Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models

Vision-language alignment in Large Vision-Language Models (LVLMs) successfully enables LLMs to understand visual input. However, we find that existing vision-language alignment methods fail to transfer the existing safety mechanism for text in LLMs to vision, which leads to vulnerabilities in toxic image. To explore the cause of this problem, we give the insightful explanation of where and how the safety mechanism of LVLMs operates and conduct comparative analysis between text and vision. We find that the hidden states at the specific transformer layers play a crucial role in the successful activation of safety mechanism, while the vision-language alignment at hidden states level in current methods is insufficient. This results in a semantic shift for input images compared to text in hidden states, therefore misleads the safety mechanism. To address this, we propose a novel Text-Guided vision-language Alignment method (TGA) for LVLMs. TGA retrieves the texts related to input vision and uses them to guide the projection of vision into the hidden states space in LLMs. Experiments show that TGA not only successfully transfers the safety mechanism for text in basic LLMs to vision in vision-language alignment for LVLMs without any safety fine-tuning on the visual modality but also maintains the general performance on various vision tasks (Safe and Good).

Enhancing Training Data Attribution for Large Language Models with Fitting Error Consideration

The black-box nature of large language models (LLMs) poses challenges in interpreting results, impacting issues such as data intellectual property protection and hallucination tracing. Training data attribution (TDA) methods are considered effective solutions to address these challenges. Most recent TDA methods rely on influence functions, assuming the model achieves minimized empirical risk. However, achieving this criterion is difficult, and sourcing accuracy can be compromised by fitting errors during model training. In this paper, we introduce a novel TDA method called Debias and Denoise Attribution (DDA), which enhances influence functions by addressing fitting errors. Specifically, the debias strategy seeks to improve the performance of influence functions by eliminating the knowledge bias present in the base model before fine-tuning, while the denoise strategy aims to reduce discrepancies in influence scores arising from varying degrees of fitting during the training process through smoothing techniques. Experimental results demonstrate that our method significantly outperforms existing approaches, achieving an averaged AUC of 91.64%. Moreover, DDA exhibits strong generality and scalability across various sources and different-scale models like LLaMA2, QWEN2, and Mistral.

AdaComp: Extractive Context Compression with Adaptive Predictor for Retrieval-Augmented Large Language Models

Retrieved documents containing noise will hinder RAG from detecting answer clues and make the inference process slow and expensive. Therefore, context compression is necessary to enhance its accuracy and efficiency. Existing context compression methods use extractive or generative models to retain the most query-relevant sentences or apply the information bottleneck theory to preserve sufficient information. However, these methods may face issues such as over-compression or high computational costs. We observe that the retriever often ranks relevant documents at the top, but the exact number of documents needed to answer the query is uncertain due to the impact of query complexity and retrieval quality: complex queries like multi-hop questions may require retaining more documents than simpler queries, and a low-quality retrieval may need to rely on more documents to generate accurate outputs. Therefore, determining the minimum number of required documents (compression rate) is still a challenge for RAG. In this paper, we introduce AdaComp, a low-cost extractive context compression method that adaptively determines the compression rate based on both query complexity and retrieval quality. Specifically, we first annotate the minimum top-k documents necessary for the RAG system to answer the current query as the compression rate and then construct triplets of the query, retrieved documents, and its compression rate. Then, we use this triplet dataset to train a compression-rate predictor. Experiments on three QA datasets and one conversational Muiti-doc QA dataset show that AdaComp significantly reduces inference costs while maintaining performance nearly identical to uncompressed models, achieving a balance between efficiency and performance.

MaFeRw: Query Rewriting with Multi-Aspect Feedbacks for Retrieval-Augmented Large Language Models

In a real-world RAG system, the current query often involves spoken ellipses and ambiguous references from dialogue contexts, necessitating query rewriting to better describe user's information needs. However, traditional context-based rewriting has minimal enhancement on downstream generation tasks due to the lengthy process from query rewriting to response generation. Some researchers try to utilize reinforcement learning with generation feedback to assist the rewriter, but these sparse rewards provide little guidance in most cases, leading to unstable training and generation results. We find that user's needs are also reflected in the gold document, retrieved documents and ground truth. Therefore, by feeding back these multi-aspect dense rewards to query rewriting, more stable and satisfactory responses can be achieved. In this paper, we propose a novel query rewriting method MaFeRw, which improves RAG performance by integrating multi-aspect feedback from both the retrieval process and generated results. Specifically, we first use manual data to train a T5 model for the rewriter initialization. Next, we design three metrics as reinforcement learning feedback: the similarity between the rewritten query and the gold document, the ranking metrics, and ROUGE between the generation and the ground truth. Inspired by RLAIF, we train three kinds of reward models for the above metrics to achieve more efficient training. Finally, we combine the scores of these reward models as feedback, and use PPO algorithm to explore the optimal query rewriting strategy. Experimental results on two conversational RAG datasets demonstrate that MaFeRw achieves superior generation metrics and more stable training compared to baselines.

A Factuality and Diversity Reconciled Decoding Method for Knowledge-Grounded Dialogue Generation

Grounding external knowledge can enhance the factuality of responses in dialogue generation. However, excessive emphasis on it might result in the lack of engaging and diverse expressions. Through the introduction of randomness in sampling, current approaches can increase the diversity. Nevertheless, such sampling method could undermine the factuality in dialogue generation. In this study, to discover a solution for advancing creativity without relying on questionable randomness and to subtly reconcile the factuality and diversity within the source-grounded paradigm, a novel method named DoGe is proposed. DoGe can dynamically alternate between the utilization of internal parameter knowledge and external source knowledge based on the model's factual confidence. Extensive experiments on three widely-used datasets show that DoGe can not only enhance response diversity but also maintain factuality, and it significantly surpasses other various decoding strategy baselines.

Unveil the Duality of Retrieval-Augmented Generation: Theoretical Analysis and Practical Solution

Retrieval-augmented generation (RAG) utilizes retrieved texts to enhance large language models (LLMs). However, studies show that RAG is not consistently effective and can even mislead LLMs due to noisy or incorrect retrieved texts. This suggests that RAG possesses a duality including both benefit and detriment. Although many existing methods attempt to address this issue, they lack a theoretical explanation for the duality in RAG. The benefit and detriment within this duality remain a black box that cannot be quantified or compared in an explainable manner. This paper takes the first step in theoretically giving the essential explanation of benefit and detriment in RAG by: (1) decoupling and formalizing them from RAG prediction, (2) approximating the gap between their values by representation similarity and (3) establishing the trade-off mechanism between them, to make them explainable, quantifiable, and comparable. We demonstrate that the distribution difference between retrieved texts and LLMs' knowledge acts as double-edged sword, bringing both benefit and detriment. We also prove that the actual effect of RAG can be predicted at token level. Based on our theory, we propose a practical novel method, X-RAG, which achieves collaborative generation between pure LLM and RAG at token level to preserve benefit and avoid detriment. Experiments in real-world tasks based on LLMs including OPT, LLaMA-2, and Mistral show the effectiveness of our method and support our theoretical results.

Source Echo Chamber: Exploring the Escalation of Source Bias in User, Data, and Recommender System Feedback Loop

Recently, researchers have uncovered that neural retrieval models prefer AI-generated content (AIGC), called source bias. Compared to active search behavior, recommendation represents another important means of information acquisition, where users are more prone to source bias. Furthermore, delving into the recommendation scenario, as AIGC becomes integrated within the feedback loop involving users, data, and the recommender system, it progressively contaminates the candidate items, the user interaction history, and ultimately, the data used to train the recommendation models. How and to what extent the source bias affects the neural recommendation models within feedback loop remains unknown. In this study, we extend the investigation of source bias into the realm of recommender systems, specifically examining its impact across different phases of the feedback loop. We conceptualize the progression of AIGC integration into the recommendation content ecosystem in three distinct phases-HGC dominate, HGC-AIGC coexist, and AIGC dominance-each representing past, present, and future states, respectively. Through extensive experiments across three datasets from diverse domains, we demonstrate the prevalence of source bias and reveal a potential digital echo chamber with source bias amplification throughout the feedback loop. This trend risks creating a recommender ecosystem with limited information source, such as AIGC, being disproportionately recommended. To counteract this bias and prevent its escalation in the feedback loop, we introduce a black-box debiasing method that maintains model impartiality towards both HGC and AIGC. Our experimental results validate the effectiveness of the proposed debiasing method, confirming its potential to disrupt the feedback loop.

Cocktail: A Comprehensive Information Retrieval Benchmark with LLM-Generated Documents Integration

The proliferation of Large Language Models (LLMs) has led to an influx of AI-generated content (AIGC) on the internet, transforming the corpus of Information Retrieval (IR) systems from solely human-written to a coexistence with LLM-generated content. The impact of this surge in AIGC on IR systems remains an open question, with the primary challenge being the lack of a dedicated benchmark for researchers. In this paper, we introduce Cocktail, a comprehensive benchmark tailored for evaluating IR models in this mixed-sourced data landscape of the LLM era. Cocktail consists of 16 diverse datasets with mixed human-written and LLM-generated corpora across various text retrieval tasks and domains. Additionally, to avoid the potential bias from previously included dataset information in LLMs, we also introduce an up-to-date dataset, named NQ-UTD, with queries derived from recent events. Through conducting over 1,000 experiments to assess state-of-the-art retrieval models against the benchmarked datasets in Cocktail, we uncover a clear trade-off between ranking performance and source bias in neural retrieval models, highlighting the necessity for a balanced approach in designing future IR systems. We hope Cocktail can serve as a foundational resource for IR research in the LLM era, with all data and code publicly available at \url{https://github.com/KID-22/Cocktail}.

UnKE: Unstructured Knowledge Editing in Large Language Models

Recent knowledge editing methods have primarily focused on modifying structured knowledge in large language models, heavily relying on the assumption that structured knowledge is stored as key-value pairs locally in MLP layers or specific neurons. However, this task setting overlooks the fact that a significant portion of real-world knowledge is stored in an unstructured format, characterized by long-form content, noise, and a complex yet comprehensive nature. The "knowledge locating" and "term-driven optimization" techniques conducted from the assumption used in previous methods (e.g., MEMIT) are ill-suited for unstructured knowledge. To address these challenges, we propose a novel unstructured knowledge editing method, namely UnKE, which extends previous assumptions in the layer dimension and token dimension. Firstly, in the layer dimension, we discard the "knowledge locating" step and treat first few layers as the key, which expand knowledge storage through layers to break the "knowledge stored locally" assumption. Next, we replace "term-driven optimization" with "cause-driven optimization" across all inputted tokens in the token dimension, directly optimizing the last layer of the key generator to perform editing to generate the required key vectors. By utilizing key-value pairs at the layer level, UnKE effectively represents and edits complex and comprehensive unstructured knowledge, leveraging the potential of both the MLP and attention layers. Results on newly proposed unstructure knowledge editing dataset (UnKEBench) and traditional structured datasets demonstrate that UnKE achieves remarkable performance, surpassing strong baselines.

A Taxation Perspective for Fair Re-ranking

Fair re-ranking aims to redistribute ranking slots among items more equitably to ensure responsibility and ethics. The exploration of redistribution problems has a long history in economics, offering valuable insights for conceptualizing fair re-ranking as a taxation process. Such a formulation provides us with a fresh perspective to re-examine fair re-ranking and inspire the development of new methods. From a taxation perspective, we theoretically demonstrate that most previous fair re-ranking methods can be reformulated as an item-level tax policy. Ideally, a good tax policy should be effective and conveniently controllable to adjust ranking resources. However, both empirical and theoretical analyses indicate that the previous item-level tax policy cannot meet two ideal controllable requirements: (1) continuity, ensuring minor changes in tax rates result in small accuracy and fairness shifts; (2) controllability over accuracy loss, ensuring precise estimation of the accuracy loss under a specific tax rate. To overcome these challenges, we introduce a new fair re-ranking method named Tax-rank, which levies taxes based on the difference in utility between two items. Then, we efficiently optimize such an objective by utilizing the Sinkhorn algorithm in optimal transport. Upon a comprehensive analysis, Our model Tax-rank offers a superior tax policy for fair re-ranking, theoretically demonstrating both continuity and controllability over accuracy loss. Experimental results show that Tax-rank outperforms all state-of-the-art baselines in terms of effectiveness and efficiency on recommendation and advertising tasks.