Live Streaming Risk Assessment

Understanding Behaviors
↓
Discovering Patterns
↓
Handling Evolution

A unified research line on fine-grained understanding, cross-session reasoning, and robust learning.

Yiran Qiao et al.

Institute of Computing Technology, CAS and ByteDance China

KDD 2026 SIGIR 2026 KDD 2026

LiveRisk Dataset

Public benchmark for live streaming risk assessment

Overview of our research line on live streaming risk assessment.

Research Motivation

① Hidden Risk Signals

Risk evidence is often buried within large amounts of normal user interactions.

② Recurring Risk Scripts

Similar malicious patterns repeatedly appear across different live rooms.

③ Evolving Tactics

Attackers continuously modify surface behaviors to evade detection.

Research Highlights

🧩 Live or Lie

Action-Aware Capsule MIL for Live Streaming Risk Assessment

Understanding localized risk signals from complex interactions.

Authors: Yiran Qiao et al.

KDD 2026

Paper Code Dataset

AC-MIL framework for live streaming risk assessment — AC-MIL: action-aware capsules with graph-aware relational reasoning and dual-view integration.

Live-streaming risk often hides in long, noisy interaction sequences. We propose Action-aware Capsule Multiple Instance Learning (AC-MIL), which organizes streamer–viewer behaviors into user–timeslot capsules and reasons over their relations for room-level risk detection under weak supervision.

Action Field Encoder encodes fine-grained action sequences into contextual action embeddings.
Relational Capsule Reasoner uses graph-aware transformers with temporal, user, role-guided, and auxiliary masks.
Dual-View Integrator fuses user-centric and temporal-centric perspectives into a unified room representation.
Cross-Level Risk Decoder gates multi-level views for robust early risk prediction.

@inproceedings{qiao2026live,
            title={Live or Lie: Action-Aware Capsule Multiple Instance Learning for Risk Assessment in Live Streaming Platforms},
            author={Yiran Qiao, Jing Chen, Xiang Ao, Qiwei Zhong, Yang Liu, Qing He},
            booktitle={Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1},
            pages={1182--1193},
            year={2026}
          }

🔍 Deja Vu in Plots

Retrieval-Augmented LLMs for Cross-Session Evidence

Discovering recurring scripts through cross-session evidence.

Authors: Yiran Qiao, Xiang Ao, Jing Chen, Yang Liu, Qiwei Zhong, Qing He.

SIGIR 2026

Paper Code Dataset

Retrieval-augmented LLM reasoning framework — From PatchNet warmup to cross-session retrieval, LLM reasoning, and efficient distillation.

Malicious live-streaming scripts often reappear across rooms with surface variations. We build a retrieval-augmented LLM framework that summarizes historical risk patches, retrieves cross-session evidence, and distills reasoning signals into an efficient PatchNet for deployment.

PatchNet models interaction patches with graph-aware attention over temporal, role, user, and auxiliary relations.
Cross-Session Index stores LLM-generated semantic summaries of recurring risk evidence.
Evidence-Integrated LLM Reasoning produces patch-level and session-level risk judgments with saliency.
Cross-Granularity Distillation transfers LLM supervision to PatchNet for fast inference without retrieval overhead.

@inproceedings{qiao2026dejavu,
  title={Deja Vu in Plots: Retrieval-Augmented LLMs for Cross-Session Evidence},
  author={Yiran Qiao, Xiang Ao, Jing Chen, Yang Liu, Qiwei Zhong, Qing He},
  booktitle={Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2026}
}

🦎 Outsmarting the Chameleon

Counterfactual Decoupling for Tactical OOD Shifts

Building robust models against evolving tactics.

Authors: Yiran Qiao, Jing Chen, Jiaqi Xu, Yang Liu, Qiwei Zhong, Xiang Ao.

KDD 2026

Paper Code Dataset

Counterfactual decoupling for tactical OOD shifts — Intent–packaging disentanglement with counterfactual consistency and post-hoc calibration.

Attackers frequently change surface tactics while preserving malicious intent. We introduce a counterfactual decoupling framework that separates intent from packaging in latent space and enforces consistency under tactical distribution shifts.

Latent Representation Disentanglement splits session embeddings into intent and packaging factors.
Counterfactual Consistency Decoupling (CCD) aligns risky intent with counterfactual reconstructions at representation and prediction levels.
Post-hoc Magnitude Calibration adapts packaging statistics at test time for evolving environments.

@inproceedings{qiao2026chameleon,
  title={Outsmarting the Chameleon: Counterfactual Decoupling for Tactical OOD Shifts},
  author={Yiran Qiao, Jing Chen, Jiaqi Xu, Yang Liu, Qiwei Zhong, Xiang Ao},
  booktitle={Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2},
  year={2026}
}

Looking Forward

🤖 LLM Agents

👥 Social Simulation

📈 Dynamic User Modeling