Chain-of-Thought (CoT) reasoning has enabled Large Language Model (LLM) to achieve remarkable performance in various NLP tasks, including arithmetic problem-solving. However, this success does not generalize to small language model (sLM) like T5, due to their limited capacity and absence of emergent abilities associated with larger models. Recent works to enhance sLM through knowledge distillation have yielded some improvements but still face significant limitations, particularly high ambiguity from the variability in natural language expressions and substantial computational costs. In this paper, we investigate why sLM perform poorly on arithmetic reasoning tasks and hypothesize that natural language format variability introduces high ambiguity for these smaller models. Based on this hypothesis, we conduct experiments with equation-only format, which is a reasoning format that unifies arithmetic reasoning previously expressed in natural language formats into mathematical equations. Experiment results demonstrate that equation-only format effectively boosts the arithmetic reasoning abilities of sLM, especially in very small models like T5-Tiny.
Knowledge graph reasoning plays a vital role in various applications and has garnered considerable attention. Recently, path-based methods have achieved impressive performance. However, they may face limitations stemming from constraints in message-passing neural networks, such as missing paths and information over-squashing. In this paper, we revisit the application of transformers for knowledge graph reasoning to address the constraints faced by path-based methods and propose a novel method KnowFormer. KnowFormer utilizes a transformer architecture to perform reasoning on knowledge graphs from the message-passing perspective, rather than reasoning by textual information like previous pretrained language model based methods. Specifically, we define the attention computation based on the query prototype of knowledge graph reasoning, facilitating convenient construction and efficient optimization. To incorporate structural information into the self-attention mechanism, we introduce structure-aware modules to calculate query, key, and value respectively. Additionally, we present an efficient attention computation method for better scalability. Experimental results demonstrate the superior performance of KnowFormer compared to prominent baseline methods on both transductive and inductive benchmarks.
Iterative human engagement is a common and effective means of leveraging the advanced language processing power of large language models (LLMs). Using well-structured prompts in a conversational manner, human users can effectively influence an LLM to develop more thoughtful and accurate responses. Motivated by this insight, we propose the Iteration of Thought (IoT) framework for enhancing LLM responses by generating "thought"-provoking prompts vis a vis an input query and the current iteration of an LLM's response. Unlike static or semi-static approaches, e.g. Chain of Thought (CoT) or Tree of Thoughts (ToT), IoT adapts its reasoning path dynamically, based on evolving context, and without generating alternate explorative thoughts which are ultimately discarded. The three components of the IoT framework are (1) an Inner Dialogue Agent (IDA) responsible for generating instructive, context-specific prompts; (2) an LLM Agent (LLMA) that processes these prompts to refine its responses; and (3) an iterative prompting loop that implements a conversation between the former two components. We introduce two variants of our framework: Autonomous Iteration of Thought (AIoT), where an LLM decides when to stop iterating, and Guided Iteration of Thought (GIoT), which always forces a fixed number iterations. We investigate the performance of IoT across various datasets, spanning complex reasoning tasks from the GPQA dataset, explorative problem-solving in Game of 24, puzzle solving in Mini Crosswords, and multi-hop question answering from the HotpotQA dataset. Our results show that IoT represents a viable paradigm for autonomous response refinement in LLMs, showcasing significant improvements over CoT and thereby enabling more adaptive and efficient reasoning systems that minimize human intervention.
Unified information extraction (UIE) aims to extract diverse structured information from unstructured text. While large language models (LLMs) have shown promise for UIE, they require significant computational resources and often struggle to generalize to unseen tasks. We propose RUIE (Retrieval-based Unified Information Extraction), a framework that leverages in-context learning for efficient task generalization. RUIE introduces a novel demonstration selection mechanism combining LLM preferences with a keyword-enhanced reward model, and employs a bi-encoder retriever trained through contrastive learning and knowledge distillation. As the first trainable retrieval framework for UIE, RUIE serves as a universal plugin for various LLMs. Experimental results on eight held-out datasets demonstrate RUIE's effectiveness, with average F1-score improvements of 19.22 and 3.22 compared to instruction-tuning methods and other retrievers, respectively.
In the realm of Large Language Model (LLM) inference, the inherent structure of transformer models coupled with the multi-GPU tensor parallelism strategy leads to a sequential execution of computation and communication. This results in substantial underutilization of computing resources during the communication phase. To mitigate this inefficiency, various techniques have been developed to optimize the use of computational power throughout the communication process. These strategies primarily involve overlapping matrix computations and communications, as well as interleaving micro-batches across different requests. Nonetheless, these approaches either fall short of achieving ideal overlap or impose certain limitations on their application. To overcome these challenges, this paper introduces a novel strategy for computation-communication overlap that operates at the sequence level. This method not only enhances the degree of overlap but also minimizes the constraints on its applicability. Experimental evaluations conducted using 30b/70b models have demonstrated significant improvements in efficiency. Specifically, the proposed technique has been shown to reduce time consumption by approximately 35% on 4090 GPU and by roughly 15% on A800 GPU during the prefill stage of LLM inference.
Recent advancements in large language models (LLMs) and agent technologies offer promising solutions to the simulation of social science experiments, but the availability of data of real-world population required by many of them still poses as a major challenge. This paper explores a novel framework that leverages census data and LLMs to generate virtual populations, significantly reducing resource requirements and bypassing privacy compliance issues associated with real-world data, while keeping a statistical truthfulness. Drawing on real-world census data, our approach first generates a persona that reflects demographic characteristics of the population. We then employ LLMs to enrich these personas with intricate details, using techniques akin to those in image generative models but applied to textual data. Additionally, we propose a framework for the evaluation of the feasibility of our method with respect to capability of LLMs based on personality trait tests, specifically the Big Five model, which also enhances the depth and realism of the generated personas. Through preliminary experiments and analysis, we demonstrate that our method produces personas with variability essential for simulating diverse human behaviors in social science experiments. But the evaluation result shows that only weak sign of statistical truthfulness can be produced due to limited capability of current LLMs. Insights from our study also highlight the tension within LLMs between aligning with human values and reflecting real-world complexities. Thorough and rigorous test call for further research. Our codes are released at https://github.com/baiyuqi/agentic-society.git
Large language models (LLMs) have significantly advanced audio processing through audio codecs that convert audio into discrete tokens, enabling the application of language modeling techniques to audio data. However, audio codecs often operate at high frame rates, resulting in slow training and inference, especially for autoregressive models. To address this challenge, we present the Low Frame-rate Speech Codec (LFSC): a neural audio codec that leverages finite scalar quantization and adversarial training with large speech language models to achieve high-quality audio compression with a 1.89 kbps bitrate and 21.5 frames per second. We demonstrate that our novel codec can make the inference of LLM-based text-to-speech models around three times faster while improving intelligibility and producing quality comparable to previous models.
There has been a surge of interest in language model agents that can navigate virtual environments such as the web or desktop. To navigate such environments, agents benefit from information on the various elements (e.g., buttons, text, or images) present. It remains unclear which element attributes have the greatest impact on agent performance, especially in environments that only provide a graphical representation (i.e., pixels). Here we find that the ordering in which elements are presented to the language model is surprisingly impactful--randomizing element ordering in a webpage degrades agent performance comparably to removing all visible text from an agent's state representation. While a webpage provides a hierarchical ordering of elements, there is no such ordering when parsing elements directly from pixels. Moreover, as tasks become more challenging and models more sophisticated, our experiments suggest that the impact of ordering increases. Finding an effective ordering is non-trivial. We investigate the impact of various element ordering methods in web and desktop environments. We find that dimensionality reduction provides a viable ordering for pixel-only environments. We train a UI element detection model to derive elements from pixels and apply our findings to an agent benchmark--OmniACT--where we only have access to pixels. Our method completes more than two times as many tasks on average relative to the previous state-of-the-art.
The recent success of large language models (LLMs) and the scaling law has led to a widespread adoption of larger models. Particularly in the healthcare industry, there is an increasing demand for locally operated LLMs due to security concerns. However, the majority of high quality open-source LLMs have a size of 70B parameters, imposing significant financial burdens on users for GPU preparation and operation. To overcome these issues, we present a medical adaptation based on the recent 7B models, which enables the operation in low computational resources. We compare the performance on medical question-answering benchmarks in two languages (Japanese and English), demonstrating that its scores reach parity with or surpass those of currently existing medical LLMs that are ten times larger. We find that fine-tuning an English-centric base model on Japanese medical dataset improves the score in both language, supporting the effect of cross-lingual knowledge transfer. We hope that this study will alleviate financial challenges, serving as a stepping stone for clinical institutions to practically utilize LLMs locally. Our evaluation code is available at https://github.com/stardust-coder/japanese-lm-med-harness.
Large language model (LLM) role-playing has gained widespread attention. Authentic character knowledge is crucial for constructing realistic LLM role-playing agents. However, existing works usually overlook the exploration of LLMs' ability to detect characters' known knowledge errors (KKE) and unknown knowledge errors (UKE) while playing roles, which would lead to low-quality automatic construction of character trainable corpus. In this paper, we propose RoleKE-Bench to evaluate LLMs' ability to detect errors in KKE and UKE. The results indicate that even the latest LLMs struggle to detect these two types of errors effectively, especially when it comes to familiar knowledge. We experimented with various reasoning strategies and propose an agent-based reasoning method, Self-Recollection and Self-Doubt (S$^2$RD), to explore further the potential for improving error detection capabilities. Experiments show that our method effectively improves the LLMs' ability to detect error character knowledge, but it remains an issue that requires ongoing attention.
Recent proposals in recommender systems represent items with their textual description, using a large language model. They show better results on standard benchmarks compared to an item ID-only model, such as Bert4Rec. In this work, we revisit the often-used Bert4Rec baseline and show that with further tuning, Bert4Rec significantly outperforms previously reported numbers, and in some datasets, is competitive with state-of-the-art models. With revised baselines for item ID-only models, this paper also establishes new competitive results for architectures that combine IDs and textual descriptions. We demonstrate this with Flare (Fusing Language models and collaborative Architectures for Recommender Enhancement). Flare is a novel hybrid sequence recommender that integrates a language model with a collaborative filtering model using a Perceiver network. Prior studies focus evaluation on datasets with limited-corpus size, but many commercially-applicable recommender systems common on the web must handle larger corpora. We evaluate Flare on a more realistic dataset with a significantly larger item vocabulary, introducing new baselines for this setting. This paper also showcases Flare's inherent ability to support critiquing, enabling users to provide feedback and refine recommendations. We leverage critiquing as an evaluation method to assess the model's language understanding and its transferability to the recommendation task.
Neurosymbolic approaches can add robustness to opaque neural systems by incorporating explainable symbolic representations. However, previous approaches have not used formal logic to contextualize queries to and validate outputs of large language models (LLMs). We propose \systemname{}, a novel neurosymbolic framework, to improve the robustness and reliability of LLMs in question-answering tasks. We provide \systemname{} with a domain-specific knowledge base, a logical reasoning system, and an integration to an existing LLM. This framework has two capabilities (1) context gathering: generating explainable and relevant context for a given query, and (2) validation: confirming and validating the factual accuracy of a statement in accordance with a knowledge base (KB). Our work opens a new area of neurosymbolic generative AI text validation and user personalization.
With the advent of the big data and large language model era, zero-shot personalized rapid customization has emerged as a significant trend. In this report, we introduce Takin AudioLLM, a series of techniques and models, mainly including Takin TTS, Takin VC, and Takin Morphing, specifically designed for audiobook production. These models are capable of zero-shot speech production, generating high-quality speech that is nearly indistinguishable from real human speech and facilitating individuals to customize the speech content according to their own needs. Specifically, we first introduce Takin TTS, a neural codec language model that builds upon an enhanced neural speech codec and a multi-task training framework, capable of generating high-fidelity natural speech in a zero-shot way. For Takin VC, we advocate an effective content and timbre joint modeling approach to improve the speaker similarity, while advocating for a conditional flow matching based decoder to further enhance its naturalness and expressiveness. Last, we propose the Takin Morphing system with highly decoupled and advanced timbre and prosody modeling approaches, which enables individuals to customize speech production with their preferred timbre and prosody in a precise and controllable manner. Extensive experiments validate the effectiveness and robustness of our Takin AudioLLM series models. For detailed demos, please refer to https://everest-ai.github.io/takinaudiollm/.
Mixture-of-Experts (MoE) models scale more effectively than dense models due to sparse computation through expert routing, selectively activating only a small subset of expert modules. However, sparse computation challenges traditional training practices, as discrete expert routing hinders standard backpropagation and thus gradient-based optimization, which are the cornerstone of deep learning. To better pursue the scaling power of MoE, we introduce GRIN (GRadient-INformed MoE training), which incorporates sparse gradient estimation for expert routing and configures model parallelism to avoid token dropping. Applying GRIN to autoregressive language modeling, we develop a top-2 16$\times$3.8B MoE model. Our model, with only 6.6B activated parameters, outperforms a 7B dense model and matches the performance of a 14B dense model trained on the same data. Extensive evaluations across diverse tasks demonstrate the potential of GRIN to significantly enhance MoE efficacy, achieving 79.4 on MMLU, 83.7 on HellaSwag, 74.4 on HumanEval, and 58.9 on MATH.
Large Language Model can reasonably understand and generate human expressions but may lack of thorough thinking and reasoning mechanisms. Recently there have been several studies which enhance the thinking ability of language models but most of them are not data-driven or training-based. In this paper, we are motivated by the cognitive mechanism in the natural world, and design a novel model architecture called TaS which allows it to first consider the thoughts and then express the response based upon the query. We design several pipelines to annotate or generate the thought contents from prompt-response samples, then add language heads in a middle layer which behaves as the thinking layer. We train the language model by the thoughts-augmented data and successfully let the thinking layer automatically generate reasonable thoughts and finally output more reasonable responses. Both qualitative examples and quantitative results validate the effectiveness and performance of TaS. Our code is available at https://anonymous.4open.science/r/TadE.
LLM-as-a-Judge and reward models are widely used alternatives of multiple-choice questions or human annotators for large language model (LLM) evaluation. Their efficacy shines in evaluating long-form responses, serving a critical role as evaluators of leaderboards and as proxies to align LLMs via reinforcement learning. However, despite their popularity, their effectiveness in diverse contexts, such as non-English prompts, factual verification, or challenging questions, remains unexplored. In this paper, we conduct a comprehensive analysis of automated evaluators, reporting several key findings on their behavior. First, we discover that English evaluation capabilities significantly influence language-specific evaluation capabilities, often more than the language proficiency itself, enabling evaluators trained in English to easily transfer their skills to other languages. Second, we identify critical shortcomings, where LLMs fail to detect and penalize errors, such as factual inaccuracies, cultural misrepresentations, and the presence of unwanted language. Finally, we find that state-of-the-art evaluators struggle with challenging prompts, in either English or Korean, underscoring their limitations in assessing or generating complex reasoning questions. We release the dataset and codes used.
Aspect-based sentiment analysis (ABSA) involves identifying sentiment towards specific aspect terms in a sentence and allows us to uncover nuanced perspectives and attitudes on particular aspects of a product, service, or topic. However, the scarcity of labeled data poses a significant challenge to training high-quality models. To address this issue, we explore the potential of data augmentation using ChatGPT, a well-performing large language model (LLM), to enhance the sentiment classification performance towards aspect terms. Specifically, we explore three data augmentation strategies based on ChatGPT: context-focused, aspect-focused, and context-aspect data augmentation techniques. Context-focused data augmentation focuses on changing the word expression of context words in the sentence while keeping aspect terms unchanged. In contrast, aspect-focused data augmentation aims to change aspect terms but keep context words unchanged. Context-Aspect data augmentation integrates the above two data augmentations to generate augmented samples. Furthermore, we incorporate contrastive learning into the ABSA tasks to improve performance. Extensive experiments show that all three data augmentation techniques lead to performance improvements, with the context-aspect data augmentation strategy performing best and surpassing the performance of the baseline models.
Next-token prediction serves as the dominant component in current neural language models. During the training phase, the model employs teacher forcing, which predicts tokens based on all preceding ground truth tokens. However, this approach has been found to create shortcuts, utilizing the revealed prefix to spuriously fit future tokens, potentially compromising the accuracy of the next-token predictor. In this paper, we introduce Semformer, a novel method of training a Transformer language model that explicitly models the semantic planning of response. Specifically, we incorporate a sequence of planning tokens into the prefix, guiding the planning token representations to predict the latent semantic representations of the response, which are induced by an autoencoder. In a minimal planning task (i.e., graph path-finding), our model exhibits near-perfect performance and effectively mitigates shortcut learning, a feat that standard training methods and baseline models have been unable to accomplish. Furthermore, we pretrain Semformer from scratch with 125M parameters, demonstrating its efficacy through measures of perplexity, in-context learning, and fine-tuning on summarization tasks.
At the beginning of 2022, a simplistic word-guessing game took the world by storm and was further adapted to many languages beyond the original English version. In this paper, we examine the strategies of daily word-guessing game players that have evolved during a period of over two years. A survey gathered from 25% of frequent players reveals their strategies and motivations for continuing the daily journey. We also explore the capability of several popular open-access large language model systems and open-source models at comprehending and playing the game in two different languages. Results highlight the struggles of certain models to maintain correct guess length and generate repetitions, as well as hallucinations of non-existent words and inflections.
This paper proposes Attention-Seeker, an unsupervised keyphrase extraction method that leverages self-attention maps from a Large Language Model to estimate the importance of candidate phrases. Our approach identifies specific components - such as layers, heads, and attention vectors - where the model pays significant attention to the key topics of the text. The attention weights provided by these components are then used to score the candidate phrases. Unlike previous models that require manual tuning of parameters (e.g., selection of heads, prompts, hyperparameters), Attention-Seeker dynamically adapts to the input text without any manual adjustments, enhancing its practical applicability. We evaluate Attention-Seeker on four publicly available datasets: Inspec, SemEval2010, SemEval2017, and Krapivin. Our results demonstrate that, even without parameter tuning, Attention-Seeker outperforms most baseline models, achieving state-of-the-art performance on three out of four datasets, particularly excelling in extracting keyphrases from long documents.