to 62.3 (+11.2), ProfileSummary 20.4 to 30.5 (+10.1), PersonaBench 50.0 to 53.2 (+3.2), PurchaseIntent 43.8 to 44.0 (+0.2).
• Small model versus large baselines: with 35B total parameters, Vela-35B-A3B outperforms DeepSeek-V4-Flash (284B, 58.6) and Qwen3.8-2.4T-A95B (2.4T, 62.3) on VelaBench. It also exceeds Qwen3.5-397B-A17B on CL-bench (29.1 vs 27.2) and CL-bench Life (12.8 vs 11.1). Kimi-K3 scores 64.0 on VelaBench, above Vela-35B-A3B.
• General capability improves: the general OOD average rises from 54.2 to 58.2 (+4.0), with gains on all five benchmarks — IFBench 62.5 to 69.2 (+6.7), CL-bench Life 7.2 to 12.8 (+5.6), GPQA-Diamond 82.8 to 87.4 (+4.6), CL-bench 26.7 to 29.1 (+2.4), AIME 2026 91.8 to 92.6 (+0.8).
• Self-generated trajectories still help: Vela-35B-A3B-RFT, trained on trajectories generated by Qwen3.6-35B-A3B itself, gains 2.0 points on VelaBench and 1.6 on the user-related OOD average while holding the general OOD average at 54.2, matching the base model.
• Scaling user behavior data helps: across 4.7K, 9.4K, 18.8K, 37.5K, and 75K Vela-Instruct examples, VelaBench rises from 51.5 at the base to 59.2 at 4.7K and 63.0 at 75K; the user-related OOD average goes from 43.5 at 4.7K to 47.5 at 75K; the general OOD average from 55.8 at 4.7K to 58.2 at 75K, with all tested scales above the base model's 54.2.
• A stronger starting point for reinforcement learning: with GSPO under the same configuration, RL initialized from an 18.8K Vela-35B-A3B checkpoint reaches 66.2 on VelaBench versus 58.5 when initialized from the base model, and 37.6 versus 29.9 on ProfileSummary — 7.7 points higher on both. RL from the base model stays below the SFT checkpoint before RL.
• Difficulty filtering contributes mainly out of domain: at the 37.5K training scale, filtering leaves VelaBench unchanged at 62.4 but improves all four user-related OOD benchmarks, with an average gain of 1.1 points (PurchaseIntent 44.6 to 45.4, ProfileSummary 29.0 to 30.5, PersonaBench 51.5 to 52.7, AdRecommendation 61.0 to 61.9).

Methodology in Plain English

The researchers start from anonymized behavior logs of 100,000 users on a social media platform, with each history spanning several weeks to several years. Profile information and direct identifiers are removed, and records containing personally identifiable information are excluded. Every event is filed under one of six categories: viewing, searching, purchasing, posting, following, or hiding a post as negative feedback.

Each history is cut into segments of 72K to 120K tokens (measured with the Qwen3.6-35B-A3B tokenizer), keeping individual records intact. Roughly the first two-thirds of each segment becomes the visible history, and the rest becomes the hidden future. Templates whose required behavior categories actually appear in the segment are instantiated: some questions give the model a candidate list and ask it to pick what the user will do next; some ask for free-form predictions; some ask for summaries of interests, style, or attributes judged by a rubric; and some show an observed future action and ask the model to explain it from earlier evidence. For any given future event, the system randomly chooses either a Predictive or an Attribution template, because including both would leak the prediction target.

Predictive answers come from the hidden future through deterministic rules, with distractors sampled from other users' pools; Open and Attribution questions are written by a composer model together with their rubrics. Scoring uses set-F1 for the two predictive types and an equal-weighted mean over rubric criteria for the other two. The composer and scorer is Qwen3.8-2.4T-A95B at the xhigh thinking level.

To keep training focused, the base model generates four responses per task with thinking disabled, and only tasks whose mean score falls in [0.2, 0.8] are kept. For each kept task, four new trajectories are generated and only the best is eligible, and only if it beats the earlier filtering score. Vela-Instruct trajectories come from Qwen3.8-2.4T-A95B, while Vela-RFT trajectories come from Qwen3.6-35B-A3B with thinking enabled.

Both models are trained with the Swift framework at a packed sequence length of 131,072 tokens, two epochs, and a learning rate of 1×10⁻⁵. Semantic and rubric-based evaluation is done by the scorer model, while PurchaseIntent, PersonaBench, AIME 2026, GPQA-Diamond, and IFBench use rule-based evaluation.

Why This Matters

Research impact. The paper reframes user activity logs — already abundant on platforms — as a supervision source that sits alongside static text and synthetic data, at a time when public human-generated text is described as a potential bottleneck for LLM scaling. It also reports that training on user behavior improves general benchmarks, not just user-facing ones, which challenges the assumption that personalization training trades off against general capability.

Real-world applications:

• Personal assistants that act on a user's long-term behavior history, such as the OpenClaw and WorkBuddy systems cited in the introduction.
• Recommendation and advertising systems that must predict what a specific user will do, corresponding to the PurchaseIntent and AdRecommendation benchmarks.
• User profile and persona summarization for personalization and audience understanding.
• Search-query generation for ad retrieval conditioned on user profile and recent queries.

Industry relevance. Behavior logs are produced continuously by content platforms and assistant products, so the training signal is regenerable rather than a fixed corpus. A 35B model matching or beating 284B and 2.4T baselines on VelaBench points to a favorable cost profile for deployment, and the positive RL-initialization result suggests the approach fits into existing post-training pipelines rather than replacing them.

Future Directions

• Extending the paradigm to more real-world tasks and scenarios beyond the single social media platform used here, as the conclusion proposes.
• Pushing the data scale further: VelaBench performance was still rising between 37.5K and 75K examples, and the authors flag potential benefits from additional scaling.
• Understanding why RL from the base model stays below the SFT checkpoint before RL, and characterizing when behavior-grounded supervision best prepares a model for reinforcement learning.
• Broader behavior modalities and sources — the related work cites benchmarks such as ATM-Bench that span multiple modalities — plus continued work on privacy-preserving construction from behavioral records.

Target Audience

Researchers and engineers working on LLM post-training, personalization, and user modeling; recommender and advertising teams with access to large behavior logs; and readers interested in alternatives to static-text and synthetic-data supervision. The paper is most useful to those already comfortable with SFT, RL, and multi-benchmark evaluation, since the framework details and comparisons assume that background.

Authors’ abstract

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.