Research map / Large language models / Language, safety and interpretability
Data and hallucination detection
Detect pretraining data, hallucinations, and unsafe training data using token-level and attribution signals.
66 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Language, safety and interpretability, keyword cluster “watermarking · generated text · watermarks”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
9 papers in 2023–24 and 54 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 1.3% to 2.9%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2020 · Detecting Asks in Social Engineering Attacks: Impact of Linguistic and Structural Knowledge (AAAI 2020 · 7 citations)
- 2022 · Towards Fine-Grained Reasoning for Fake News Detection (AAAI 2022 · 79 citations)
- 2023 · A Holistic Approach to Undesired Content Detection in the Real World (AAAI 2023 · 119 citations)
- 2024 · Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · 221 citations)
- 2025 · DABL: Detecting Semantic Anomalies in Business Processes Using Large Language Models (AAAI 2025 · 13 citations)
- 2026 · FACTGUARD: Event-Centric and Commonsense-Guided Fake News Detection (AAAI 2026 · 1 citation)
Earliest papers here
- Detecting Asks in Social Engineering Attacks: Impact of Linguistic and Structural Knowledge (AAAI 2020 · 7 citations)
- TRENDNERT: A Benchmark for Trend and Downtrend Detection in a Scientific Domain (AAAI 2020 · 3 citations)
- Towards Fine-Grained Reasoning for Fake News Detection (AAAI 2022 · 79 citations)
What has been tried
27 papers here report 284 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 4%, hurt 26%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add the LLM Judgment Predictor | Weibo21 test · macro F1 | 0.773 -> 0.784 | Helped | Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · 221 citations) |
| add semantic post augmentation | Kaggle test · average Macro-F1 (%) | 71.40 -> 72.07 | Helped | LLM vs Small Model? Large Language Model Based Text Augmentation Enhanced Personality Detection Model (AAAI 2024 · 31 citations) |
| Remove singular-value weighting from membership score | TruthfulQA · AUROC | 77.24 -> 78.64 | Hurt | HaloScope: Harnessing Unlabeled LLM Generations for Hallucination Detection (NeurIPS 2024 · 19 citations) |
| Add RC semantic-consistency loss | PolitiFact and Twitter-16 · accuracy and MacF1 | Helped | Is LLMs Hallucination Usable? LLM-based Negative Reasoning for Fake News Detection (AAAI 2025 · 13 citations) | |
| Raise learning rate from 1 × 10^-5 to 1 × 10^-4 | BoolQ, Commonsense QA, IMDB, LSAT QA, and Truthful QA · AUC | BoolQ: 0.64 -> 0.91; Commonsense QA: 0.59 -> 0.80; IMDB: 0.76 -> 0.98; LSAT QA: 0.72 -> 0.82; Truthful QA: 0.56 -> 0.74 | Helped | Detecting Pretraining Data from Large Language Models (ICLR 2024 · 13 citations) |
| Add feature internal semantic processing | Ransomware family classification · balanced-acc | 0.529 -> 0.548 | Helped | SRDC: Semantics-based Ransomware Detection and Classification with LLM-assisted Pre-training (AAAI 2025 · 8 citations) |
| add demonstration retrieval | Pheme · accuracy | 0.762 -> 0.788 | Helped | Collaborative Evolution: Multi-Round Learning Between Large and Small Language Models for Emergent Fake News Detection (AAAI 2025 · 7 citations) |
| Train the truth direction on affirmative and negated statements instead of affirmative statements alone | Negated statements · AUROC | Helped | Truth is Universal: Robust Detection of Lies in LLMs (NeurIPS 2024 · 6 citations) | |
| Use two-step fine-tuning including the curated explainable dataset rather than step 1 only | Real-World · F-1 | 0.47 -> 0.87 (rank 8); 0.65 -> 0.92 (rank 16); 0.53 -> 0.96 (rank 32) | Helped | ScamNet: Toward Explainable Large Language Model-Based Fraudulent Shopping Website Detection (AAAI 2025 · 2 citations) |
| Combine Neutral and Summary augmentations | AntiVax, Sentiment shortcut · accuracy | 58.0 -> 69.2 (54.7%↑) | Helped | Truth over Tricks: Measuring and Mitigating Shortcut Learning in Misinformation Detection (NeurIPS 2025 · 2 citations) |
| Increase question-aligned synthetic data from M=200 to M=1000 | X-Stance (avg. 10 questions) · F1 | 0.712 -> 0.723 | Helped | The Power of LLM-Generated Synthetic Data for Stance Detection in Online Political Discussions (ICLR 2025 · 2 citations) |
| Increase examples from 0 to 10 | n2c2 test set, Drug · lenient F1 | 0.525 -> 0.400 | Hurt | Entity Only vs. Inline Approaches: Evaluating LLMs for Adverse Drug Event Detection in Clinical Text (Student Abstract) (AAAI 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 26 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Twitter16 · 3 of 26 papers
- TriviaQA · 3 of 26 papers
- Pheme · 2 of 26 papers
- Wikipedia · 2 of 26 papers
- PolitiFact · 2 of 26 papers
- Twitter-15 · 2 of 26 papers
- SQuAD · 2 of 26 papers
- BioASQ · 2 of 26 papers
Baselines
- Min-K% · 4 of 26 papers
- BERT · 3 of 26 papers
- Zlib · 3 of 26 papers
- Perplexity · 3 of 26 papers
- RoBERTa · 2 of 26 papers
- SVM · 2 of 26 papers
- p(True) · 2 of 26 papers
- Semantic Entropy · 2 of 26 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 63 papers; from OpenAlex and the papers' affiliations).
- NVIDIA · 3 papers
- National University of Singapore · 2 papers
- The Education University of Hong Kong · 2 papers
- Imperial College London · 2 papers
- University of Wisconsin - Madison · 2 papers
Use it
Download BibTeX of 25 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Text watermarking (80 papers)
- Zero-shot machine-text detection (82 papers)
Elsewhere in the topic
- LLM unlearning and benchmarks (298 papers)
- LLM pruning and scaling data (285 papers)
- Controlled text generation (220 papers)
- Faithfulness of LLM explanations (169 papers)
- Data and evaluation methodology (136 papers)
- Uncertainty quantification for LLMs (136 papers)
- Pruning and circuits in LLMs (135 papers)
- Statistical frameworks for LLM evaluation (124 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).