Research map / Large language models / Language, safety and interpretability
Truth over Tricks: Measuring and Mitigating Shortcut Learning in Misinformation Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TruthOverTricks, an evaluation paradigm measuring intrinsic and extrinsic shortcut learning in misinformation detection across 14+2 benchmarks, and proposes SMF, an LLM-based rewriting augmentation (paraphrase, summary, neutral) that improves robustness of trainable detectors.
Problem
Misinformation detectors rely on superficial shortcuts (sentiment, style, topic, perplexity, LLM-injected cues) correlated with labels in training data, causing poor generalization to diverse and LLM-generated misinformation.
Approach
Unifies evaluation of shortcut learning by splitting shortcuts into intrinsic induction (distribution-shifted splits over sentiment/style/topic/perplexity) and extrinsic injection (LLM rewrites that inject sentiment, tone, word choice, age, gender cues), and mitigates them with LLM-based data augmentation (paraphrase, factual summary, sentiment neutralization) applied before training and inference
Key results
- Explicit shortcut injection drops LM-based detectors by about 80% on average, with the largest drop reaching 92.0% (e.g., DeBERTa under Word Choice accuracy falls from 80.2 to 6.4).
- SMF Summary augmentation raises DeBERTa under Vanilla injection from 11.2 to 36.2 accuracy, and Paraphrase on Word Choice improves results by up to 838.4% relative.
- LLM-based detectors (Mistralv3, Llama3) remain relatively robust to shortcut injection, with drops of only about 1-9%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Inject explicit shortcuts (sentiment, tone, or word choice) into text | Trainable misinformation detectors · 16 datasets | accuracy: (88.4%↓); (91.6%↓); (92.0%↓) | Hurt |
| Inject vanilla LLM rewrites | BERT and DeBERTa misinformation detectors · 16 datasets, Vanilla shortcut | accuracy: 78.1 -> 17.3; 80.2 -> 11.2 | Hurt |
| Inject age or gender attributes | Trainable misinformation detectors · 16 datasets, implicit shortcuts | accuracy: BERT: 73.4 (6.0%↓), 74.3 (4.9%↓); DeBERTa: 74.6 (6.9%↓), 74.0 (7.7%↓) | Hurt |
| Apply Summary augmentation | Trainable misinformation detectors · 16 datasets, Vanilla, Sentiment, Tone, and Word Choice shortcuts | accuracy: up to 350.8%↑ | Helped |
| Apply Paraphrase augmentation | Trainable misinformation detectors · 16 datasets, Word Choice shortcut | accuracy: up to 838.4%↑ | Helped |
| Apply any SMF augmentation to implicit Age or Gender attacks | Trainable misinformation detectors · 16 datasets, Age and Gender shortcuts | accuracy: 3.1%↓ to 6.9%↓ | Hurt |
| Combine Neutral and Summary augmentations (ablation) | DeBERTa · AntiVax, Sentiment shortcut | accuracy: 58.0 -> 69.2 (54.7%↑) | Helped |
| Combine Neutral and Summary augmentations (ablation) | DeBERTa · AntiVax, Vanilla shortcut | accuracy: 23.2 -> 34.2 (47.3%↑) | Helped |
| Combine Summary and Neutral augmentations (ablation) | DeBERTa · AntiVax, Tone shortcut | accuracy: 70.8 -> 79.0 (11.7%↑) | Helped |
| Combine Neutral and Summary augmentations (ablation) | DeBERTa · AntiVax, Word Choice shortcut | accuracy: 57.2 -> 74.2 (29.7%↑) | Helped |
Surprising findings
- Detectors trained on sentiment-injected misinformation gain up to 26.7% accuracy on external sentiment classification tasks, indicating they learn sentiment as a proxy.
- Implicit injections (age, gender) cause minimal degradation because LLMs fail to reliably simulate author attributes (human Kappa near 0).
Limitations
- Only text-based detectors and datasets are studied; social context based detectors are excluded.
- Only four intrinsic shortcuts and six injection attacks are covered, so not all shortcuts are measured.
- Implicit injections largely fail to inject the intended attribute, limiting that part of the evaluation.
Experimental setup
Datasets: RumourEval Pheme Twitter15 Twitter16 Celebrity FakeNews Politifact Gossipcop Tianchi MultiLingual AntiVax COCO
Benchmarks: RumourEval Pheme Twitter15 Twitter16 Celebrity FakeNews Politifact Gossipcop Tianchi MultiLingual AntiVax COCO
Metrics: accuracy expected calibration error ROUGE-L BERTScore Cohen's Kappa
Baselines: BERT DeBERTa CMTR DisC CATCH F3
Builds on: F3 CATCH AutoBencher
Closest papers by meaning
- SNIFFER: Multimodal Large Language Model for Explainable Out-of-Context Misinformation Detection (CVPR 2024 · similarity 0.87)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.86)
- Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning (NeurIPS 2025 · similarity 0.86)
- Text-Guided Fine-grained Counterfactual Inference for Short Video Fake News Detection (AAAI 2025 · similarity 0.86)
- Liars' Bench: Evaluating Lie Detectors for Language Models (NeurIPS 2026 · similarity 0.85)
- Drifting Away from Truth: GenAI-Driven News Diversity Challenges LVLM-Based Misinformation Detection (AAAI 2026 · similarity 0.85)
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.