Vela Uses Real Behavior Data From 100K Users as Supervision, Up to +11.5 on User Tasks
AllSpark-Research · hf · 2026-10-14
AllSpark Research introduces Vela, a framework built on the idea of "user behavior as supervision": training LLMs' user-dimension intelligence with real user behavior data—inferring preferences, intentions, and future actions from behavior records spanning weeks to years:
- Data pipeline: 100K tasks under 91 real-task templates are generated from the behavior records of 100K anonymized users; after difficulty filtering and trajectory acceptance criteria, roughly 75K tasks with training trajectories are retained
- Results: after fine-tuning Qwen3.6-35B-A3B, user-related tasks improve by up to +11.5 (63.0 on VelaBench, surpassing DeepSeek-V4-Flash and others), with an average +4.0 across 5 general benchmarks; more behavior data yields continued gains
- RL finding: SFT followed by reinforcement learning beats direct RL from the base model by up to 7.7 percentage points
Conclusion: real-world user behavior is a scalable source of supervision that improves both user-centric capabilities and broader LLM intelligence.
More from Research
- Chollet: Ideal AI-for-Science Outcome Is Millions of AI-Aided Scientists, Not a Few Rich Labs — fchollet · 2026-10-14
- Feeling a dipole's electric field along an ellipse: 2.08kq at the end, 0.13kq on top — TinfoilTricorn · 2026-10-14
- Alibaba's Representation Guidance: plain pixel-diffusion Transformers beat bigger models — bdsqlsz · 2026-10-14
- Visualizing the Fourier series of a sawtooth wave, with Gibbs overshoot hitting ~9% — TinfoilTricorn · 2026-10-14
- Gemma Embedding 2 hits SOTA on Vietnamese speech after 3,000 hours of audio, 2 hours on 8xH100 — stablequan · 2026-10-14
- KILVO: robust multimodal odometry for humanoid robots fusing encoders, IMU, LiDAR and vision — rsasaki0109 · 2026-10-14