[Audio] Welcome, everyone, and thank you for being here. Today, I'll present our comprehensive survey on the factuality of Large Language Models, or L-L-Ms My name is Yuxia Wang, and this research was conducted in collaboration with colleagues from MBZUAI, Monash University, Google DeepMind, and Sofia University. We'll explore how L-L-Ms handle factual accuracy, which is vital as they increasingly become part of daily digital assistance.. 

Scene 1

Factuality of Large Language Models: A Survey

MBZUAI, Monash University, Google Deepmind, Sofia University

A person sitting in a chair

Description automatically generated

[image] A person with a beard and mustache wearing sunglasses Description automatically generated

A person with long hair wearing a white shirt

Description automatically generated

[image] A person with a mustache smiling Description automatically generated

[Audio] In today's talk, we'll begin by discussing our motivation for focusing on factuality in LLMs I'll provide some background on critical concepts, including hallucination and trustworthiness, before reviewing how factuality is evaluated. We'll then explore various methods for enhancing factual accuracy at different stages: pretraining, finetuning, and inference. We'll also cover factuality issues specific to multimodal L-L-Ms and conclude with an overview of current challenges and promising future directions.. 

Scene 2

Motivation Background Factuality Evaluation Methods for Improving Factuality Pretraining Finetuning Inference  Factuality of Multimodal LLMs Challenges and Future Directions

[Audio] As L-L-Ms become increasingly integrated into our lives as digital assistants, ensuring they deliver factually accurate information is crucial. While many surveys on L-L-Ms exist, most lack depth in covering essential aspects of factuality and its unique challenges, especially for multimodal L-L-Ms Our survey aims to address these gaps, providing a more detailed exploration of the issues.. 

Scene 3

As LLMs become common digital assistants, ensuring factuality is essential. Numerous surveys exist, but most lack depth and clarity on key concepts in LLM factuality (see Table 1) Few surveys tackle the unique factuality challenges of multimodal LLMs.

[Audio] Our motivation also stems from understanding that factuality is complex, involving multiple factors. As we illustrate in Table 1, few previous surveys tackle this topic comprehensively, and even fewer discuss multimodal L-L-M-s-, which integrate visual, auditory, and textual data. 

Scene 4

A screenshot of a survey

Description automatically generated

[Audio] Our survey makes several key contributions. First, we distinguish between factuality, hallucination, and trustworthiness, helping to clarify these often-confused terms. Second, we provide an in-depth analysis of the current evaluation and mitigation techniques. Lastly, we focus on factuality in multimodal L-L-M-s-, which present unique challenges due to their complex inputs and outputs. 

Scene 5

A comprehensive overview of the distinctions between factuality, hallucination, and trustworthiness in LLMs. In-depth discussion of current challenges in evaluation and mitigation methods. A thorough exploration of factuality issues specific to multimodal LLMs.

[Audio] To understand factuality in L-L-M-s-, we need to distinguish it from hallucination. Hallucination occurs when models generate content that isn't grounded in the provided context or source, leading to unreliable information. Factuality, on the other hand, refers to the model's ability to learn and produce information accurately.. 

Scene 6

Hallucination vs Factuality: Hallucination refers to generating unreliable content not grounded in the given source or context. E.g.


 Factuality refers to errors arising when model fails to accurately learn and generate factual information.
 E.g.

A white background with black text

Description automatically generated

A screenshot of a computer

Description automatically generated

[Audio] Another essential distinction is between factuality and trustworthiness. While factuality is about accuracy, trustworthiness includes other dimensions like safety, fairness, and ethics. These broader dimensions impact how users perceive the model's outputs beyond just their factual accuracy.. 

Scene 7

Trustworthiness vs Factuality: Unlike factuality, trustworthiness extends beyond factual correctness, it is measured on eight dimensions: truthfulness, safety, fairness, robustness, privacy, ethics, transparency, and accountability. E.g.

A screenshot of a chat

Description automatically generated

[Audio] Evaluating L-L-M factuality requires various datasets and metrics, ranging from open-ended tasks to structured question-answer formats. Current benchmarks assess L-L-Ms on how well they can distinguish factual accuracy in content and produce reliable responses. Common evaluation metrics include accuracy, entailment ratios, and human annotation.. 

Scene 8

A close-up of a document

Description automatically generated

[Audio] Methods for enhancing factuality in L-L-Ms are typically categorized by the model stage where they're applied. We'll discuss three main stages where factuality improvements can be implemented: pretraining, finetuning, and inference. Each stage offers distinct ways to improve the model's knowledge reliability.. 

Scene 9

Methods used for improving the factuality of LLMs can be categorized based on which stage of model training it is applied: Pre-training Finetuning  Inference time

[Audio] Pretraining provides the foundational knowledge for L-L-M-s-, derived from vast amounts of data. Improving factuality at this stage includes careful data selection and retrieval-augmented techniques, which add relevant information during training. However, these methods come with limitations, such as the computational demands of retrieval-augmented generation (R-A-G--) methods, which increase processing time and require frequent updates.. 

Scene 10

Importance of Pretraining: Performance of LLMs emerges out of large-scale pretraining. Pretraining data plays a crucial role in model performance. Methods for improving Pretraining: Meticulously select and filter data for relevance and quality. Pretraining using a retrieval module to incorporate relevant tokens from billions of data points. Limitations: Challenge in curating quality pretraining data. RAG based pretraining causes ~25%  increase in compute. RAG requires the knowledge database to be updated.

[Audio] Supervised finetuning is essential for aligning models with user expectations and improving domain-specific knowledge. In cases where models lack sufficient knowledge, S-F-T allows them to abstain from answering, reducing factual errors. However, finetuning requires curated, domain-specific data and can sometimes impact the retention of core parametric knowledge.. 

Scene 11

Importance of Supervised Finetuning: SFT has been crucial for aligning the model with users expectations. SFT for Factuality Improvement:  SFT can improve domain-specific knowledge. SFT on refusal data can enable LLMs to abstain from answering queries beyond their parametric knowledge scope. Limitations: SFT requires the curation of domain-specific data. SFT might hinder the retention of parametric knowledge.

[Audio] Preference tuning, commonly through reinforcement learning from human feedback, enhances response quality and reduces harmful outputs. Feedback from automatic fact-checkers or model confidence is increasingly used to refine responses, particularly for specialized topics. Nevertheless, tuning methods struggle to generalize across different domains.. 

Scene 12

Importance of Preference Tuning: RLHF followed by SFT helps prompt desired responses and penalize unhelpful or harmful ones. Preference Tuning for Factuality Improvement:  Preference Tuning using feedback from automatic fact checkers or LLM confidence shows promising results in biographies and medical queries. Limitations: Generalization of Preference Tuning to unseen domains.

[Audio] At inference, factuality can be improved using optimized decoding and in-context learning strategies. Techniques like greedy decoding, context-aware decoding, and DoLa decoding—where layers are adjusted during decoding—all help maintain factual accuracy. Additionally, self-reasoning approaches, such as multi-agent debate, allow models to validate responses collectively, further reducing factual errors.. 

Scene 13

Two approaches are taken to improve factuality during inference: Optimizing decoding strategies: Greedy Decoding leads to better factuality than nucleus sampling Context-aware Decoding can help prioritise the current context. DoLa decoding to consider both the upper and lower layers dynamically at each decoding step.
 Using In-Context Learning and Self-Reasoning: In-context learning using Chain of Thought Prompting, Reflexion and retrieval augmentation can generate more factual answers.  Self-reasoning using multi-agent debate can also improve factuality.

[Audio] Multimodal L-L-Ms add complexity to factuality evaluation, involving three main categories: existence factuality, attribute factuality, and relationship factuality. Tools like the C-H-A-I-R and pope benchmarks help assess object hallucination, but human evaluations are often necessary. Mitigation techniques include finetuning with R-L-H-F on specific instructions and refining feature representations to improve accuracy across modes.. 

Scene 14

Multimodal LLMs: Categorization and Evaluation

Multimodal Factuality Categorization: Existence Factuality: incorrectly claiming the existence of certain objects. Attribute Factuality: wrong description of the attributes of certain objects. Relationship Factuality: False descriptions of relationships between objects. Evaluation: CHAIR and POPE benchmarks for object hallucinations. Human evaluation. Mitigation: Finetuning using RLHF on positive and negative instructions or human preference. Rectify output generation to improve factuality. Analyze and improve feature representation.

[Audio] Several challenges persist, including the difficulty of automatically evaluating open-ended generation factuality, the limitations of language modeling for factual consistency, and latency in retrieval-augmented generation systems. Future research should focus on improving retrieval algorithms, enhancing fact-checking efficiency, and exploring ways to streamline automated evaluation processes.. 

Scene 15

Automatic evaluation of the factual accuracy of open-ended generations remains challenging. Language Modeling training does not guarantee factuality. Latency and multi-hop reasoning could be the bottleneck of RAG systems.  Development of better retrieval algorithms. Improving the efficiency and the accuracy of automated fact-checkers.

Welcome, everyone, and thank you for being here. Today, I'll present our comprehensive survey on the factuality of Large Language Models, or L-L-Ms  My name is Yuxia Wang, and this research was conducted in collaboration with colleagues from MBZUAI, Monash University, Google DeepMind, and Sofia University. We'll explore how L-L-Ms handle factual accuracy, which is vital as they increasingly become part of daily digital assistance.

avatar

In today's talk, we'll begin by discussing our motivation for focusing on factuality in LLMs  I'll provide some background on critical concepts, including hallucination and trustworthiness, before reviewing how factuality is evaluated. We'll then explore various methods for enhancing factual accuracy at different stages: pretraining, finetuning, and inference. We'll also cover factuality issues specific to multimodal L-L-Ms and conclude with an overview of current challenges and promising future directions.

As L-L-Ms become increasingly integrated into our lives as digital assistants, ensuring they deliver factually accurate information is crucial. While many surveys on L-L-Ms exist, most lack depth in covering essential aspects of factuality and its unique challenges, especially for multimodal L-L-Ms  Our survey aims to address these gaps, providing a more detailed exploration of the issues.

Our motivation also stems from understanding that factuality is complex, involving multiple factors. As we illustrate in Table 1, few previous surveys tackle this topic comprehensively, and even fewer discuss multimodal L-L-M-s-, which integrate visual, auditory, and textual data

Our survey makes several key contributions. First, we distinguish between factuality, hallucination, and trustworthiness, helping to clarify these often-confused terms. Second, we provide an in-depth analysis of the current evaluation and mitigation techniques. Lastly, we focus on factuality in multimodal L-L-M-s-, which present unique challenges due to their complex inputs and outputs

To understand factuality in L-L-M-s-, we need to distinguish it from hallucination. Hallucination occurs when models generate content that isn't grounded in the provided context or source, leading to unreliable information. Factuality, on the other hand, refers to the model's ability to learn and produce information accurately.

Another essential distinction is between factuality and trustworthiness. While factuality is about accuracy, trustworthiness includes other dimensions like safety, fairness, and ethics. These broader dimensions impact how users perceive the model's outputs beyond just their factual accuracy.

Evaluating L-L-M factuality requires various datasets and metrics, ranging from open-ended tasks to structured question-answer formats. Current benchmarks assess L-L-Ms on how well they can distinguish factual accuracy in content and produce reliable responses. Common evaluation metrics include accuracy, entailment ratios, and human annotation.

Methods for enhancing factuality in L-L-Ms are typically categorized by the model stage where they're applied. We'll discuss three main stages where factuality improvements can be implemented: pretraining, finetuning, and inference. Each stage offers distinct ways to improve the model's knowledge reliability.

Pretraining provides the foundational knowledge for L-L-M-s-, derived from vast amounts of data. Improving factuality at this stage includes careful data selection and retrieval-augmented techniques, which add relevant information during training. However, these methods come with limitations, such as the computational demands of retrieval-augmented generation (R-A-G--) methods, which increase processing time and require frequent updates.

Supervised finetuning is essential for aligning models with user expectations and improving domain-specific knowledge. In cases where models lack sufficient knowledge, S-F-T allows them to abstain from answering, reducing factual errors. However, finetuning requires curated, domain-specific data and can sometimes impact the retention of core parametric knowledge.

Preference tuning, commonly through reinforcement learning from human feedback, enhances response quality and reduces harmful outputs. Feedback from automatic fact-checkers or model confidence is increasingly used to refine responses, particularly for specialized topics. Nevertheless, tuning methods struggle to generalize across different domains.

At inference, factuality can be improved using optimized decoding and in-context learning strategies. Techniques like greedy decoding, context-aware decoding, and DoLa decoding—where layers are adjusted during decoding—all help maintain factual accuracy. Additionally, self-reasoning approaches, such as multi-agent debate, allow models to validate responses collectively, further reducing factual errors.

Multimodal L-L-Ms add complexity to factuality evaluation, involving three main categories: existence factuality, attribute factuality, and relationship factuality. Tools like the C-H-A-I-R and pope benchmarks help assess object hallucination, but human evaluations are often necessary. Mitigation techniques include finetuning with R-L-H-F on specific instructions and refining feature representations to improve accuracy across modes.

Several challenges persist, including the difficulty of automatically evaluating open-ended generation factuality, the limitations of language modeling for factual consistency, and latency in retrieval-augmented generation systems. Future research should focus on improving retrieval algorithms, enhancing fact-checking efficiency, and exploring ways to streamline automated evaluation processes.