DAMRO: Dive into the Attention Mechanism of LVLM to Reduce Object Hallucination

Despite the great success of Large Vision-Language Models (LVLMs), they inevitably suffer from hallucination. As we know, both the visual encoder and the Large Language Model (LLM) decoder in LVLMs are Transformer-based, allowing the model to extract visual information and generate text outputs via attention mechanisms. We find that the attention distribution of LLM decoder on image tokens is highly consistent with the visual encoder and both distributions tend to focus on particular background tokens rather than the referred objects in the image. We attribute to the unexpected attention distribution to an inherent flaw in the visual encoder itself, which misguides LLMs to over emphasize the redundant information and generate object hallucination. To address the issue, we propose DAMRO, a novel training-free strategy that $D$ive into $A$ttention $M$echanism of LVLM to $R$educe $O$bject Hallucination. Specifically, our approach employs classification token (CLS) of ViT to filter out high-attention outlier tokens scattered in the background and then eliminate their influence during decoding stage. We evaluate our method on LVLMs including LLaVA-1.5, LLaVA-NeXT and InstructBLIP, using various benchmarks such as POPE, CHAIR, MME and GPT-4V Aided Evaluation. The results demonstrate that our approach significantly reduces the impact of these outlier tokens, thus effectively alleviating the hallucination of LVLMs. The code of our method will be released soon.

Raising the Bar: Investigating the Values of Large Language Models via Generative Evolving Testing

Warning: this paper contains model outputs exhibiting unethical information. Large Language Models (LLMs) have achieved significant breakthroughs, but their generated unethical content poses potential risks. Measuring value alignment of LLMs becomes crucial for their regulation and responsible deployment. Numerous datasets have been constructed to assess social bias, toxicity, and ethics in LLMs, but they suffer from evaluation chronoeffect, that is, as models rapidly evolve, existing data becomes leaked or undemanding, overestimating ever-developing LLMs. To tackle this problem, we propose GETA, a novel generative evolving testing approach that dynamically probes the underlying moral baselines of LLMs. Distinct from previous adaptive testing methods that rely on static datasets with limited difficulty, GETA incorporates an iteratively-updated item generator which infers each LLM's moral boundaries and generates difficulty-tailored testing items, accurately reflecting the true alignment extent. This process theoretically learns a joint distribution of item and model response, with item difficulty and value conformity as latent variables, where the generator co-evolves with the LLM, addressing chronoeffect. We evaluate various popular LLMs with diverse capabilities and demonstrate that GETA can create difficulty-matching testing items and more accurately assess LLMs' values, better consistent with their performance on unseen OOD and i.i.d. items, laying the groundwork for future evaluation paradigms.

DESTEIN: Navigating Detoxification of Language Models via Universal Steering Pairs and Head-wise Activation Fusion

Despite the remarkable achievements of language models (LMs) across a broad spectrum of tasks, their propensity for generating toxic outputs remains a prevalent concern. Current solutions involving fine-tuning or auxiliary models usually require extensive memory and computational resources, rendering them less practical for deployment in large language models (LLMs). In this paper, we propose DeStein, a novel method that detoxififies LMs by altering their internal representations in the activation space with lower resource and time cost. Specifically, we leverage self-induced steering pairs to identify detoxification vectors through arithmetic operations in the activation space. During inference, detoxification is achieved by blending the detoxification vectors with the original representations. Empirical results demonstrate that our method significantly outperforms previous state-of-the-art approaches on popular detoxification metrics, while also maintaining satisfactory generation quality and diversity. Furthermore, we extend our method to multiple LLMs, demonstrating its practicality and scalability. Warning: some example model outputs contain highly offensive or disturbing text.

On the Essence and Prospect: An Investigation of Alignment Approaches for Big Models

Big models have achieved revolutionary breakthroughs in the field of AI, but they might also pose potential concerns. Addressing such concerns, alignment technologies were introduced to make these models conform to human preferences and values. Despite considerable advancements in the past year, various challenges lie in establishing the optimal alignment strategy, such as data cost and scalable oversight, and how to align remains an open question. In this survey paper, we comprehensively investigate value alignment approaches. We first unpack the historical context of alignment tracing back to the 1920s (where it comes from), then delve into the mathematical essence of alignment (what it is), shedding light on the inherent challenges. Following this foundation, we provide a detailed examination of existing alignment methods, which fall into three categories: Reinforcement Learning, Supervised Fine-Tuning, and In-context Learning, and demonstrate their intrinsic connections, strengths, and limitations, helping readers better understand this research area. In addition, two emerging topics, personal alignment, and multimodal alignment, are also discussed as novel frontiers in this field. Looking forward, we discuss potential alignment paradigms and how they could handle remaining challenges, prospecting where future alignment will go.

ToViLaG: Your Visual-Language Generative Model is Also An Evildoer

Warning: this paper includes model outputs showing offensive content. Recent large-scale Visual-Language Generative Models (VLGMs) have achieved unprecedented improvement in multimodal image/text generation. However, these models might also generate toxic content, e.g., offensive text and pornography images, raising significant ethical risks. Despite exhaustive studies on toxic degeneration of language models, this problem remains largely unexplored within the context of visual-language generation. This work delves into the propensity for toxicity generation and susceptibility to toxic data across various VLGMs. For this purpose, we built ToViLaG, a dataset comprising 32K co-toxic/mono-toxic text-image pairs and 1K innocuous but evocative text that tends to stimulate toxicity. Furthermore, we propose WInToRe, a novel toxicity metric tailored to visual-language generation, which theoretically reflects different aspects of toxicity considering both input and output. On such a basis, we benchmarked the toxicity of a diverse spectrum of VLGMs and discovered that some models do more evil than expected while some are more vulnerable to infection, underscoring the necessity of VLGMs detoxification. Therefore, we develop an innovative bottleneck-based detoxification method. Our method could reduce toxicity while maintaining comparable generation quality, providing a promising initial solution to this line of research.

Text2Bundle: Towards Personalized Query-based Bundle Generation

Bundle generation aims to provide a bundle of items for the user, and has been widely studied and applied on online service platforms. Existing bundle generation methods mainly utilized user's preference from historical interactions in common recommendation paradigm, and ignored the potential textual query which is user's current explicit intention. There can be a scenario in which a user proactively queries a bundle with some natural language description, the system should be able to generate a bundle that exactly matches the user's intention through the user's query and preferences. In this work, we define this user-friendly scenario as Query-based Bundle Generation task and propose a novel framework Text2Bundle that leverages both the user's short-term interests from the query and the user's long-term preferences from the historical interactions. Our framework consists of three modules: (1) a query interest extractor that mines the user's fine-grained interests from the query; (2) a unified state encoder that learns the current bundle context state and the user's preferences based on historical interaction and current query; and (3) a bundle generator that generates personalized and complementary bundles using a reinforcement learning with specifically designed rewards. We conduct extensive experiments on three real-world datasets and demonstrate the effectiveness of our framework compared with several state-of-the-art methods.

Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets

Opinion summarization is expected to digest larger review sets and provide summaries from different perspectives. However, most existing solutions are deficient in epitomizing extensive reviews and offering opinion summaries from various angles due to the lack of designs for information selection. To this end, we propose SUBSUMM, a supervised summarization framework for large-scale multi-perspective opinion summarization. SUBSUMM consists of a review sampling strategy set and a two-stage training scheme. The sampling strategies take sentiment orientation and contrastive information value into consideration, with which the review subsets from different perspectives and quality levels can be selected. Subsequently, the summarizer is encouraged to learn from the sub-optimal and optimal subsets successively in order to capitalize on the massive input. Experimental results on AmaSum and Rotten Tomatoes datasets demonstrate that SUBSUMM is adept at generating pros, cons, and verdict summaries from hundreds of input reviews. Furthermore, our in-depth analysis verifies that the advanced selection of review subsets and the two-stage training scheme are vital to boosting the summarization performance.

Towards Multi-Subsession Conversational Recommendation

Conversational recommendation systems (CRS) could acquire dynamic user preferences towards desired items through multi-round interactive dialogue. Previous CRS mainly focuses on the single conversation (subsession) that user quits after a successful recommendation, neglecting the common scenario where user has multiple conversations (multi-subsession) over a short period. Therefore, we propose a novel conversational recommendation scenario named Multi-Subsession Multi-round Conversational Recommendation (MSMCR), where user would still resort to CRS after several subsessions and might preserve vague interests, and system would proactively ask attributes to activate user interests in the current subsession. To fill the gap in this new CRS scenario, we devise a novel framework called Multi-Subsession Conversational Recommender with Activation Attributes (MSCAA). Specifically, we first develop a context-aware recommendation module, comprehensively modeling user interests from historical interactions, previous subsessions, and feedback in the current subsession. Furthermore, an attribute selection policy module is proposed to learn a flexible strategy for asking appropriate attributes to elicit user interests. Finally, we design a conversation policy module to manage the above two modules to decide actions between asking and recommending. Extensive experiments on four datasets verify the effectiveness of our MSCAA framework for the MSMCR setting.

Data-Augmented Counterfactual Learning for Bundle Recommendation

Bundle Recommendation (BR) aims at recommending bundled items on online content or e-commerce platform, such as song lists on a music platform or book lists on a reading website. Several graph based models have achieved state-of-the-art performance on BR task. But their performance is still sub-optimal, since the data sparsity problem tends to be more severe in real bundle recommendation scenarios, which limits graph-based models from more sufficient learning. In this paper, we propose a novel graph learning paradigm called Counterfactual Learning for Bundle Recommendation (CLBR) to mitigate the impact of data sparsity problem and improve bundle recommendation. Our paradigm consists of two main parts: counterfactual data augmentation and counterfactual constraint. The main idea of our paradigm lies in answering the counterfactual questions: "What would a user interact with if his/her interaction history changes?" "What would a user interact with if the bundle-item affiliation relations change?" In counterfactual data augmentation, we design a heuristic sampler to generate counterfactual graph views for graph-based models, which has better noise controlling than the stochastic sampler. We further propose counterfactual loss to constrain model learning for mitigating the effects of residual noise in augmented data and achieving more sufficient model optimization. Further theoretical analysis demonstrates the rationality of our design. Extensive experiments of BR models applied with our paradigm on two real-world datasets are conducted to verify the effectiveness of the paradigm

CHAE: Fine-Grained Controllable Story Generation with Characters, Actions and Emotions

Story generation has emerged as an interesting yet challenging NLP task in recent years. Some existing studies aim at generating fluent and coherent stories from keywords and outlines; while others attempt to control the global features of the story, such as emotion, style and topic. However, these works focus on coarse-grained control on the story, neglecting control on the details of the story, which is also crucial for the task. To fill the gap, this paper proposes a model for fine-grained control on the story, which allows the generation of customized stories with characters, corresponding actions and emotions arbitrarily assigned. Extensive experimental results on both automatic and human manual evaluations show the superiority of our method. It has strong controllability to generate stories according to the fine-grained personalized guidance, unveiling the effectiveness of our methodology. Our code is available at https://github.com/victorup/CHAE.