·
Machine Learning Researcher |Li Auto (理想汽车)
Applied LLM researcher at Li Auto. First author of MoFE-Time (arXiv:2507.06502), a mixture-of-frequency-domain-experts architecture for time-series forecasting, and co-author of asLLR (arXiv:2510.21713) — a decoder-only LLM combining CTR and QA losses to rank automotive sales leads, reporting a +9.5% sales lift in production A/B testing.
Biography
Yiwen Liu is an applied machine-learning researcher at Li Auto (理想汽车), a Chinese new-energy-vehicle manufacturer. Yiwen's public bibliography, verified via Semantic Scholar's author cluster anchored on arXiv:2510.21713, is three 2025 papers — all co-authored with the same Li Auto research group (Yin Sun, Junjie Song, Chenyu Zhang, Siqi Chen, Yuji Cao recur as collaborators): • MoFE-Time (arXiv:2507.06502, July 2025) — Yiwen is first author. A mixture-of-frequency-domain-experts architecture for time-series forecasting that combines time and frequency characteristics in a pretraining–finetuning paradigm. • asLLR (arXiv:2510.21713, September 2025) — Yiwen is second author (first: Yin Sun). A decoder-only LLM that jointly optimises a CTR loss and a Question-Answering loss to rank sales leads from CRM data mixing tabular signals with free-text sales-customer interaction logs. Reports AUC 0.8127 (+0.0231 over traditional CTR baselines) on a 340k-sample new-energy-vehicle dataset and a +9.5% sales-volume lift in production A/B testing. • Balancing Rewards in Text Summarization (arXiv:2510.19325, October 2025) — Yiwen is second author (first: Junjie Song). Applies multi-objective reinforcement learning with HyperVolume Optimization to balance consistency, coherence, relevance and fluency in LLM-based summarization. Biographical details beyond this publication record (education, prior employers, affiliations) are not independently verified and have been intentionally omitted rather than guessed — the name 'Yiwen Liu' is shared across many unrelated researchers in academic indexes.
Co-authored (2nd of 8) decoder-only LLM that jointly optimises CTR loss and QA loss to rank sales leads, fusing CRM tabular features with natural-language sales-customer interaction logs. Reports AUC 0.8127 (+0.0231 over traditional CTR baselines) on a 340k-sample new-energy-vehicle dataset and a +9.5% sales-volume lift in production A/B testing.
First-authored (1st of 9) mixture-of-experts architecture that combines time-domain and frequency-domain expert routing in a pretraining–finetuning paradigm for time-series forecasting. Motivated by the observation that existing LLM-based forecasters rarely model both time and frequency characteristics jointly.
Co-authored (2nd of 7) HyperVolume-based multi-objective RL approach for LLM summarization that balances consistency, coherence, relevance and fluency in a single training signal rather than scalarising the objectives.
Research generated May 14, 2026