We design, implement, and evaluate adversarial decoding, a new, generic text generation technique that produces readable documents for different adversarial objectives. Prior methods either produce easily detectable gibberish, or cannot handle objectives that include embedding similarity. In particular, they only work for direct attacks (such as jailbreaking) and cannot produce adversarial text for realistic indirect injection, e.g., documents that (1) are retrieved in RAG systems in response to broad classes of queries, and also (2) adversarially influence subsequent generation. We also show that fluency (low perplexity) is not sufficient to evade filtering. We measure the effectiveness of adversarial decoding for different objectives, including RAG poisoning, jailbreaking, and evasion of defensive filters, and demonstrate that it outperforms existing methods while producing readable adversarial documents.
In this work, we propose Erasure of Language Memory (ELM), an approach for concept-level unlearning built on the principle of matching the distribution defined by an introspective classifier. Our key insight is that effective unlearning should leverage the model's ability to evaluate its own knowledge, using the model itself as a classifier to identify and reduce the likelihood of generating content related to undesired concepts. ELM applies this framework to create targeted low-rank updates that reduce generation probabilities for concept-specific content while preserving the model's broader capabilities. We demonstrate ELM's efficacy on biosecurity, cybersecurity, and literary domain erasure tasks. Comparative analysis shows that ELM achieves superior performance across key metrics, including near-random scores on erased topic assessments, maintained coherence in text generation, preserved accuracy on unrelated benchmarks, and robustness under adversarial attacks. Our code, data, and trained models are available at https://elm.baulab.info
Generating commonsense assertions within a given story context remains a difficult task for modern language models. Previous research has addressed this problem by aligning commonsense inferences with stories and training language generation models accordingly. One of the challenges is determining which topic or entity in the story should be the focus of an inferred assertion. Prior approaches lack the ability to control specific aspects of the generated assertions. In this work, we introduce "hinting," a data augmentation technique that enhances contextualized commonsense inference. "Hinting" employs a prefix prompting strategy using both hard and soft prompts to guide the inference process. To demonstrate its effectiveness, we apply "hinting" to two contextual commonsense inference datasets: ParaCOMET and GLUCOSE, evaluating its impact on both general and context-specific inference. Furthermore, we evaluate "hinting" by incorporating synonyms and antonyms into the hints. Our results show that "hinting" does not compromise the performance of contextual commonsense inference while offering improved controllability.
Existing automatic prompt engineering methods are typically designed for discriminative tasks, where new task prompts are iteratively refined with limited feedback from a single metric reflecting a single aspect. However, these approaches are suboptimal for generative tasks, which require more nuanced guidance beyond a single numeric metric to improve the prompt and optimize multiple aspects of the generated text. To address these challenges, we propose a novel multi-aspect Critique-Suggestion-guided automatic Prompt Optimization (CriSPO) approach. CriSPO introduces a critique-suggestion module as its core component. This module spontaneously discovers aspects, and compares generated and reference texts across these aspects, providing specific suggestions for prompt modification. These clear critiques and actionable suggestions guide a receptive optimizer module to make more substantial changes, exploring a broader and more effective search space. To further improve CriSPO with multi-metric optimization, we introduce an Automatic Suffix Tuning (AST) extension to enhance the performance of task prompts across multiple metrics. We evaluate CriSPO on 4 state-of-the-art LLMs across 4 summarization and 5 QA datasets. Extensive experiments show 3-4% ROUGE score improvement on summarization and substantial improvement of various metrics on QA. Code available at https://github.com/amazon-science/crispo
Knowledge Graphs (KGs) have been used to organize large datasets into structured, interconnected information, enhancing data analytics across various fields. In the legislative context, one potential natural application of KGs is modeling the intricate set of interconnections that link laws and their articles with each other and the broader legislative context. At the same time, the rise of large language models (LLMs) such as GPT has opened new opportunities in legal applications, such as text generation and document drafting. Despite their potential, the use of LLMs in legislative contexts is critical since it requires the absence of hallucinations and reliance on up-to-date information, as new laws are published on a daily basis. This work investigates how Legislative Knowledge Graphs and LLMs can synergize and support legislative processes. We address three key questions: the benefits of using KGs for legislative systems, how LLM can support legislative activities by ensuring an accurate output, and how we can allow non-technical users to use such technologies in their activities. To this aim, we develop Legis AI Platform, an interactive platform focused on Italian legislation that enhances the possibility of conducting legislative analysis and that aims to support lawmaking activities.
The advancement of text generation models has granted us the capability to produce coherent and convincing text on demand. Yet, in real-life circumstances, individuals do not continuously generate text or voice their opinions. For instance, consumers pen product reviews after weighing the merits and demerits of a product, and professional analysts issue reports following significant news releases. In essence, opinion expression is typically prompted by particular reasons or signals. Despite long-standing developments in opinion mining, the appropriate timing for expressing an opinion remains largely unexplored. To address this deficit, our study introduces an innovative task - the identification of news-triggered opinion expressing timing. We ground this task in the actions of professional stock analysts and develop a novel dataset for investigation. Our approach is decision-focused, leveraging text generation models to steer the classification model, thus enhancing overall performance. Our experimental findings demonstrate that the text generated by our model contributes fresh insights from various angles, effectively aiding in identifying the optimal timing for opinion expression.
Large Language Models (LLMs) have shown the impressive capability of answering questions in a wide range of scenarios. However, when LLMs face different types of questions, it is worth exploring whether LLMs are aware that some questions have limited answers and need to respond more deterministically but some do not. We refer to this as question awareness of LLMs. The lack of question awareness in LLMs leads to two phenomena that LLMs are: (1) too casual to answer non-open-ended questions or (2) too boring to answer open-ended questions. In this paper, we first evaluate the question awareness in LLMs. The experimental results show that LLMs have the issues of lacking awareness of questions in certain domains, e.g. factual knowledge, resulting in hallucinations during the generation. To mitigate these, we propose a method called Question Awareness Temperature Sampling (QuATS). This method enhances the question awareness of LLMs by adaptively adjusting the output distributions based on question features. The automatic adjustment in QuATS eliminates the need for manual temperature tuning in text generation and consistently improves model performance in various benchmarks.
Probing learned concepts in large language models (LLMs) is crucial for understanding how semantic knowledge is encoded internally. Training linear classifiers on probing tasks is a principle approach to denote the vector of a certain concept in the representation space. However, the single vector identified for a concept varies with both data and training, making it less robust and weakening its effectiveness in real-world applications. To address this challenge, we propose an approach to approximate the subspace representing a specific concept. Built on linear probing classifiers, we extend the concept vectors into Gaussian Concept Subspace (GCS). We demonstrate GCS's effectiveness through measuring its faithfulness and plausibility across multiple LLMs with different sizes and architectures. Additionally, we use representation intervention tasks to showcase its efficacy in real-world applications such as emotion steering. Experimental results indicate that GCS concept vectors have the potential to balance steering performance and maintaining the fluency in natural language generation tasks.
This paper investigates the transformative potential of Generative AI (Gen-AI) technologies, particularly large language models, within the building industry. By leveraging these advanced AI tools, the study explores their application across key areas such as automated compliance checking and building design assistance. The research highlights how Gen-AI can automate labor-intensive processes, significantly improving efficiency and reducing costs in building practices. The paper first discusses the two widely applied fundamental models-Transformer and Diffusion model-and summarizes current pathways for accessing Gen-AI models and the most common techniques for customizing them. It then explores applications for text generation, such as compliance checking, control support, data mining, and building simulation input file editing. Additionally, it examines image generation, including direct generation through diffusion models and indirect generation through language model-supported template creation based on existing Computer-Aided Design or other design tools with rendering. The paper concludes with a comprehensive analysis of the current capabilities of Gen-AI in the building industry, outlining future directions for research and development, with the goal of paving the way for smarter, more effective, and responsive design, construction, and operational practices.
Generative models lack rigorous statistical guarantees for their outputs and are therefore unreliable in safety-critical applications. In this work, we propose Sequential Conformal Prediction for Generative Models (SCOPE-Gen), a sequential conformal prediction method producing prediction sets that satisfy a rigorous statistical guarantee called conformal admissibility control. This guarantee states that with high probability, the prediction sets contain at least one admissible (or valid) example. To this end, our method first samples an initial set of i.i.d. examples from a black box generative model. Then, this set is iteratively pruned via so-called greedy filters. As a consequence of the iterative generation procedure, admissibility of the final prediction set factorizes as a Markov chain. This factorization is crucial, because it allows to control each factor separately, using conformal prediction. In comparison to prior work, our method demonstrates a large reduction in the number of admissibility evaluations during calibration. This reduction is important in safety-critical applications, where these evaluations must be conducted manually by domain experts and are therefore costly and time consuming. We highlight the advantages of our method in terms of admissibility evaluations and cardinality of the prediction sets through experiments in natural language generation and molecular graph extension tasks.
Despite the success of Transformers, handling long contexts remains challenging due to the limited length generalization and quadratic complexity of self-attention. Thus Transformers often require post-training with a larger attention window, significantly increasing computational and memory costs. In this paper, we propose a novel attention mechanism based on dynamic context, Grouped Cross Attention (GCA), which can generalize to 1000 times the pre-training context length while maintaining the ability to access distant information with a constant attention window size. For a given input sequence, we split it into chunks and use each chunk to retrieve top-k relevant past chunks for subsequent text generation. Specifically, unlike most previous works that use an off-the-shelf retriever, our key innovation allows the retriever to learn how to retrieve past chunks that better minimize the auto-regressive loss of subsequent tokens in an end-to-end manner. Such a mechanism accommodates retrieved chunks with a fixed-size attention window to achieve long-range information access, significantly reducing computational and memory costs during training and inference. Experiments show that GCA-based models achieve near-perfect accuracy in passkey retrieval for 16M context lengths, which is 1000 times the training length.
Natural language and visualization are two complementary modalities of human communication that play a crucial role in conveying information effectively. While visualizations help people discover trends, patterns, and anomalies in data, natural language descriptions help explain these insights. Thus, combining text with visualizations is a prevalent technique for effectively delivering the core message of the data. Given the rise of natural language generation (NLG), there is a growing interest in automatically creating natural language descriptions for visualizations, which can be used as chart captions, answering questions about charts, or telling data-driven stories. In this survey, we systematically review the state of the art on NLG for visualizations and introduce a taxonomy of the problem. The NLG tasks fall within the domain of Natural Language Interfaces (NLI) for visualization, an area that has garnered significant attention from both the research community and industry. To narrow down the scope of the survey, we primarily concentrate on the research works that focus on text generation for visualizations. To characterize the NLG problem and the design space of proposed solutions, we pose five Wh-questions, why and how NLG tasks are performed for visualizations, what the task inputs and outputs are, as well as where and when the generated texts are integrated with visualizations. We categorize the solutions used in the surveyed papers based on these "five Wh-questions." Finally, we discuss the key challenges and potential avenues for future research in this domain.
Originally proposed for handling time series data, Auto-regressive Decision Trees (ARDTs) have not yet been explored for language modeling. This paper delves into both the theoretical and practical applications of ARDTs in this new context. We theoretically demonstrate that ARDTs can compute complex functions, such as simulating automata, Turing machines, and sparse circuits, by leveraging "chain-of-thought" computations. Our analysis provides bounds on the size, depth, and computational efficiency of ARDTs, highlighting their surprising computational power. Empirically, we train ARDTs on simple language generation tasks, showing that they can learn to generate coherent and grammatically correct text on par with a smaller Transformer model. Additionally, we show that ARDTs can be used on top of transformer representations to solve complex reasoning tasks. This research reveals the unique computational abilities of ARDTs, aiming to broaden the architectural diversity in language model development.
Fraud is a prevalent offence that extends beyond financial loss, causing psychological and physical harm to victims. The advancements in online communication technologies alowed for online fraud to thrive in this vast network, with fraudsters increasingly using these channels for deception. With the progression of technologies like AI, there is a growing concern that fraud will scale up, using sophisticated methods, like deep-fakes in phishing campaigns, all generated by language generation models like ChatGPT. However, the application of AI in detecting and analyzing online fraud remains understudied. We conduct a Systematic Literature Review on AI and NLP techniques for online fraud detection. The review adhered the PRISMA-ScR protocol, with eligibility criteria including relevance to online fraud, use of text data, and AI methodologies. We screened 2,457 academic records, 350 met our eligibility criteria, and included 223. We report the state-of-the-art NLP techniques for analysing various online fraud categories; the training data sources; the NLP algorithms and models built; and the performance metrics employed for model evaluation. We find that current research on online fraud is divided into various scam activitiesand identify 16 different frauds that researchers focus on. This SLR enhances the academic understanding of AI-based detection methods for online fraud and offers insights for policymakers, law enforcement, and businesses on safeguarding against such activities. We conclude that focusing on specific scams lacks generalization, as multiple models are required for different fraud types. The evolving nature of scams limits the effectiveness of models trained on outdated data. We also identify issues in data limitations, training bias reporting, and selective presentation of metrics in model performance reporting, which can lead to potential biases in model evaluation.
Large language models (LLMs) have demonstrated prominent reasoning capabilities in recommendation tasks by transforming them into text-generation tasks. However, existing approaches either disregard or ineffectively model the user-item high-order interactions. To this end, this paper presents an enhanced LLM-based recommender (ELMRec). We enhance whole-word embeddings to substantially enhance LLMs' interpretation of graph-constructed interactions for recommendations, without requiring graph pre-training. This finding may inspire endeavors to incorporate rich knowledge graphs into LLM-based recommenders via whole-word embedding. We also found that LLMs often recommend items based on users' earlier interactions rather than recent ones, and present a reranking solution. Our ELMRec outperforms state-of-the-art (SOTA) methods in both direct and sequential recommendations.
The emergence of Vision Language Models (VLMs) is a significant advancement in integrating computer vision with Large Language Models (LLMs) to produce detailed text descriptions based on visual inputs, yet it introduces new security vulnerabilities. Unlike prior work that centered on single modalities or classification tasks, this study introduces TrojVLM, the first exploration of backdoor attacks aimed at VLMs engaged in complex image-to-text generation. Specifically, TrojVLM inserts predetermined target text into output text when encountering poisoned images. Moreover, a novel semantic preserving loss is proposed to ensure the semantic integrity of the original image content. Our evaluation on image captioning and visual question answering (VQA) tasks confirms the effectiveness of TrojVLM in maintaining original semantic content while triggering specific target text outputs. This study not only uncovers a critical security risk in VLMs and image-to-text generation but also sets a foundation for future research on securing multimodal models against such sophisticated threats.
Large Vision-Language Models (LVLMs) have shown remarkable performance on many visual-language tasks. However, these models still suffer from multimodal hallucination, which means the generation of objects or content that violates the images. Many existing work detects hallucination by directly judging whether an object exists in an image, overlooking the association between the object and semantics. To address this issue, we propose Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding (HELPD). This framework incorporates hallucination feedback at both object and sentence semantic levels. Remarkably, even with a marginal degree of training, this approach can alleviate over 15% of hallucination. Simultaneously, HELPD penalizes the output logits according to the image attention window to avoid being overly affected by generated text. HELPD can be seamlessly integrated with any LVLMs. Our experiments demonstrate that the proposed framework yields favorable results across multiple hallucination benchmarks. It effectively mitigates hallucination for different LVLMs and concurrently improves their text generation quality.
ChatGPT and other state-of-the-art large language models (LLMs) are rapidly transforming multiple fields, offering powerful tools for a wide range of applications. These models, commonly trained on vast datasets, exhibit human-like text generation capabilities, making them useful for research tasks such as ideation, literature review, coding, drafting, and outreach. We conducted a study involving 13 astronomers at different career stages and research fields to explore LLM applications across diverse tasks over several months and to evaluate their performance in research-related activities. This work was accompanied by an anonymous survey assessing participants' experiences and attitudes towards LLMs. We provide a detailed analysis of the tasks attempted and the survey answers, along with specific output examples. Our findings highlight both the potential and limitations of LLMs in supporting research while also addressing general and research-specific ethical considerations. We conclude with a series of recommendations, emphasizing the need for researchers to complement LLMs with critical thinking and domain expertise, ensuring these tools serve as aids rather than substitutes for rigorous scientific inquiry.
Brain CT report generation is significant to aid physicians in diagnosing cranial diseases. Recent studies concentrate on handling the consistency between visual and textual pathological features to improve the coherence of report. However, there exist some challenges: 1) Redundant visual representing: Massive irrelevant areas in 3D scans distract models from representing salient visual contexts. 2) Shifted semantic representing: Limited medical corpus causes difficulties for models to transfer the learned textual representations to generative layers. This study introduces a Pathological Clue-driven Representation Learning (PCRL) model to build cross-modal representations based on pathological clues and naturally adapt them for accurate report generation. Specifically, we construct pathological clues from perspectives of segmented regions, pathological entities, and report themes, to fully grasp visual pathological patterns and learn cross-modal feature representations. To adapt the representations for the text generation task, we bridge the gap between representation learning and report generation by using a unified large language model (LLM) with task-tailored instructions. These crafted instructions enable the LLM to be flexibly fine-tuned across tasks and smoothly transfer the semantic representation for report generation. Experiments demonstrate that our method outperforms previous methods and achieves SoTA performance. Our code is available at "https://github.com/Chauncey-Jheng/PCRL-MRG".
Ensembling various LLMs to unlock their complementary potential and leverage their individual strengths is highly valuable. Previous studies typically focus on two main paradigms: sample-level and token-level ensembles. Sample-level ensemble methods either select or blend fully generated outputs, which hinders dynamic correction and enhancement of outputs during the generation process. On the other hand, token-level ensemble methods enable real-time correction through fine-grained ensemble at each generation step. However, the information carried by an individual token is quite limited, leading to suboptimal decisions at each step. To address these issues, we propose SweetSpan, a span-level ensemble method that effectively balances the need for real-time adjustments and the information required for accurate ensemble decisions. Our approach involves two key steps: First, we have each candidate model independently generate candidate spans based on the shared prefix. Second, we calculate perplexity scores to facilitate mutual evaluation among the candidate models and achieve robust span selection by filtering out unfaithful scores. To comprehensively evaluate ensemble methods, we propose a new challenging setting (ensemble models with significant performance gaps) in addition to the standard setting (ensemble the best-performing models) to assess the performance of model ensembles in more realistic scenarios. Experimental results in both standard and challenging settings across various language generation tasks demonstrate the effectiveness, robustness, and versatility of our approach compared with previous ensemble methods.