The recent explosion in popularity of large language models (LLMs) has inspired learning engineers to incorporate them into adaptive educational tools that automatically score summary writing. Understanding and evaluating LLMs is vital before deploying them in critical learning environments, yet their unprecedented size and expanding number of parameters inhibits transparency and impedes trust when they underperform. Through a collaborative user-centered design process with several learning engineers building and deploying summary scoring LLMs, we characterized fundamental design challenges and goals around interpreting their models, including aggregating large text inputs, tracking score provenance, and scaling LLM interpretability methods. To address their concerns, we developed iScore, an interactive visual analytics tool for learning engineers to upload, score, and compare multiple summaries simultaneously. Tightly integrated views allow users to iteratively revise the language in summaries, track changes in the resulting LLM scores, and visualize model weights at multiple levels of abstraction. To validate our approach, we deployed iScore with three learning engineers over the course of a month. We present a case study where interacting with iScore led a learning engineer to improve their LLM's score accuracy by three percentage points. Finally, we conducted qualitative interviews with the learning engineers that revealed how iScore enabled them to understand, evaluate, and build trust in their LLMs during deployment.
Visible-infrared person re-identification (VI-ReID) is challenging due to considerable cross-modality discrepancies. Existing works mainly focus on learning modality-invariant features while suppressing modality-specific ones. However, retrieving visible images only depends on infrared samples is an extreme problem because of the absence of color information. To this end, we present the Refer-VI-ReID settings, which aims to match target visible images from both infrared images and coarse language descriptions (e.g., "a man with red top and black pants") to complement the missing color information. To address this task, we design a Y-Y-shape decomposition structure, dubbed YYDS, to decompose and aggregate texture and color features of targets. Specifically, the text-IoU regularization strategy is firstly presented to facilitate the decomposition training, and a joint relation module is then proposed to infer the aggregation. Furthermore, the cross-modal version of k-reciprocal re-ranking algorithm is investigated, named CMKR, in which three neighbor search strategies and one local query expansion method are explored to alleviate the modality bias problem of the near neighbors. We conduct experiments on SYSU-MM01, RegDB and LLCM datasets with our manually annotated descriptions. Both YYDS and CMKR achieve remarkable improvements over SOTA methods on all three datasets. Codes are available at https://github.com/dyhBUPT/YYDS.
The advent of Large Language Models (LLMs) has led to remarkable progress on a wide range of natural language processing tasks. Despite the advances, these large-sized models still suffer from hallucinating information in their output, which poses a major issue in automatic text summarization, as we must guarantee that the generated summary is consistent with the content of the source document. Previous research addresses the challenging task of detecting hallucinations in the output (i.e. inconsistency detection) in order to evaluate the faithfulness of the generated summaries. However, these works primarily focus on English and recent multilingual approaches lack German data. This work presents absinth, a manually annotated dataset for hallucination detection in German news summarization and explores the capabilities of novel open-source LLMs on this task in both fine-tuning and in-context learning settings. We open-source and release the absinth dataset to foster further research on hallucination detection in German.
Large language models (LLMs) are highly effective in various natural language processing (NLP) tasks. However, they are susceptible to producing unreliable conjectures in ambiguous contexts called hallucination. This paper presents a new method for evaluating LLM hallucination in Question Answering (QA) based on the unanswerable math word problem (MWP). To support this approach, we innovatively develop a dataset called Unanswerable Math Word Problem (UMWP) which comprises 5200 questions across five categories. We developed an evaluation methodology combining text similarity and mathematical expression detection to determine whether LLM considers the question unanswerable. The results of extensive experiments conducted on 31 LLMs, including GPT-3, InstructGPT, LLaMA, and Claude, demonstrate that in-context learning and reinforcement learning with human feedback (RLHF) training significantly enhance the model's ability to avoid hallucination. We show that utilizing MWP is a reliable and effective approach to assess hallucination. Our code and data are available at https://github.com/Yuki-Asuuna/UMWP.
This study investigates the generation of synthetic disinformation by OpenAI's Large Language Models (LLMs) through prompt engineering and explores their responsiveness to emotional prompting. Leveraging various LLM iterations using davinci-002, davinci-003, gpt-3.5-turbo and gpt-4, we designed experiments to assess their success in producing disinformation. Our findings, based on a corpus of 19,800 synthetic disinformation social media posts, reveal that all LLMs by OpenAI can successfully produce disinformation, and that they effectively respond to emotional prompting, indicating their nuanced understanding of emotional cues in text generation. When prompted politely, all examined LLMs consistently generate disinformation at a high frequency. Conversely, when prompted impolitely, the frequency of disinformation production diminishes, as the models often refuse to generate disinformation and instead caution users that the tool is not intended for such purposes. This research contributes to the ongoing discourse surrounding responsible development and application of AI technologies, particularly in mitigating the spread of disinformation and promoting transparency in AI-generated content.
Neural machine translation (NMT) has progressed rapidly in the past few years, promising improvements and quality translations for different languages. Evaluation of this task is crucial to determine the quality of the translation. Overall, insufficient emphasis is placed on the actual sense of the translation in traditional methods. We propose a bidirectional semantic-based evaluation method designed to assess the sense distance of the translation from the source text. This approach employs the comprehensive multilingual encyclopedic dictionary BabelNet. Through the calculation of the semantic distance between the source and its back translation of the output, our method introduces a quantifiable approach that empowers sentence comparison on the same linguistic level. Factual analysis shows a strong correlation between the average evaluation scores generated by our method and the human assessments across various machine translation systems for English-German language pair. Finally, our method proposes a new multilingual approach to rank MT systems without the need for parallel corpora.
There has been an increasing interest in the alignment of large language models (LLMs) with human values. However, the safety issues of their integration with a vision module, or vision language models (VLMs), remain relatively underexplored. In this paper, we propose a novel jailbreaking attack against VLMs, aiming to bypass their safety barrier when a user inputs harmful instructions. A scenario where our poisoned (image, text) data pairs are included in the training data is assumed. By replacing the original textual captions with malicious jailbreak prompts, our method can perform jailbreak attacks with the poisoned images. Moreover, we analyze the effect of poison ratios and positions of trainable parameters on our attack's success rate. For evaluation, we design two metrics to quantify the success rate and the stealthiness of our attack. Together with a list of curated harmful instructions, a benchmark for measuring attack efficacy is provided. We demonstrate the efficacy of our attack by comparing it with baseline methods.
Large language models (LLMs) are shaping a new user interface (UI) paradigm in writing tools by enabling users to generate text through prompts. This paradigm shifts some creative control from the user to the system, thereby diminishing the user's authorship and autonomy in the writing process. To restore autonomy, we introduce Textfocals, a UI prototype designed to investigate a human-centered approach that emphasizes the user's role in writing. Textfocals supports the writing process by providing LLM-generated summaries, questions, and advice (i.e., LLM views) in a sidebar of a text editor, encouraging reflection and self-driven revision in writing without direct text generation. Textfocals' UI affordances, including contextually adaptive views and scaffolding for prompt selection and customization, offer a novel way to interact with LLMs where users maintain full authorship of their writing. A formative user study with Textfocals showed promising evidence that this approach might help users develop underdeveloped ideas, cater to the rhetorical audience, and clarify their writing. However, the study also showed interaction design challenges related to document navigation and scoping, prompt engineering, and context management. Our work highlights the breadth of the design space of writing support interfaces powered by generative AI that maintain authorship integrity.
The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model's diverse response styles such as starting with "Sure" or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output, too and ii) caution against relying solely on first-token evaluation.
Diffusion models benefit from instillation of task-specific information into the score function to steer the sample generation towards desired properties. Such information is coined as guidance. For example, in text-to-image synthesis, text input is encoded as guidance to generate semantically aligned images. Proper guidance inputs are closely tied to the performance of diffusion models. A common observation is that strong guidance promotes a tight alignment to the task-specific information, while reducing the diversity of the generated samples. In this paper, we provide the first theoretical study towards understanding the influence of guidance on diffusion models in the context of Gaussian mixture models. Under mild conditions, we prove that incorporating diffusion guidance not only boosts classification confidence but also diminishes distribution diversity, leading to a reduction in the differential entropy of the output distribution. Our analysis covers the widely adopted sampling schemes including DDPM and DDIM, and leverages comparison inequalities for differential equations as well as the Fokker-Planck equation that characterizes the evolution of probability density function, which may be of independent theoretical interest.