arXiv preprint 2611.01710
Find the risk.
Trace it back.
CST: Lightweight Risk Triage with Source-Grounded Evidence for Web Discussions
CST helps moderators decide where to look, what to read, and how to check the evidence in the original discussion. The pipeline runs on a CPU.
1 Fudan University · 2 University of Shanghai for Science and Technology · 3 Alibaba Group
† Corresponding author
karma / flagging
Character features, sentence embeddings and reply links inform the queue.
Illustrative thread from Figures 3 and 8 · View paper figures
Method
Three decisions in one review.
CST ranks discussions by risk, then separately selects and organizes evidence for review. Grouping leaves the risk scores unchanged.
Explore the examples
What changes when you group the evidence?
Use the illustrative threads from Figures 1 and 2 to separate reply context, message selection, and presentation.
Figure 1
Same reply, different stance.
Change Alex’s message and watch what Sam’s “I agree.” endorses.
Sam supports the removal.
Figure 2a
Spend the same budget differently.
The discussion and risk order are fixed. Pick how many messages a reviewer can read.
At two messages, group round-robin reaches both topics, while risk order stays on the team topic.
Risk order
Group round-robin
Figure 2b
Rearrange the same three messages.
Switch layouts. The message IDs, wording, risk scores, and source links stay fixed in the paper comparison.
team
m1Leave the team.m2Nobody wants you on the team.
accent
m3Your accent is ridiculous.
m1Leave the team.m3Your accent is ridiculous.m2Nobody wants you on the team.
Evaluation
Compare the results.
Explore the paper’s evidence recall, classification performance, and end-to-end CPU cost.
136 discussions · 1,960 dangerous messages
| Selector | 1 | 3 | 5 | 10 |
|---|---|---|---|---|
| Original order | 1.38 | 4.13 | 7.14 | 14.49 |
| Risk order | 4.29 | 11.43 | 17.86 | 31.12 |
| MMR | 4.29 | 11.48 | 18.37 | 32.14 |
| Group round-robin | 4.29 | 12.55 | 20.05 | 34.39 |
Figure 4. Recall is measured across all dangerous messages in the selected cohort. Every discussion receives the selected budget.
Average precision on the test set
Table 3. XLM-R is fully fine-tuned. CST uses a frozen sentence encoder with character features.
Mean total processing time · lower is better
Figure 6 and Section 5.7. Three four-thread passes over 136 SCCD discussions and 8,028 messages, including scoring and evidence outputs.
Paper figures
See the original figures.
Figure 8
From a phrase back to its reply.
- Original discussionKeep the intervening reply in the thread.
- Phrase and message IDLocate “flagging” in the original text of m9.
- Reply pathTrace the chain from m8 through m11.
Citation
Cite CST.
Read the full paper on arXiv or use the BibTeX entry in your bibliography.
@misc{wang2026cstlightweightrisktriage,
title={CST: Lightweight Risk Triage with Source-Grounded Evidence for Web Discussions},
author={Wang, Zhihua and Gao, Yifeng and Feng, Yunhao and Pan, Rui and
Zhao, Baiheng and Chen, Zixing and Li, Nan},
year={2026},
eprint={2611.01710},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2611.01710}
}