Community-Aware Assessment of Social Textual Engagement and Resonance: A Human-Centric Perspective on User-Generated Content Evaluation
Paper โข 2606.01897 โข Published
MEDEA: Multimodal Engagement-Driven Evaluation Architecture
This repository contains the MEDEA, as described in:
Community-Aware Assessment of Social Textual Engagement and Resonance: A Human-Centric Perspective on User-Generated Content Evaluation
Tianjiao Li*, Kai Zhao*, Xiang Li, Yang Liu, Huyang Sun
ACL 2026
MEDEA is a multimodal large language model trained to assess whether User-Generated Content (UGC) achieves positive community resonance. Unlike traditional video quality assessment methods that focus on aesthetic fidelity, MEDEA evaluates content quality through social-cognitive reasoning โ simulating diverse viewer personas and their empathetic reactions before making a quality judgment.
Model Description
- Base model: Qwen3-VL-8B-Thinking
- Training: Supervised Fine-Tuning + Process-Supervised Reinforcement Learning (GRPO with Social Alignment Reward)
- Task: CASTER (Community-Aware Assessment of Social Textual Engagement and Resonance)
- Input: Video frames, cover image, title, tags, category metadata, ASR transcripts
- Output: Social-CoT reasoning paths + binary quality label (High-Quality / Low-Quality)
- Languages: Chinese and English bilingual support
Key Features
- Social Chain-of-Thought (Social-CoT): Instantiates diverse viewer personas to simulate authentic community reactions before aggregating them into a final quality judgment.
- Social Alignment Reward: RL training ensures generated reasoning paths are grounded in genuine human social cognition rather than generic templates.
- Community Perspective-Taking: The model reasons about why content resonates with specific audience segments, providing interpretable and empathetic evaluation traces.
Performance
Results on CASTER-Bench:
| High-Quality F1 | Low-Quality F1 | Macro F1 | |
|---|---|---|---|
| MEDEA | 0.650 | 0.847 | 0.749 |
| Claude-4.5-opus (Long-CoT) | 0.528 | 0.517 | 0.522 |
| Qwen3-VL-Plus (Social-CoT) | 0.508 | 0.647 | 0.578 |
| GPT-5.2 | 0.506 | 0.489 | 0.498 |
Deployment with vLLM
vllm serve your-org/MEDEA \
--served-model-name "MEDEA" \
--max-model-len 20480 \
--tensor-parallel-size 8 \
--gpu-memory-utilization 0.8 \
--dtype float16 \
--port 8006
Inference
Input Format
Prepare a JSON file where each sample has input_text and images fields:
{
"input_text": "<ไปปๅกๅฎไน>\nไฝ ๆฏไธไธชB็ซๅ
ๅฎนๅๆไธๅฎถใ่ฏท็ปผๅ็ปๅฎ็จฟไปถ็่ง่ง็นๅพๅๆๆฌๅ
ๅฎน๏ผ็ถๅๆ นๆฎ่ง้ขๅ
ๅฎนๆจๆต่ฏ่ฎบๅบ้ๅฏน่ฏฅ่ง้ขๅไฝ่ดจ้ๅฏ่ฝๅบ็ฐ็่ฏ่ฎบ๏ผ่ณๅฐ15ๆก๏ผๅๅ
ถๅพๅ๏ผๆๅๆป็ป่ฏ่ฎบๅบ็ๆปๅพๅๅนถ็ดๆฅ่พๅบ็ปๆใ\n\n### ่พๅ
ฅๆฐๆฎ\n1. ๅฐ้ข: ็จฟไปถ็ๅฐ้ขๅพๅ\n2. ๅ
ณ้ฎๅธง: 7ๅผ ่ง้ข็ๅ
ณ้ฎๅธงๅพๅ\n3. ๆ ้ข: {title}\n4. ๆ ็ญพ: {tag}\n5. ASR: {asr}\n6. ไธ็บงๅๅบ: {new_tid_name}\n7. ไบ็บงๅๅบ: {new_sub_tid_name}\n8. ๆถ้ฟ: {duration}\n9. ๅ่พจ็: {resolution}\n10. ๆฏๅฆ็ซๅฑ: {vertical}\n\n### ่ฏ่ฎบๅบๆปๅพๅๅคๆญๅๅ\n1. ๆจกๆ็่ฏ่ฎบๅฟ
้กป่ณๅฐๆ15ๆก๏ผไธ่ฝ้ๅค๏ผๅนถไธๆ็กฎๅบ็ฐๅจๆ่่ฟ็จไธญ\n2. ๅ่ฎพๅจๆจกๆ็่ฏ่ฎบไธญ๏ผๆฏๆ็ฑป่ฏ่ฎบๅ
ฑXๆก๏ผๅๅฏน็ฑป่ฏ่ฎบๅ
ฑYๆก\n3. ่ฎก็ฎ Sigma-ๆ ๅๅๅทฎ้ข๏ผSkellam z ๅๆฐ๏ผz = (X - Y) / sqrt(X + Y)\n4. ๅฆๆ z โฅ 1.5 ๅ่ฏ่ฎบๅบๆปๅพๅไธบ\"ๆฏๆ\"๏ผๅฆๅไธบ\"้ๆ็กฎๆฏๆ\"\n5. ๅจ่พๅบๆถ๏ผzๅผๅฟ
้กป็จboxed{}ๅ
่ฃน๏ผไพๅฆ\"z = boxed{-2}\"\n6. ๆป็ปๆถ็ๆฏๆ็ฑป่ฏ่ฎบๆฐๅๅๅฏน็ฑป่ฏ่ฎบๆฐ้่ฆๅๆ่่ฟ็จไธญๆจกๆ็่ฏ่ฎบๆฐไธฅๆ ผๅน้
๏ผไธ่ฝ้ๆ่ๆ\n\n<ๅฝๅไปปๅก>\nๅฐ้ข: <image>\nๅ
ณ้ฎๅธง: <image><image><image><image><image><image><image>\nๆ ้ข: \"็ปไธ่งๆดฒไธ็น้ๆผ\"\nๆ ็ญพ: \"ๆธธๆ,ๅฐๅปๆธธๆ,ไธ่งๆดฒ่กๅจ\"\nASR: \"\"\nไธ็บงๅๅบ: \"ๆธธๆ\"\nไบ็บงๅๅบ: \"ๅฐๅปๆธธๆ\"\nๆถ้ฟ: \"28็ง\"\nๅ่พจ็: \"2160x3840\"\nๆฏๅฆ็ซๅฑ: \"ๆฏ\"\n\n่ฏท็ดๆฅ่พๅบ็ปๆ๏ผๆฏๆ/้ๆ็กฎๆฏๆ๏ผ๏ผไธ่ฆๅ
ๅซไปปไฝ้ขๅค่งฃ้ใ",
"images": [
"/path/to/cover.jpg",
"/path/to/frame_0.jpg",
"/path/to/frame_1.jpg",
"/path/to/frame_2.jpg",
"/path/to/frame_3.jpg",
"/path/to/frame_4.jpg",
"/path/to/frame_5.jpg",
"/path/to/frame_6.jpg"
]
}
Each <image> placeholder corresponds to an image path in the images list (1 cover + 7 key frames). The {title}, {tag}, etc. in the task definition are already filled in the <ๅฝๅไปปๅก> / <current_task> section.
MEDEA supports both Chinese and English prompts. Below is an English prompt example:
{
"input_text": "<task_definition>\nYou are a Bilibili content analysis expert. Please integrate the visual features and text content of a given video, then infer at least 15 possible comments and their sentiment regarding the quality of the video creation that might appear in the comment section. Finally, summarize the overall sentiment of the comment section and output the result directly.\n\n### Input Data\n1. Cover: Cover image of the video\n2. Keyframes: 7 keyframe images from the video\n3. Title: {title}\n4. Tags: {tag}\n5. ASR: {asr}\n6. Primary Category: {new_tid_name}\n7. Secondary Category: {new_sub_tid_name}\n8. Duration: {duration}\n9. Resolution: {resolution}\n10. Is Vertical: {vertical}\n\n### Criteria for Overall Comment Section Sentiment\n1. There must be at least 15 simulated comments, they must not be repetitive, and they must be explicitly shown in the thinking process.\n2. Assume that among the simulated comments, there are X supportive comments and Y opposing comments.\n3. Calculate the Skellam z-score: z = (X - Y) / sqrt(X + Y)\n4. If z โฅ 1.5, the overall comment section sentiment is \"supportive\"; otherwise, it is \"not clearly supportive\".\n5. When outputting, the z-value must be enclosed in boxed{} brackets, e.g., \"z = boxed{-2}\".\n6. The number of supportive and opposing comments in the summary must strictly match the number of simulated comments in the thinking process and not be fabricated.\n\n<current_task>\nCover: <image>\nKeyframes: <image><image><image><image><image><image><image>\nTitle: \"Epic moment in Delta Force\"\nTags: \"gaming,shooter,Delta Force\"\nASR: \"\"\nPrimary Category: \"Gaming\"\nSecondary Category: \"Shooter\"\nDuration: \"28 seconds\"\nResolution: \"2160x3840\"\nIs Vertical: \"Yes\"\n\nPlease output the result directly (supportive/not clearly supportive), without any additional explanation.",
"images": [
"/path/to/cover.jpg",
"/path/to/frame_0.jpg",
"/path/to/frame_1.jpg",
"/path/to/frame_2.jpg",
"/path/to/frame_3.jpg",
"/path/to/frame_4.jpg",
"/path/to/frame_5.jpg",
"/path/to/frame_6.jpg"
]
}
Batch Inference Script
import base64
import json
import mimetypes
from concurrent.futures import ThreadPoolExecutor, as_completed
from openai import OpenAI
from tqdm import tqdm
client = OpenAI(base_url="http://localhost:8006/v1", api_key="EMPTY")
def file_to_data_url(path):
mime, _ = mimetypes.guess_type(path)
mime = mime or "image/jpeg"
with open(path, "rb") as f:
b64 = base64.b64encode(f.read()).decode()
return f"data:{mime};base64,{b64}"
def build_multimodal_content(text, image_paths):
parts = text.split("<image>")
content = []
for i, seg in enumerate(parts):
if seg.strip():
content.append({"type": "text", "text": seg.strip()})
if i < len(image_paths):
content.append({"type": "image_url", "image_url": {"url": file_to_data_url(image_paths[i])}})
return content
def infer(sample):
user_msg = next(m for m in sample["messages"] if m["role"] == "user")
content = build_multimodal_content(user_msg["content"], sample.get("images", []))
resp = client.chat.completions.create(
model="MEDEA",
messages=[{"role": "user", "content": content}],
temperature=0.6,
max_tokens=4096,
top_p=0.95,
)
return resp.choices[0].message.content.strip()
# Batch inference
samples = json.load(open("test_data.json"))
with ThreadPoolExecutor(max_workers=32) as ex:
futures = {ex.submit(infer, s): i for i, s in enumerate(samples)}
for fut in tqdm(as_completed(futures), total=len(samples)):
i = futures[fut]
samples[i]["pred"] = fut.result()
json.dump(samples, open("results.json", "w"), ensure_ascii=False, indent=2)
Output Format
The model generates structured Social-CoT reasoning with simulated community reactions, followed by a final quality label:
Click to expand full example output
<think>
ๅบไบ็ปๅฎ็จฟไปถไฟกๆฏ๏ผ่ฏฅ่ง้ขๅฑไบใๆธธๆใๅๅบไธ็ใๅฐๅปๆธธๆใ็ฑปๅซ๏ผๆ ้ขใ็ปไธ่งๆดฒไธ็น้ๆผ๏ผ"็"๏ผๅคง็บขๅ
๏ผใๆ็กฎๆๅบ่ง้ขไธป้ขๆฏๅ
ณไบๆธธๆๅ
่ทๅพๅคง้็จๆ็ฉๅใ่ง้ขๆถ้ฟใ28็งใ๏ผๅ่พจ็ใ2160x3840ใ๏ผไธไธบใ็ซๅฑใๆๆใASRๅ
ๅฎนๆชๆไพใๅ
ณ้ฎๅธงๅๅฐ้ขๆธ
ๆฐๅฑ็คบไบ่ง้ขๅ
ๅฎน๏ผๅณ้่ฟใๆๆบๆๆๆพ็คบๅฑใ็ๆนๅผ๏ผๅ็ฐไบใไธ่งๆดฒ่กๅจใๆธธๆไธญไธไธชใ่ๅ
ใ่ขซๆๅฎๅ๏ผ้้ข่ฃ
ๆปกไบใไธปๆๅฆๅ
ๆจกๅใๅ็จไฟกๆฏ็ป็ซฏใๆพๅกใๅป็ๆบๅจไบบใๆตฎๅ่ฎพๅคใ้ป้็ผ็ฒใๅฎ้ชๆชๆขฐใๅ่ดตๆ่กจใ็็ผ ็ญใ่ฟไบ็จๆ็ใ็บข่ฒๅ่ดจใๆธธๆ็ฉ่ต๏ผไธๆ ้ขไธญ็"ๅคง็บขๅ
"ๆฆๅฟตๅฎ็พๅฅๅใ
่งไผๅจ่ง็่ฟ็ฑปไปฅๆธธๆ็จๆ็ฉๅๅฑ็คบไธบไธป้ข็่ง้ขๆถ๏ผ้ๅธธไผๅ
ณๆณจไปฅไธๅ ไธชๆน้ข๏ผ้ฆๅ
ๆฏๅ
ๅฎนๆฏๅฆ็ๅฎ๏ผๅ
ถๆฌกๆฏๅฑ็คบ็็ฉๅ็จๆๅบฆๆฏๅฆ่ถณๅค้ๆผ๏ผๅ่
ๆฏๅฏนupไธป่ฟๆฐๆๅฎๅ็่ฏไปท๏ผๆๅๅฏ่ฝไนไผๅ
ณๆณจ่ง้ขๆฌ่บซ็ๅถไฝ่ดจ้ใ
ๅฝ่งไผ็ๅฐใ่ง้ขๆ ้ข"็"๏ผๅคง็บขๅ
๏ผไธๅฑๅนไธๅ
ๆปก็บข่ฒ็จๆ็ฉๅ็่ๅ
ใๆถ๏ผไผๅ ๅ
ถ่ง่งๅฒๅปๅไบง็ๅผบ็็่ฏๅฎใ
ใ็็็นๅซ็ปๅ๏ผๆพๅฟๅฒ[ๆๆ็ผ][ๆๆ็ผ][ๆๆ็ผ]ใ(ๆฏๆ่ฏ่ฎบ1)
้ๅฏนใ่ง้ขไธญ่ๅ
่ขซ็จๆ็บข่ฒ็ฉๅๅกซๆปกใ็ๆฏ่ฑก๏ผ่งไผไผๅฐๅ
ถไธๆ ้ขใๅคง็บขๅ
ใ่็ณป่ตทๆฅ๏ผ่กจ่พพๅฏนๅ
ๅฎน็่ฎคๅฏๅ่ตๅนใ
ใ็้ผ๏ผ็็บขๅ
๐งงใ(ๆฏๆ่ฏ่ฎบ2)
ๅฝ่งไผ็ๅฐใๅฆๆญคไนๅค็็จๆ็ฉๅๆฑ้ไธๅ
ใๆถ๏ผไผ็ฑ่กทๅฐ่ตๅนupไธป็ๅฎๅๆ่ฟๆฐใ
ใๆๅป[ๆๆ็ผ]่ฟไนๆๅฎๅใ(ๆฏๆ่ฏ่ฎบ3)
ใ่ง้ขๅ
ๅฎน็ดๆฅๅฑ็ฐไบไธ่ๅ
็บข่ฒ็จๆ็ฉๅ็้ๆผๆๆใ๏ผ่ฟ่ถณไปฅ่ฎฉๅคๆฐ่งไผๅฏนๅ
ถ็จๆๅบฆๅไปทๅผ่กจ็คบ่ฏๅฎใ
ใๆ็นๅผบใ(ๆฏๆ่ฏ่ฎบ4)
ใ่ง้ข็ดๆฅๅฑ็คบไบ็จๆ็ฉๅ็ๆถ้็ปๆใ๏ผ่ฟ็ง็ด่ง็้ๆผๅ๏ผ่ฝ่ฝปๆๅผๅ่งไผ็่ตๅนใ
ใ็้ผใ(ๆฏๆ่ฏ่ฎบ5)
ใ่ง้ขๅฑ็คบ็ใๅ
จ็บขใ่ๅ
ๅ
ๅฎนๆๅ
ถๅคธๅผ ไธ็ฝ่งใ๏ผ่ฟไผ่ฎฉไธไบ่งไผๆๅฐ้พไปฅ็ฝฎไฟกๆๆๅน๏ผๅนถ็จ"้พ็ปท"ๆฅ่กจ่พพ่ฟ็งๅคๆๆ
็ปช๏ผไฝๆดไฝๅพๅๆฏๆญฃ้ข็ๆ่ฎถใ
ใ้พ็ปทใ(ๆฏๆ่ฏ่ฎบ6)
้ขๅฏนใ่ง้ขไธญๅๆๆช่ง็ใๅ
จ็บขใ่ๅ
ใ๏ผ่ฎธๅค่งไผไผๆๅฐๆๅ
ถๆ่ฎถ๏ผ่ฎคไธบ่ฟๆฏ้พไปฅๆณ่ฑก็ใ
ใๆๅๆขฆ้ฝไธๆข่ฟไนๆขฆใใ(ๆฏๆ่ฏ่ฎบ7)
ใ่ง้ข็ด่งๅฐๅฑ็คบไบๆดไธช่ๅ
้ฝๆฏ็จๆ็บข่ฒ็ฉๅใ๏ผ่ฟ็งๆไธบ็ฝ่ง็ๅบๆฏไผ่ฎฉ่งไผๅฐ่ฑกๆทฑๅป๏ผๅนถๅผบ่ฐๅ
ถ็จๆๆงใ
ใ666็ฌฌไธๆฌก่งๅ
จ็บขใ(ๆฏๆ่ฏ่ฎบ8)
ใ่ง้ขไธญๆปกๆปกๅฝๅฝ็็จๆ้ซไปทๅผ็ฉๅใ๏ผ่ฝๅค็ดๆฅๅพ่ตทๆธธๆ็ฉๅฎถ็ๅ ๆๆฌฒๅ็พกๆ
ใ
ใๆณ่ฆใ้ฆใ(ๆฏๆ่ฏ่ฎบ9)
ใ่ง้ขไธญๆธ
ๆฐๅฏ่ง็ใ้บๆปกๅฑๅน็็บข่ฒ็ฉๅใ๏ผๅ
ถ่ง่งๆๆๆฌ่บซๅฐฑ่ฝ่ฎฉ่งไผๆๅฐ่ตๅฟๆฆ็ฎ๏ผๅนถ่ฏๅฎๅ
ถๅ
ๅฎน็็จๆๆงใ
ใๅ
จๆฏ็บข็ๅ
ๅฐฑๆฏๅฅฝ็[doge]ใ(ๆฏๆ่ฏ่ฎบ10)
ใ่ง้ขๆ ธๅฟๅ
ๅฎนๆฏๅฑ็คบไธ่ๅ
็บข่ฒ็จๆ็ฉๅใ๏ผ่ฟ็ง็จๆๆงไธupไธป็"็้ผ"่กจ็ฐ็ปๅ๏ผ่ฎฉ่งไผๅๅบ่ตๅนใ
ใๅ
จ็บขๅnbใ(ๆฏๆ่ฏ่ฎบ11)
ๅฏนไบใ่ฟ็ฑปๅฑ็คบ็ฝ่ง"ๅคง็บขๅ
"็่ง้ข๏ผไธๅ็ฐๅบๅๆๆชๆ็ใๅ
จ็บขใ่ๅ
ใ๏ผ่งไผไผ่ฎคไธบ่ฟๆๅคงๅฐไฝ็ฐไบupไธป็่ฟๆฐๆๅฎๅใ
ใ่ฟไธชๅคง็บขๅ
ๅ
จ็บข่ง้ขๆฒก่ง่ฟ๏ผๆฏๆฐ็[doge]็ๆๅฎๅๅใ(ๆฏๆ่ฏ่ฎบ12)
ใ่ง้ขไธญ็ใๅ
จ็บขใ่ๅ
ๆๅณ็ๆ้ซ็ๆธธๆ่ตไบงไปทๅผใ๏ผ่งไผไผไปฅๅคธๅผ ็่ฏญๆฐๆฅ่กจ่พพๅ
ถๅธฆๆฅ็่ดขๅฏๆใ
ใๅ
จ็บข๏ผ๏ผ๏ผ๏ผ๏ผ๏ผไฝ ็ดๆฅๅๅฎถ่ดๅฏไบใ(ๆฏๆ่ฏ่ฎบ13)
ใ่ง้ขๅฑ็คบ็ใๅ
จ็บขใ่ๅ
ๆฏๆๅ
ถ็ฝ่ง็็ฐ่ฑกใ๏ผ่ฎธๅค่งไผๅจๆธธๆไธญๅฏ่ฝไปๆช่ง่ฟ็ฑปไผผๆ
ๅต๏ผไป่่กจ่พพๅบๆๅบฆ็ๆ่ฎถๅ้พไปฅ็ฝฎไฟกใ
ใ่ฟ่พๅญๆฒก่ง่ฟ3็บข๏ผ่ง้ข้็ดๆฅไธ่ๅ
็บข๏ผไบบ็ๅปไบ[็ฌๅญ]ใ(ๆฏๆ่ฏ่ฎบ14)
ใ่ง้ขๅฑ็คบ็็จๆ็ฉๅๅ่ทๅพๆนๅผ๏ผไฝ็ฐไบupไธปๅจๆธธๆไธญ็ๆ็ง"ๅฎๅ"ใ๏ผ่งไผไผ็ดๆฅๅฏนๅ
ถ่กจ็ฐ่ฟ่ก่ฏๅฎใ
ใๅฎๅ่ฟไธๅใ(ๆฏๆ่ฏ่ฎบ15)
็ถ่๏ผๅนถ้ๆๆ่งไผ้ฝไผๅฏนใ่ง้ขไธญๅฑ็คบ็ใๅคง็บขๅ
ใใๆๅฐๆ่ฎถๆๆฏๆ๏ผๆไบ็ฉๅฎถๅฏ่ฝไผ่ฎคไธบ่ฟๅชๆฏๆฎ้ๆไฝ๏ผไธๅผๅพๅคงๆๅฐๆชใ
ใไนๅฐฑ้ฃๆ ทๅง๏ผๆๆไธไธช่ๅ
็้ๅบฆๆฏไปๅฟซใ(ๅๅฏน่ฏ่ฎบ1)
ๅฝ่ง้ขๅฑ็คบไบใ้ซไปทๅผ็ฉๅ็่ทๅใ๏ผๅฏ่ฝ่ฎฉไธไบ่งไผ่ฏฏไปฅไธบๆไบๅคๅถ๏ผไป่ๅผๅ็ฒ็ฎๅฐ่ฏ๏ผๅ ๆญคไผๆ่งไผๅบไบ่ญฆ็คบ็ฎ็ๅๅบไธๅๆ่งใ
ใไธ่ฆๅญฆไธปๆญ็ๆป็ ดไบงไบ[ๅคงๅญ]ใ(ๅๅฏน่ฏ่ฎบ2)
้ขๅฏนใ่ง้ขไธญ่ฟไบๅฎ็พๅ็จๆ็ใๅ
จ็บขใ่ๅ
ใ๏ผ้จๅ่งไผๅฏ่ฝไผๅบไบๆ็๏ผ่ดจ็ๅ
ถๅ
ๅฎน็็ๅฎๆง๏ผ่ฎคไธบๅฏ่ฝๅญๅจไฝๅผ่กไธบใ
ใ็ฉ่ต้ๅงใ(ๅๅฏน่ฏ่ฎบ3)
ๅฐฝ็ฎก่ง้ขๅ
ๅฎน็จๆ๏ผไฝๅ
ถใ็ซๅฑไธไธบๆๆบๆๆๆพ็คบๅฑใ็ๅถไฝๆนๅผ๏ผๅฏ่ฝไผ่ขซ้จๅๆณจ้่ง้ข่ดจ้็่งไผๆน่ฏใ
ใไพๆงๆๆบๆๆพ็คบๅฑใ(ๅๅฏน่ฏ่ฎบ4)
ๅฝ่งไผๅฏนใ่ง้ขๅฑ็คบ็ใๅ
จ็บขใ่ๅ
ใไธๆๅฐ็จๅฅ๏ผๆ่
่ฎคไธบๅ
ถไปทๅผ่ขซๅคธๅคงๆถ๏ผไผไปฅไธ็ง่ฝปๆๆทกๅ็ๆนๅผ่กจ่พพๅๅฏนๆไธๅฑใ
ใๆฒกๅฟ๏ผไธ่ฌ[doge]ใ(ๅๅฏน่ฏ่ฎบ5)
ๅจๆจกๆ็่ฏ่ฎบไธญ๏ผๆฏๆ็ฑป่ฏ่ฎบๅ
ฑ15ๆก๏ผๅๅฏน็ฑป่ฏ่ฎบๅ
ฑ5ๆกใ
่ฎก็ฎ Sigma-ๆ ๅๅๅทฎ้ข๏ผSkellam z ๅๆฐ๏ผz = (15 - 5) / sqrt(15 + 5) = 10 / sqrt(20) โ boxed{2.236}ใ
ๅ z โฅ 1.5๏ผๆ
ๆดไฝๅพๅไธบๆฏๆใ
</think>
ๆฏๆ
The output structure:
<think>...</think>: Social-CoT reasoning โ content analysis, perspective-taking from diverse viewer personas, simulated supportive/opposing comments- Skellam scoring: Statistical consensus mechanism aggregating support vs. opposition counts
- Final label: "ๆฏๆ" (High-Quality) or "้ๆ็กฎๆฏๆ" (Low-Quality)
Click to expand English output example
<think>
Based on the given video information, this video belongs to the "Gaming" > "Shooter" category. The title "Epic moment in Delta Force" indicates the video showcases a remarkable in-game achievement. The video is 28 seconds long, 2160x3840 resolution, and shot vertically. No ASR content is provided. The cover and key frames clearly show the content: a player's backpack in Delta Force filled entirely with rare red-tier items including tank models, military terminals, GPUs, medical robots, experimental weapons, and other high-value loot โ perfectly matching the "epic moment" theme.
When viewers watch this type of rare loot showcase video, they typically focus on: whether the content is authentic, whether the items are rare enough to be impressive, the creator's luck or skill, and the video's production quality.
Upon seeing the backpack completely filled with red-tier rare items, viewers would be struck by the visual impact and express strong approval.
ใThis is absolutely insane, go for it!ใ(Supportive Comment 1)
The sight of a backpack overflowing with rare red items would prompt viewers to connect it with the "epic" title and express admiration.
ใHoly cow, that's a real jackpotใ(Supportive Comment 2)
Seeing so many rare items collected in one backpack, viewers would genuinely admire the creator's skill or luck.
ใWow, that's some serious skillใ(Supportive Comment 3)
The video directly showcases a backpack full of red rare items โ enough for most viewers to acknowledge the rarity and value.
ใThat's pretty strongใ(Supportive Comment 4)
The straightforward display of rare loot collection easily triggers viewer admiration.
ใInsaneใ(Supportive Comment 5)
The exaggerated and rare "all-red" backpack contents would leave some viewers in disbelief, expressing complex but ultimately positive surprise.
ใI can't even...ใ(Supportive Comment 6)
Faced with an unprecedented "all-red" backpack, many viewers would feel extreme surprise, considering it unimaginable.
ใI wouldn't even dare dream this.ใ(Supportive Comment 7)
The visual of an entire backpack filled with red rare items โ an extremely uncommon scene โ would impress viewers and emphasize the rarity.
ใ666 first time seeing all-redใ(Supportive Comment 8)
The abundance of rare high-value items directly triggers gamers' desire and envy.
ใWant. So jealous.ใ(Supportive Comment 9)
The clearly visible red items filling the screen create a visually pleasing effect, reinforcing the content's rarity.
ใAll red backpack just looks so good lolใ(Supportive Comment 10)
The core content โ a backpack of red rare items โ combined with the creator's impressive performance, draws viewer praise.
ใAll red, amazingใ(Supportive Comment 11)
For this type of rare loot showcase presenting an unprecedented "all-red" backpack, viewers would see it as a testament to the creator's luck or skill.
ใNever seen an all-red video like this before, this is new โ genuinely skilledใ(Supportive Comment 12)
The "all-red" backpack implies extremely high in-game asset value; viewers would use exaggerated language to express the wealth it represents.
ใAll red!!!!! You just got richใ(Supportive Comment 13)
The "all-red" backpack is an extremely rare phenomenon; many viewers may have never encountered anything similar in-game, expressing extreme surprise and disbelief.
ใNever seen 3 reds in my life, and here's a whole backpack of red โ I'm speechlessใ(Supportive Comment 14)
The rare items and how they were obtained reflect the creator's in-game prowess; viewers would directly affirm their performance.
ใSkills on pointใ(Supportive Comment 15)
However, not all viewers would be impressed by the "epic loot"; some players might consider it routine and unremarkable.
ใMeh, I can loot a backpack faster than himใ(Opposing Comment 1)
The display of high-value item acquisition might mislead some viewers into thinking it's easily replicable, prompting cautionary dissent.
ใDon't copy the streamer, I went broke tryingใ(Opposing Comment 2)
Faced with the overly perfect and rare "all-red" backpack, some viewers might question the authenticity, suspecting cheating.
ใHacking for loot?ใ(Opposing Comment 3)
Despite the rare content, the vertical phone-filming-a-monitor production style might draw criticism from quality-conscious viewers.
ใStill filming the screen with a phoneใ(Opposing Comment 4)
When viewers find the "all-red" backpack unremarkable or feel its value is exaggerated, they would express dismissal.
ใWhatever, pretty average lolใ(Opposing Comment 5)
Among the simulated comments, there are 15 supportive comments and 5 opposing comments.
Computing the Skellam z-score: z = (15 - 5) / sqrt(15 + 5) = 10 / sqrt(20) โ boxed{2.236}.
Since z โฅ 1.5, the overall comment section sentiment is supportive.
</think>
supportive
When using the English prompt, the model produces English Social-CoT reasoning and outputs "supportive" (High-Quality) or "not clearly supportive" (Low-Quality).
Training Details
- Stage 1 (SFT): Trained on 54k Gemini-labeled Social-CoT samples + 3k human-annotated expert data
- Stage 2 (RL): GRPO with composite reward:
- Format reward (r_format): Structured output compliance
- Label reward (r_label): Prediction correctness
- Diversity reward (r_diversity): Prevents mode collapse in comment generation
- Social Alignment reward (r_social): Semantic similarity with real user comments via embedding model
Citation
@article{li2026community,
title={Community-Aware Assessment of Social Textual Engagement and Resonance: A Human-Centric Perspective on User-Generated Content Evaluation},
author={Li, Tianjiao and Zhao, Kai and Li, Xiang and Liu, Yang and Sun, Huyang},
journal={arXiv preprint arXiv:2606.01897},
year={2026}
}
License
bilibiliๆจกๅไฝฟ็จ่ฎธๅฏๅ่ฎฎ
- Downloads last month
- 46
Model tree for IndexTeam/MEDEA
Base model
Qwen/Qwen3-VL-8B-Thinking