LiveSim from CAS and ByteDance models how livestream shills triple viewer conversion, accepted at EMNLP 2026
量子位 · wechat · 2026-09-10
Researchers from the Chinese Academy of Sciences and ByteDance present LiveSim, an EMNLP 2026 paper introducing dynamic user simulation for livestream ecosystems.
- Core idea: Instead of a static user profile, LiveSim treats personas as behavioral hypotheses continuously corrected by interaction. Its RBHS mechanism probes model predictions against real logs and converts errors into "environment-behavior patches" describing how cues like social proof and host scripts shift interest, trust, desire, and fatigue.
- Results: On real Douyin logs (1,963 users, 14,391 user-livestream pairs), with Doubao 1.8 the behavioral distribution gap (A-JSD) drops 43.45%, risky-behavior F1 rises 13.17%, and trajectory consistency improves 21.14% — versus only 2.35% for next-action hit rate.
- Shill effect: In closed-loop multi-agent simulation, raising shill ratio from 0 to 40% lifts viewer conversion from 20.61% to 56.97% (nearly 3x); the first 10% of shills drives over half the increase, and 62–77% of conversions happen within the first 10 rounds.
- Protection: Context-aware personalized warnings cut conversion from 18.19% to 8.61% (58.68% protection success), showing lightweight state estimates suffice for timely intervention.
- The team also released the LiveRisk benchmark and related KDD/SIGIR 2026 risk-assessment work.
More from Research
- Flag-planting on arXiv was normal for a decade — LLMs just made it worse — RexDouglass · 2026-09-21
- Meta researcher proposes reviewer score flagging papers as "a waste of everyone's time" — burny_tech · 2026-09-21
- Jev and the bittersweet lesson: can differently sized models serve different apps? — amankhan · 2026-09-21
- OceanMoE: Conditional Sparse MoE Cuts Error in Long-Horizon Ocean Forecasting — burny_tech · 2026-09-21
- Fudan team claims first observation of a quantum state in mitochondria of living cells — coherence · 2026-09-21
- Ethan Mollick: social science urgently needs fast, AI-informed, forward-looking research — robseamans · 2026-09-21