To what extent can LLMs be used as part of a cognitive model of language generation? In this paper, we approach this question by exploring a neuro-symbolic implementation of an algorithmic cognitive model of referential expression generation by Dale & Reiter (1995). The symbolic task analysis implements the generation as an iterative procedure that scaffolds symbolic and gpt-3.5-turbo-based modules. We compare this implementation to an ablated model and a one-shot LLM-only baseline on the A3DS dataset (Tsvilodub & Franke, 2023). We find that our hybrid approach is cognitively plausible and performs well in complex contexts, while allowing for more open-ended modeling of language generation in a larger domain.
Hallucinations pose a challenge to the application of large language models (LLMs) thereby motivating the development of metrics to evaluate factual precision. We observe that popular metrics using the Decompose-Then-Verify framework, such as \FActScore, can be manipulated by adding obvious or repetitive subclaims to artificially inflate scores. This observation motivates our new customizable plug-and-play subclaim selection component called Core, which filters down individual subclaims according to their uniqueness and informativeness. We show that many popular factual precision metrics augmented by Core are substantially more robust on a wide range of knowledge domains. We release an evaluation framework supporting easy and modular use of Core and various decomposition strategies, which we recommend adoption by the community. We also release an expansion of the FActScore biography dataset to facilitate further studies of decomposition-based factual precision evaluation.
The matrix quantization entails representing matrix elements in a more space-efficient form to reduce storage usage, with dequantization restoring the original matrix for use. We formulate the Quantization Error Minimization (QEM) problem as minimizing the distance between a matrix before and after quantization, under the condition that the quantized matrix occupies the same memory space. Matrix quantization is crucial in various applications, including Large Language Models (LLMs) weight quantization, vector databases, KV cache quantization, graph compression, and image compression. Recent advancements in LLMs, such as GPT-4 and BERT, have highlighted the importance of matrix compression due to the large size of parameters and KV cache, which are stored as matrices. We propose Quantum Entanglement Trees (QET) to address the QEM problem by leveraging the local orderliness of matrix elements, involving iterative element swapping to form a locally ordered matrix. This matrix is then grouped and quantized by columns. To enhance QET, we introduce two optimizations: further quantizing residuals to reduce MSE, and using masking and batch processing to accelerate the algorithm. Experimental results demonstrate that QET can effectively reduce MSE to 5.05%, 13.33%, and 11.89% of the current best method on the LLM dataset, K cache, and V cache, respectively. Our contributions include the abstraction of the QEM problem, the design of the QET algorithm, and the proposal of two optimizations to improve accuracy and speed.
Recent work shows Large Language Models (LLMs) struggle to understand natural language constraints for various text generation tasks in zero- and few-shot settings. While, in the code domain, there is wide usage of constraints in code format to maintain the integrity of code written in Domain-Specific Languages (DSLs) like JSON and YAML which are widely used for system-level programming tasks in enterprises. Given that LLMs are increasingly used for system-level code tasks, evaluating if they can comprehend these code constraints is crucial. However, no work has been done to evaluate their controllability over code constraints. Hence, we introduce ConCodeEval, a first-of-its-kind benchmark having two novel tasks for code constraints across five representations. Our findings suggest that language models struggle with code constraints. Code languages that perform excellently for normal code tasks do not perform well when the same languages represent fine-grained constraints.
The hallucination problem of Large Language Models (LLMs) significantly limits their reliability and trustworthiness. Humans have a self-awareness process that allows us to recognize what we don't know when faced with queries. Inspired by this, our paper investigates whether LLMs can estimate their own hallucination risk before response generation. We analyze the internal mechanisms of LLMs broadly both in terms of training data sources and across 15 diverse Natural Language Generation (NLG) tasks, spanning over 700 datasets. Our empirical analysis reveals two key insights: (1) LLM internal states indicate whether they have seen the query in training data or not; and (2) LLM internal states show they are likely to hallucinate or not regarding the query. Our study explores particular neurons, activation layers, and tokens that play a crucial role in the LLM perception of uncertainty and hallucination risk. By a probing estimator, we leverage LLM self-assessment, achieving an average hallucination estimation accuracy of 84.32\% at run time.
In this paper, we propose a method for document summarization using auxiliary information. This approach effectively summarizes descriptions related to specific images, tables, and appendices within lengthy texts. Our experiments demonstrate that leveraging high-quality OCR data and initially extracted information from the original text enables efficient summarization of the content related to described objects. Based on these findings, we enhanced popular text generation model models by incorporating additional auxiliary branches to improve summarization performance. Our method achieved top scores of 4.33 and 4.66 in the long caption and short caption tracks, respectively, of the 2024 SciCAP competition, ranking highest in both categories.
Free-text rationales justify model decisions in natural language and thus become likable and accessible among approaches to explanation across many tasks. However, their effectiveness can be hindered by misinterpretation and hallucination. As a perturbation test, we investigate how large language models (LLMs) perform rationale generation under the effects of readability level control, i.e., being prompted for an explanation targeting a specific expertise level, such as sixth grade or college. We find that explanations are adaptable to such instruction, though the requested readability is often misaligned with the measured text complexity according to traditional readability metrics. Furthermore, the generated rationales tend to feature medium level complexity, which correlates with the measured quality using automatic metrics. Finally, our human annotators confirm a generally satisfactory impression on rationales at all readability levels, with high-school-level readability being most commonly perceived and favored.
In this paper, we present our champion solution to the Global Artificial Intelligence Technology Innovation Competition Track 1: Medical Imaging Diagnosis Report Generation. We select CPT-BASE as our base model for the text generation task. During the pre-training stage, we delete the mask language modeling task of CPT-BASE and instead reconstruct the vocabulary, adopting a span mask strategy and gradually increasing the number of masking ratios to perform the denoising auto-encoder pre-training task. In the fine-tuning stage, we design iterative retrieval augmentation and noise-aware similarity bucket prompt strategies. The retrieval augmentation constructs a mini-knowledge base, enriching the input information of the model, while the similarity bucket further perceives the noise information within the mini-knowledge base, guiding the model to generate higher-quality diagnostic reports based on the similarity prompts. Surprisingly, our single model has achieved a score of 2.321 on leaderboard A, and the multiple model fusion scores are 2.362 and 2.320 on the A and B leaderboards respectively, securing first place in the rankings.
This paper investigates the subtle and often concealed biases present in Large Language Models (LLMs), focusing on implicit biases that may remain despite passing explicit bias tests. Implicit biases are significant because they influence the decisions made by these systems, potentially perpetuating stereotypes and discrimination, even when LLMs appear to function fairly. Traditionally, explicit bias tests or embedding-based methods are employed to detect bias, but these approaches can overlook more nuanced, implicit forms of bias. To address this, we introduce two novel psychological-inspired methodologies: the LLM Implicit Association Test (IAT) Bias and the LLM Decision Bias, designed to reveal and measure implicit biases through prompt-based and decision-making tasks. Additionally, open-ended generation tasks with thematic analysis of word generations and storytelling provide qualitative insights into the model's behavior. Our findings demonstrate that the LLM IAT Bias correlates with traditional methods and more effectively predicts downstream behaviors, as measured by the LLM Decision Bias, offering a more comprehensive framework for detecting subtle biases in AI systems. This research advances the field of AI ethics by proposing new methods to continually assess and mitigate biases in LLMs, highlighting the importance of qualitative and decision-focused evaluations to address challenges that previous approaches have not fully captured.
Large Language Models (LLMs) generate text by sampling the next token from a probability distribution over the vocabulary at each decoding step. Popular sampling methods like top-p (nucleus sampling) often struggle to balance quality and diversity, especially at higher temperatures which lead to incoherent or repetitive outputs. We propose min-p sampling, a dynamic truncation method that adjusts the sampling threshold based on the model's confidence by using the top token's probability as a scaling factor. Our experiments on benchmarks including GPQA, GSM8K, and AlpacaEval Creative Writing show that min-p sampling improves both the quality and diversity of generated text across different model families (Mistral and Llama 3) and model sizes (1B to 123B parameters), especially at higher temperatures. Human evaluations further show a clear preference for min-p sampling, in both text quality and creativity. Min-p sampling has been adopted by popular open-source LLM frameworks, including Hugging Face Transformers, VLLM, and many others, highlighting its considerable impact on improving text generation quality.
In the current landscape of automatic language generation, there is a need to understand, evaluate, and mitigate demographic biases as existing models are becoming increasingly multilingual. To address this, we present the initial eight languages from the MASSIVE MULTILINGUAL HOLISTICBIAS (MMHB) dataset and benchmark consisting of approximately 6 million sentences representing 13 demographic axes. We propose an automatic construction methodology to further scale up MMHB sentences in terms of both language coverage and size, leveraging limited human annotation. Our approach utilizes placeholders in multilingual sentence construction and employs a systematic method to independently translate sentence patterns, nouns, and descriptors. Combined with human translation, this technique carefully designs placeholders to dynamically generate multiple sentence variations and significantly reduces the human translation workload. The translation process has been meticulously conducted to avoid an English-centric perspective and include all necessary morphological variations for languages that require them, improving from the original English HOLISTICBIAS. Finally, we utilize MMHB to report results on gender bias and added toxicity in machine translation tasks. On the gender analysis, MMHB unveils: (1) a lack of gender robustness showing almost +4 chrf points in average for masculine semantic sentences compared to feminine ones and (2) a preference to overgeneralize to masculine forms by reporting more than +12 chrf points in average when evaluating with masculine compared to feminine references. MMHB triggers added toxicity up to 2.3%.
In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide array of text-centric tasks. However, their `large' scale introduces significant computational and storage challenges, particularly in managing the key-value states of the transformer, which limits their wider applicability. Therefore, we propose to adaptively release resources from caches and rebuild the necessary key-value states. Particularly, we accomplish this by a lightweight controller module to approximate an ideal top-$K$ sparse attention. This module retains the tokens with the highest top-$K$ attention weights and simultaneously rebuilds the discarded but necessary tokens, which may become essential for future decoding. Comprehensive experiments in natural language generation and modeling reveal that our method is not only competitive with full attention in terms of performance but also achieves a significant throughput improvement of up to 221.8%. The code for replication is available on the https://github.com/WHUIR/ADORE.
The modern saying, "You Are What You Eat" resonates on a profound level, reflecting the intricate connection between our identities and the food we consume. Our project, Deep Image-to-Recipe Translation, is an intersection of computer vision and natural language generation that aims to bridge the gap between cherished food memories and the art of culinary creation. Our primary objective involves predicting ingredients from a given food image. For this task, we first develop a custom convolutional network and then compare its performance to a model that leverages transfer learning. We pursue an additional goal of generating a comprehensive set of recipe steps from a list of ingredients. We frame this process as a sequence-to-sequence task and develop a recurrent neural network that utilizes pre-trained word embeddings. We address several challenges of deep learning including imbalanced datasets, data cleaning, overfitting, and hyperparameter selection. Our approach emphasizes the importance of metrics such as Intersection over Union (IoU) and F1 score in scenarios where accuracy alone might be misleading. For our recipe prediction model, we employ perplexity, a commonly used and important metric for language models. We find that transfer learning via pre-trained ResNet-50 weights and GloVe embeddings provide an exceptional boost to model performance, especially when considering training resource constraints. Although we have made progress on the image-to-recipe translation, there is an opportunity for future exploration with advancements in model architectures, dataset scalability, and enhanced user interaction.
Recent approaches to controlled text generation (CTG) often involve manipulating the weights or logits of base language models (LMs) at decoding time. However, these methods are inapplicable to latest black-box LMs and ineffective at preserving the core semantics of the base LM's original generations. In this work, we propose Locate&Edit(L&E), an efficient and flexible energy-based approach to CTG, which edits text outputs from a base LM using off-the-shelf energy models. Given text outputs from the base LM, L&E first locates spans that are most relevant to constraints (e.g., toxicity) utilizing energy models, and then edits these spans by replacing them with more suitable alternatives. Importantly, our method is compatible with black-box LMs, as it requires only the text outputs. Also, since L&E doesn't mandate specific architecture for its component models, it can work with a diverse combination of available off-the-shelf models. Moreover, L&E preserves the base LM's original generations, by selectively modifying constraint-related aspects of the texts and leaving others unchanged. These targeted edits also ensure that L&E operates efficiently. Our experiments confirm that L&E achieves superior semantic preservation of the base LM generations and speed, while simultaneously obtaining competitive or improved constraint satisfaction. Furthermore, we analyze how the granularity of energy distribution impacts CTG performance and find that fine-grained, regression-based energy models improve constraint satisfaction, compared to conventional binary classifier energy models.
Uncertainty quantification (UQ) in natural language generation (NLG) tasks remains an open challenge, exacerbated by the closed-source nature of the latest large language models (LLMs). This study investigates applying conformal prediction (CP), which can transform any heuristic uncertainty notion into rigorous prediction sets, to black-box LLMs in open-ended NLG tasks. We introduce a novel uncertainty measure based on self-consistency theory, and then develop a conformal uncertainty criterion by integrating the uncertainty condition aligned with correctness into the CP algorithm. Empirical evaluations indicate that our uncertainty measure outperforms prior state-of-the-art methods. Furthermore, we achieve strict control over the correctness coverage rate utilizing 7 popular LLMs on 4 free-form NLG datasets, spanning general-purpose and medical scenarios. Additionally, the calibrated prediction sets with small size further highlights the efficiency of our method in providing trustworthy guarantees for practical open-ended NLG applications.
RLHF-aligned LMs have shown unprecedented ability on both benchmarks and long-form text generation, yet they struggle with one foundational task: next-token prediction. As RLHF models become agent models aimed at interacting with humans, they seem to lose their world modeling -- the ability to predict what comes next in arbitrary documents, which is the foundational training objective of the Base LMs that RLHF adapts. Besides empirically demonstrating this trade-off, we propose a potential explanation: to perform coherent long-form generation, RLHF models restrict randomness via implicit blueprints. In particular, RLHF models concentrate probability on sets of anchor spans that co-occur across multiple generations for the same prompt, serving as textual scaffolding but also limiting a model's ability to generate documents that do not include these spans. We study this trade-off on the most effective current agent models, those aligned with RLHF, while exploring why this may remain a fundamental trade-off between models that act and those that predict, even as alignment techniques improve.
With an increasing social demand for fine-grained sentiment analysis (SA), implicit sentiment analysis (ISA) poses a significant challenge with the absence of salient cue words in expressions. It necessitates reliable reasoning to understand how the sentiment is aroused and thus determine implicit sentiments. In the era of Large Language Models (LLMs), Encoder-Decoder (ED) LLMs have gained popularity to serve as backbone models for SA applications, considering impressive text comprehension and reasoning ability among diverse tasks. On the other hand, Decoder-only (DO) LLMs exhibit superior natural language generation and in-context learning capabilities. However, their responses may contain misleading or inaccurate information. To identify implicit sentiment with reliable reasoning, this study proposes RVISA, a two-stage reasoning framework that harnesses the generation ability of DO LLMs and the reasoning ability of ED LLMs to train an enhanced reasoner. Specifically, we adopt three-hop reasoning prompting to explicitly furnish sentiment elements as cues. The generated rationales are utilized to fine-tune an ED LLM into a skilled reasoner. Additionally, we develop a straightforward yet effective verification mechanism to ensure the reliability of the reasoning learning. We evaluated the proposed method on two benchmark datasets and achieved state-of-the-art results in ISA performance.
Self-consistency (SC), leveraging multiple samples from LLMs, shows significant gains on various reasoning tasks but struggles with free-form generation due to the difficulty of aggregating answers. Its variants, UCS and USC, rely on sample selection or voting mechanisms to improve output quality. These methods, however, face limitations due to their inability to fully utilize the nuanced consensus knowledge present within multiple candidate samples, often resulting in suboptimal outputs. We propose Fine-Grained Self-Consistency (FSC) to addresses these limitations by extracting and integrating segment-level commonalities from candidate samples, enhancing the performance of LLMs both in open-ended and reasoning tasks. Based on this, we present two additional strategies: candidate filtering, which enhances overall quality by identifying highly similar candidate sets, and merging, which reduces input token requirements by combining similar samples. The effectiveness of FSC is demonstrated through extensive experiments on various tasks, including summarization, code generation, and mathematical reasoning, using GPT-3.5-turbo and GPT-4. The results indicate significant improvements over baseline methods, showcasing the potential of FSC to optimize output quality by effectively synthesizing fine-grained consensus knowledge from multiple samples.
Retrieval-augmented text generation (RAG) addresses the common limitations of large language models (LLMs), such as hallucination, by retrieving information from an updatable external knowledge base. However, existing approaches often require dedicated backend servers for data storage and retrieval, thereby limiting their applicability in use cases that require strict data privacy, such as personal finance, education, and medicine. To address the pressing need for client-side dense retrieval, we introduce MeMemo, the first open-source JavaScript toolkit that adapts the state-of-the-art approximate nearest neighbor search technique HNSW to browser environments. Developed with modern and native Web technologies, such as IndexedDB and Web Workers, our toolkit leverages client-side hardware capabilities to enable researchers and developers to efficiently search through millions of high-dimensional vectors in the browser. MeMemo enables exciting new design and research opportunities, such as private and personalized content creation and interactive prototyping, as demonstrated in our example application RAG Playground. Reflecting on our work, we discuss the opportunities and challenges for on-device dense retrieval. MeMemo is available at https://github.com/poloclub/mememo.
Advancing representation learning in specialized fields like medicine remains challenging due to the scarcity of expert annotations for text and images. To tackle this issue, we present a novel two-stage framework designed to extract high-quality factual statements from free-text radiology reports in order to improve the representations of text encoders and, consequently, their performance on various downstream tasks. In the first stage, we propose a \textit{Fact Extractor} that leverages large language models (LLMs) to identify factual statements from well-curated domain-specific datasets. In the second stage, we introduce a \textit{Fact Encoder} (CXRFE) based on a BERT model fine-tuned with objective functions designed to improve its representations using the extracted factual data. Our framework also includes a new embedding-based metric (CXRFEScore) for evaluating chest X-ray text generation systems, leveraging both stages of our approach. Extensive evaluations show that our fact extractor and encoder outperform current state-of-the-art methods in tasks such as sentence ranking, natural language inference, and label extraction from radiology reports. Additionally, our metric proves to be more robust and effective than existing metrics commonly used in the radiology report generation literature. The code of this project is available at \url{https://github.com/PabloMessina/CXR-Fact-Encoder}.