Research map / Label-efficient and robust learning / OOD and anomaly detection
Bot Meets Shortcut: How Can LLMs Aid in Handling Unknown Invariance OOD Scenarios?
Shiyan Zheng, Herun Wan, Minnan Luo, Junhang Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on shortcut-correlated sentiment data instead of standard data (ablation) | RoBERTa+MLP · Twibot-20 shortcut test | accuracy: .682 -> .051 | Hurt |
| train on shortcut-correlated sentiment data instead of standard data (ablation) | RoBERTa+MLP · Cresci-2017 shortcut test | accuracy: .884 -> .287 | Hurt |
| train on shortcut-correlated topic data instead of standard data (ablation) | RoBERTa+MLP · Cresci-2015 shortcut test | accuracy: .985 -> .941 | No clear effect |
| add AMR+CIGA causal debiasing | RoBERTa shortcut-trained · Cresci-2017 shortcut test (sentiment) | accuracy: .287 -> .271 | No clear effect |
| add AMR+CIGA causal debiasing | RoBERTa shortcut-trained · Cresci-2015 shortcut test (topic) | accuracy: .941 -> .774 | Hurt |
| add LLM counterfactual Text-Level debiasing | RoBERTa shortcut-trained · Twibot-20 shortcut test (sentiment) | accuracy: .051 -> .229 | Improved |
| add Dataset-Level counterfactual debiasing | RoBERTa shortcut-trained · Twibot-20 shortcut test (topic) | accuracy: .164 -> .607 | Improved |
| add Model-Level contrastive fine-tuning of the feature extractor | RoBERTa shortcut-trained · Twibot-20 shortcut test (sentiment) | accuracy: .051 -> .415 | Improved |
| add Model-Level debiasing | RoBERTa shortcut-trained · Cresci-2017 shortcut test (topic) | accuracy: .587 -> .839 | Improved |
| Dataset-Level and Model-Level debiasing on values shortcut | RoBERTa shortcut-trained · Cresci-2015 shortcut test (values) | accuracy: .659 -> .633 (dataset), .615 (model); but Text-Level .762 | Mixed |
| all three LLM-based mitigation levels on text-based models | RoBERTa-based detectors · shortcut scenarios | relative accuracy: average 59% relative improvement (text), 53% (graph) | Improved |
Closest papers by meaning
- BotSim: LLM-Powered Malicious Social Botnet Simulation (AAAI 2025 · similarity 0.85)
- A Statistical Framework for Ranking LLM-based Chatbots (ICLR 2025 · similarity 0.84)
- Exploring and Mitigating Adversarial Manipulation of Voting-Based Leaderboards (ICML 2025 · similarity 0.84)
- Truth over Tricks: Measuring and Mitigating Shortcut Learning in Misinformation Detection (NeurIPS 2025 · similarity 0.84)
- Evaluating LLM-contaminated Crowdsourcing Data Without Ground Truth (NeurIPS 2025 · similarity 0.83)
- DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
