Enhancing the attribution in large language models (LLMs) is a crucial task. One feasible approach is to enable LLMs to cite external sources that support their generations. However, existing datasets and evaluation methods in this domain still exhibit notable limitations. In this work, we formulate the task of attributed query-focused summarization (AQFS) and present WebCiteS, a Chinese dataset featuring 7k human-annotated summaries with citations. WebCiteS derives from real-world user queries and web search results, offering a valuable resource for model training and evaluation. Prior works in attribution evaluation do not differentiate between groundedness errors and citation errors. They also fall short in automatically verifying sentences that draw partial support from multiple sources. We tackle these issues by developing detailed metrics and enabling the automatic evaluator to decompose the sentences into sub-claims for fine-grained verification. Our comprehensive evaluation of both open-source and proprietary models on WebCiteS highlights the challenge LLMs face in correctly citing sources, underscoring the necessity for further improvement. The dataset and code will be open-sourced to facilitate further research in this crucial field.

With recent advancements in large language models (LLMs), alignment has emerged as an effective technique for keeping LLMs consensus with human intent. Current methods primarily involve direct training through Supervised Fine-tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), both of which require substantial computational resources and extensive ground truth data. This paper explores an efficient method for aligning black-box large models using smaller models, introducing a model-agnostic and lightweight Bayesian Persuasion Alignment framework. We formalize this problem as an optimization of the signaling strategy from the small model's perspective. In the persuasion process, the small model (Advisor) observes the information item (i.e., state) and persuades large models (Receiver) to elicit improved responses. The Receiver then generates a response based on the input, the signal from the Advisor, and its updated belief about the information item. Through training using our framework, we demonstrate that the Advisor can significantly enhance the performance of various Receivers across a range of tasks. We theoretically analyze our persuasion framework and provide an upper bound on the Advisor's regret, confirming its effectiveness in learning the optimal signaling strategy. Our Empirical results demonstrates that GPT-2 can significantly improve the performance of various models, achieving an average enhancement of 16.1% in mathematical reasoning ability and 13.7% in code generation. We hope our work can provide an initial step toward rethinking the alignment framework from the Bayesian Persuasion perspective.

Large language models are known to hallucinate when faced with unfamiliar queries, but the underlying mechanism that govern how models hallucinate are not yet fully understood. In this work, we find that unfamiliar examples in the models' finetuning data -- those that introduce concepts beyond the base model's scope of knowledge -- are crucial in shaping these errors. In particular, we find that an LLM's hallucinated predictions tend to mirror the responses associated with its unfamiliar finetuning examples. This suggests that by modifying how unfamiliar finetuning examples are supervised, we can influence a model's responses to unfamiliar queries (e.g., say ``I don't know''). We empirically validate this observation in a series of controlled experiments involving SFT, RL, and reward model finetuning on TriviaQA and MMLU. Our work further investigates RL finetuning strategies for improving the factuality of long-form model generations. We find that, while hallucinations from the reward model can significantly undermine the effectiveness of RL factuality finetuning, strategically controlling how reward models hallucinate can minimize these negative effects. Leveraging our previous observations on controlling hallucinations, we propose an approach for learning more reliable reward models, and show that they improve the efficacy of RL factuality finetuning in long-form biography and book/movie plot generation tasks.

Large language models (LLMs) have shown remarkable performance on many tasks in different domains. However, their performance in closed-book biomedical machine reading comprehension (MRC) has not been evaluated in depth. In this work, we evaluate GPT on four closed-book biomedical MRC benchmarks. We experiment with different conventional prompting techniques as well as introduce our own novel prompting method. To solve some of the retrieval problems inherent to LLMs, we propose a prompting strategy named Implicit Retrieval Augmented Generation (RAG) that alleviates the need for using vector databases to retrieve important chunks in traditional RAG setups. Moreover, we report qualitative assessments on the natural language generation outputs from our approach. The results show that our new prompting technique is able to get the best performance in two out of four datasets and ranks second in rest of them. Experiments show that modern-day LLMs like GPT even in a zero-shot setting can outperform supervised models, leading to new state-of-the-art (SoTA) results on two of the benchmarks.

Recent text-to-image (T2I) models have had great success, and many benchmarks have been proposed to evaluate their performance and safety. However, they only consider explicit prompts while neglecting implicit prompts (hint at a target without explicitly mentioning it). These prompts may get rid of safety constraints and pose potential threats to the applications of these models. This position paper highlights the current state of T2I models toward implicit prompts. We present a benchmark named ImplicitBench and conduct an investigation on the performance and impacts of implicit prompts with popular T2I models. Specifically, we design and collect more than 2,000 implicit prompts of three aspects: General Symbols, Celebrity Privacy, and Not-Safe-For-Work (NSFW) Issues, and evaluate six well-known T2I models' capabilities under these implicit prompts. Experiment results show that (1) T2I models are able to accurately create various target symbols indicated by implicit prompts; (2) Implicit prompts bring potential risks of privacy leakage for T2I models. (3) Constraints of NSFW in most of the evaluated T2I models can be bypassed with implicit prompts. We call for increased attention to the potential and risks of implicit prompts in the T2I community and further investigation into the capabilities and impacts of implicit prompts, advocating for a balanced approach that harnesses their benefits while mitigating their risks.

From paired image-text training to text-only training for image captioning, the pursuit of relaxing the requirements for high-cost and large-scale annotation of good quality data remains consistent. In this paper, we propose Text-only Synthesis for Image Captioning (ToCa), which further advances this relaxation with fewer human labor and less computing time. Specifically, we deconstruct caption text into structures and lexical words, which serve as the fundamental components of the caption. By combining different structures and lexical words as inputs to the large language model, massive captions that contain various patterns of lexical words are generated. This method not only approaches the target domain but also surpasses it by generating new captions, thereby enhancing the zero-shot generalization ability of the model. Considering the different levels of data access in the real world, we define three synthesis scenarios: cross-domain synthesis, in-domain synthesis, and data-efficient synthesis. Experiments in these scenarios demonstrate the generalizability, transferability and practicability of ToCa with a nearly 5 CIDEr improvement for zero-shot cross-domain captioning and a maximum increase of over 20 CIDEr for data-efficient captioning.

The advent of Large Language Models (LLMs) has altered the Natural Language Processing (NLP) landscape, empowering professionals across diverse disciplines with their remarkable ability to generate human-like text. Models like OpenAI's GPT [44], Meta's Llama [45], and Anthropic's Claude 3 [4] have become indispensable collaborators in many peoples' daily lives; giving rise to innovative products such as ChatGPT for general use, GitHub Copilot for code generation, DALL-E 2 for image creation, and a multitude of voice generators, including OpenAI's text-to-speech API and ElevenLabs's Generative Voice AI. Currently, LLMs are being experimentally applied across various fields, yielding mixed results [3]. While some applications seem questionable, others exhibit spectacular outcomes. One of the most contentious applications is using LLMs for tasks necessitating mathematical reasoning. Given LLMs' inherently probabilistic nature, this application was once deemed implausible. However, recent findings suggest a shift in perspective, particularly with LLMs boasting vast parameter counts [1]. As LLMs continue to scale, new capabilities emerge [48]. Crucially, these opportunities are contingent upon the thoughtful design of prompts, which helps mitigate the risk of LLMs providing irrelevant or inaccurate responses [47]. 1

Multi-modal Large Language Models (MLLMs) have demonstrated remarkable capabilities in understanding and generating content across various modalities, such as images and text. However, their interpretability remains a challenge, hindering their adoption in critical applications. This research proposes a novel approach to enhance the interpretability of MLLMs by focusing on the image embedding component. We combine an open-world localization model with a MLLM, thus creating a new architecture able to simultaneously produce text and object localization outputs from the same vision embedding. The proposed architecture greatly promotes interpretability, enabling us to design a novel saliency map to explain any output token, to identify model hallucinations, and to assess model biases through semantic adversarial perturbations.

Effectively aligning Large Language Models (LLMs) with human-centric values while preventing the degradation of abilities acquired through Pre-training and Supervised Fine-tuning (SFT) poses a central challenge in Reinforcement Learning from Human Feedback (RLHF). In this paper, we first discover that interpolating RLHF and SFT model parameters can adjust the trade-off between human preference and basic capabilities, thereby reducing the alignment tax at the cost of alignment reward. Inspired by this, we propose integrating the RL policy and SFT models at each optimization step in RLHF to continuously regulate the training direction, introducing the Online Merging Optimizer. Specifically, we merge gradients with the parameter differences between SFT and pretrained models, effectively steering the gradient towards maximizing rewards in the direction of SFT optimization. We demonstrate that our optimizer works well with different LLM families, such as Qwen and LLaMA, across various model sizes ranging from 1.8B to 8B, various RLHF algorithms like DPO and KTO, and existing model merging methods. It significantly enhances alignment reward while mitigating alignment tax, achieving higher overall performance across 14 benchmarks.

In the domain of code generation, self-debugging is crucial. It allows LLMs to refine their generated code based on execution feedback. This is particularly important because generating correct solutions in one attempt proves challenging for complex tasks. Prior works on self-debugging mostly focus on prompting methods by providing LLMs with few-shot examples, which work poorly on small open-sourced LLMs. In this work, we propose a training framework that significantly improves self-debugging capability of LLMs. Intuitively, we observe that a chain of explanations on the wrong code followed by code refinement helps LLMs better analyze the wrong code and do refinement. We thus propose an automated pipeline to collect a high-quality dataset for code explanation and refinement by generating a number of explanations and refinement trajectories and filtering via execution verification. We perform supervised fine-tuning (SFT) and further reinforcement learning (RL) on both success and failure trajectories with a novel reward design considering code explanation and refinement quality. SFT improves the pass@1 by up to 15.92% and pass@10 by 9.30% over four benchmarks. RL training brings additional up to 3.54% improvement on pass@1 and 2.55% improvement on pass@10. The trained LLMs show iterative refinement ability, and can keep refining code continuously. Lastly, our human evaluation shows that the LLMs trained with our framework generate more useful code explanations and help developers better understand bugs in source code.