HotComment - User Behavior Simulation (Video): leaderboard
HotComment (2026) asks models to write comments on Bilibili videos that would become popular. The score is the engagement that a Qwen3-14B agent simulation of the platform audience estimates for the generated comment, higher is better.
Metric: User Behavior Simulation score: the engagement a Qwen3-14B agent simulation of the platform audience (interested and casual users by demographic priors) estimates, for comments generated zero-shot for held-out HotComment test items (a time- and topic-stratified 10% split) on Bilibili videos; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral-Small-3.1-24B-Instruct [8bit] | 76.08 | |
| 2 | Gemini2.5-Image (HotComment checkpoint unspecified) | 73.09 | |
| 3 | GPT-4o | 73.01 | #751 |
| 4 | GPT-4o Mini | 71.07 | #1667 |
| 5 | Qwen3-VL-8B-Instruct [8bit] | 70.58 | |
| 6 | HotComment Llama-3.2-Vision-Instruct (size unspecified) [8bit] | 65.08 | |
| 7 | Qwen3-VL-4B-Instruct [8bit] | 64.22 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=hotcomment-user-behavior-simulation-video · How It Works · Data refreshed daily, snapshot 2026-11-21.