·
Yiwen Liu|6 questions
Generated by a 20-agent research pipeline that gathers intelligence from GitHub, arXiv, ORCID, HuggingFace, and blog posts, then synthesizes domain-specific interview questions grounded in the person's actual work.
asLLR jointly optimises a CTR loss and a QA loss inside a decoder-only LLM. How is the QA supervision signal constructed from raw CRM interaction logs — what plays the role of 'question' and what plays the role of 'answer'?
MoFE-Time uses a mixture of frequency-domain experts on top of an LLM backbone. How does the expert-routing gate decide between time-domain and frequency-domain experts — is it learned per-token, per-window, or per-series?
asLLR reports +0.0231 AUC over traditional CTR baselines offline and +9.5% sales volume online. Which baselines did it beat in the A/B test, and how do you interpret the gap between offline AUC and online revenue impact?
Decoder-only LLMs are expensive to serve at CRM-event latency. What inference strategy does asLLR use in production at Li Auto — full-LLM scoring, distillation, cached embeddings, or something else?
Your three 2025 papers (asLLR, MoFE-Time, Balancing Rewards) share an overlapping author group but cover very different problems — ranking, forecasting, summarization. Is this one applied-ML team rotating through Li Auto's product needs, or are you collaborating across separate sub-teams?
asLLR was trained and evaluated on a single new-energy-vehicle brand's CRM data. What would have to be true for the approach to transfer to other industries with similar lead-scoring problems (real estate, B2B SaaS, insurance)?