Awesome-LLM-Hallucination

My personal paper selection for LLM Hallucination.

---

1. Analysis

• [Analysis] Why Does ChatGPT Fall Short in Providing Truthful Answers? link arXiv'23

This paper attributes untruthful answers from LLMs into four types of errors: 1) comprehension error of user queries; 2) specificity errors - fail to provide answer in the an expected level of specificity; 3) reasoning errors - mistakes in logical reasoning; 4) facutality errors - mainly result from knowledge memorization and recall.

• [Analysis] Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models link arXiv'23

This paper propose SAT Probe, a method probing attention patterns, that can predict factual errors.

• [Analysis] Do Large Language Models Know What They Don’t Know? link ACL'23-findings

Would LLMs have the self-awareness of admiting their own limitations? Such as, identiying unknowable questions or questions that they have no idea about? This paper developed a new dataset contains five types of unanswerable questions and their answerable counterparts. They also introduce a simple similarity method to classify questions into answerable and unanswerable.

• [Analysis] Do Language Models Know When They're Hallucinating References? link EACL'24-finding

This paper focuses on the hallucination on references/citations/authors. And found they seem to aware they are hallucinating.

2. Fine-tuning

• [Fine-tuning] Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue link ACL'23

This paper was motivated by an observation: current LLMs training often solely rely on golden samples, but in real life, the query given by users are often not accurate and noisy, therefore it's important for models to learn how to handle incorrect knowledge in user queries.

Therefore the authors propose to train the model on examples with different levels of incorrect knowledge (augmentatively altered (minor changes) and contrastively altered (major errors) knowledge samples). Alongside the incorrect knowledge, it designs responses that are euphemistic or vaguely correct.

• [Fine-tuning] Fine-tuning Language Models for Factuality link arXiv'23

This paper generate a factuality dataset automatically, by automated factuality judging systems (external knowledge base or model’s confidence scores). Then models are fine-tuned with DPO to improve factuality.

• [Fine-tuning] Teaching Language Models to Hallucinate Less with Synthetic Tasks link arXiv'23

This paper propose to train LLMs on simple synthetic tasks where huallucination are common and easy to detect, with prompt-tuning method. Then the tuned-prompt are be used directly in real-life tasks.

• [Fine-tuning] GRATH: Gradual Self-Truthifying for Large Language Models link arXiv'24

Again, another paper that generate correct-incorrect answer pairs, and fine-tune the model with DPO.

3. Decoding

• [Decoding] Faithfulness-Aware Decoding Strategies for Abstractive Summarization link EACL'23

This paper propose two new decoding methods: 1) rank beams based on faithfulness score; 2) lookahead heuristics: predict unfinished beams' faithfulness score to find the more faithful continuation.

Among these, factuality errors are the most common and severe. The paper argues that we can use simple synthetic tasks to probe models' capabilities on knowledge memorization and recall. The paper also suggests we can use external knowledge base to improve knowledge memorization and triggers or cues to improve knowledge recall.

• [Decoding] DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models link arXiv'23

This paper identifies that factul knowledge are distributioned unevenly across layers in LLMs, and the earlier layers are more likely to produce factual errors. Therefore, the authors contrast project of later layers against the earlier layers (on the vocab space) during predicting the next token.

• [Decoding] KCTS: Knowledge-Constrained Tree Search Decoding with Token-Level Hallucination Detection link arXiv'23.10

This paper propose to use a knowledge classifier score and Monte Carlo Tree Search to guide the decoding process.

4. Post processing

• [Post-processing] Self-Contradictory Hallucinations of Large Language Models: Evaluation, Detection, and Mitigation link ICLR'24

This paper propose methods to automatically detect and mitigate self-contradictions in LLMs outputs. It is training-free so is thus applicable to black-box models.