Large language models (LLMs) excellently generate human-like text, but also raise concerns about misuse in fake news and academic dishonesty. Decoding-based watermark, particularly the watermark based on the GumbelMax trick (GM watermark), is a standout solution for safeguarding machine-generated texts due to its notable detectability. However, GM watermark encounters a major challenge with generation diversity, always yielding identical outputs for the same prompt, negatively impacting generation diversity and user experience. To overcome this limitation, we introduce a new type of GM watermark, the Logits-Addition watermark, as well as three variants that aim to enhance diversity, particularly the GumbelSoft watermark (i.e., the softmax variant of the Logits-Addition watermark). When assessed for detectability in high diversity settings, our Gumbelsoft demonstrates superior performance, with its AUROC score exceeding those of the two alternative variants by a margin of 0.1 to 0.3 and outperforming other decoding-based watermarking methods by a minimum of 0.1.
论文检索
输入标题、作者或关键词，从 1,915 篇学术成果中精准定位
With recent rapid growth of large language models (LLMs), discrete speech tokenization has played an important role for injecting speech into LLMs. However, this discretization gives rise to a loss of information, consequently impairing overall performance. To improve the performance of these discrete speech tokens, we present RepCodec, a novel speech representation codec for semantic speech tokenization. In contrast to audio codecs which reconstruct the raw audio, RepCodec learns a vector quantization codebook through reconstructing speech representations from speech encoders like HuBERT or data2vec. Together, the speech encoder, the codec encoder and the vector quantization codebook form a pipeline for converting speech waveforms into semantic tokens. The extensive experiments illustrate that RepCodec, by virtue of its enhanced information retention capacity, significantly outperforms the widely used k-means clustering approach in both speech understanding and generation. Furthermore, this superiority extends across various speech encoders and languages, affirming the robustness of RepCodec.We believe our method can facilitate large language modeling research on speech processing.
PRP-Graph: Pairwise Ranking Prompting to LLMs with Graph Aggregation for Effective Text Re-ranking
PDF ↗Pairwise Ranking Prompting (PRP) demonstrates impressive effectiveness in zero-shot document re-ranking tasks with large language models (LLMs). However, in the existing methods, PRP only outputs the same label for the comparison results of different confidence intervals without considering the uncertainty of pairwise comparison, which implies an underutilization of the generation probability information of LLMs. To bridge this gap, we propose PRP-Graph, a novel pairwise re-ranking approach, based on a refined scoring PRP unit that exploits the output probabilities of target labels to capture the degree of certainty of the comparison results. Specifically, the PRP-Graph consists of two stages, namely ranking graph construction and ranking graph aggregation. Extensive experiments conducted on the BEIR benchmark demonstrate the superiority of our approach over existing PRP-based methods. Comprehensive analysis reveals that the PRP-Graph displays strong robustness towards the initial ranking order and delivers exceptional re-ranking results with acceptable efficiency. Our code and data are available at https://github.com/Memelank/PRP-Graph.
Zero-shot dialogue state tracking (DST) seeks to enable dialogue systems to transition to unfamiliar domains without manual annotation or extensive retraining. Prior research has approached this objective by embedding prompts into language models (LMs). Common methodologies include integrating prompts at the input layer or introducing learnable variables at each transformer layer. Nonetheless, each strategy exhibits inherent limitations. Prompts integrated at the input layer risk underutilization, with their impact potentially diminishing across successive transformer layers. Conversely, the addition of learnable variables to each layer can complicate the training process and increase inference latency. To tackle the issues mentioned above, this paper proposes Dual Low-Rank Adaptation (DualLoRA), a plug-and-play architecture designed for zero-shot DST. DualLoRA incorporates two distinct Low-Rank Adaptation (LoRA) components, targeting both dialogue context processing and prompt optimization, to ensure the comprehensive influence of prompts throughout the transformer model layers. This is achieved without incurring additional inference latency, showcasing an efficient integration into existing architectures. Through rigorous evaluation on the MultiWOZ and SGD datasets, DualLoRA demonstrates notable improvements across multiple domains, outperforming traditional baseline methods in zero-shot settings.
Large language models are first pre-trained on trillions of tokens and then instruction-tuned or aligned to specific preferences. While pre-training remains out of reach for most researchers due to the compute required, fine-tuning has become affordable thanks to parameter-efficient methods such as LoRA and QLoRA. Alignment is known to be sensitive to the many factors involved, including the quantity and quality of data, the alignment method, and the adapter rank. However, there has not yet been an extensive study of their effect on downstream performance. To address this gap, we conduct an in-depth investigation of the impact of popular choices for three crucial axes: (i) the alignment dataset (HH-RLHF and BeaverTails), (ii) the alignment technique (SFT and DPO), and (iii) the model (LLaMA-1, Vicuna-v1.3, Mistral-7b, and Mistral-7b-Instruct). Our extensive setup spanning over 300 experiments reveals consistent trends and unexpected findings. We observe how more informative data helps with preference alignment, cases where supervised fine-tuning outperforms preference optimization, and how aligning to a distinct preference boosts performance on downstream tasks. Through our in-depth analyses, we put forward key guidelines to help researchers perform more effective parameter-efficient LLM alignment.
Forgetting before Learning: Utilizing Parametric Arithmetic for Knowledge Updating in Large Language Models
PDF ↗Recent advancements in Large Language Models (LLMs) have showcased their remarkable capabilities in text understanding and generation. However, even stronger LLMs are susceptible to acquiring erroneous or obsolete information from the training corpus. Direct secondary fine-tuning with data containing new knowledge may be ineffective in updating knowledge due to the conflict between old and new knowledge. In this paper, we propose a new paradigm for fine-tuning called F-Learning (Forgetting before Learning), which employs parametric arithmetic to facilitate the forgetting of old knowledge and learning of new knowledge. Experimental results on two publicly available datasets demonstrate that our proposed F-Learning can obviously improve the knowledge updating performance of both full fine-tuning and LoRA fine-tuning, simultaneously outperforming the existing baselines in most cases. Moreover, we have also discovered that forgetting old knowledge by subtracting the parameters of LoRA can yield a similar effect to subtracting the parameters of full fine-tuning, and occasionally even surpass it significantly.
Large language models (LLMs) demonstrate remarkable multilingual capabilities without being pre-trained on specially curated multilingual parallel corpora.It remains a challenging problem to explain the underlying mechanisms by which LLMs process multilingual texts.In this paper, we delve into the composition of Transformer architectures in LLMs to pinpoint language-specific regions.Specially, we propose a novel detection method, language activation probability entropy (LAPE), to identify language-specific neurons within LLMs.Based on LAPE, we conduct comprehensive experiments on several representative LLMs, such as LLaMA-2, BLOOM, and Mistral. Our findings indicate that LLMs’ proficiency in processing a particular language is predominantly due to a small subset of neurons, primarily situated in the models’ top and bottom layers.Furthermore, we showcase the feasibility to “steer” the output language of LLMs by selectively activating or deactivating language-specific neurons. Our research provides important evidence to the understanding and exploration of the multilingual capabilities of LLMs.
In the era of data-driven decision-making, the complexity of data analysis necessitates advanced expertise and tools of data science, presenting significant challenges even for specialists. Large Language Models (LLMs) have emerged as promising aids as data science agents, assisting humans in data analysis and processing. Yet their practical efficacy remains constrained by the varied demands of real-world applications and complicated analytical process. In this paper, we introduce DSEval – a novel evaluation paradigm, as well as a series of innovative benchmarks tailored for assessing the performance of these agents throughout the entire data science lifecycle. Incorporating a novel bootstrapped annotation method, we streamline dataset preparation, improve the evaluation coverage, and expand benchmarking comprehensiveness. Our findings uncover prevalent obstacles and provide critical insights to inform future advancements in the field.
The recent success of large language models (LLMs) has attracted widespread interest to develop role-playing conversational agents personalized to the characteristics and styles of different speakers to enhance their abilities to perform both general and special purpose dialogue tasks. However, the ability to personalize the generated utterances to speakers, whether conducted by human or LLM, has not been well studied. To bridge this gap, our study introduces a novel evaluation challenge: speaker verification in agent-generated conversations, which aimed to verify whether two sets of utterances originate from the same speaker. To this end, we assemble a large dataset collection encompassing thousands of speakers and their utterances. We also develop and evaluate speaker verification models under experiment setups. We further utilize the speaker verification models to evaluate the personalization abilities of LLM-based role-playing models. Comprehensive experiments suggest that the current role-playing models fail in accurately mimicking speakers, primarily due to their inherent linguistic characteristics.
New intent discovery (NID) is an important problem for deploying practical dialogue systems, which trains intent classifiers on a semi-supervised corpus where unlabeled user utterances contain both known and novel intents. Most existing NID algorithms place hope on the sample similarity to cluster unlabeled corpus to known or new samples. Lacking supervision on new intents, we experimentally find the intent classifier fails to fully distinguish new intents since they tend to assemble into intertwined centers.To address this problem, we propose a novel GeoID framework that learns geometry-aware representations to maximally separate all intents. Specifically, we are motivated by the recent findings on Neural Collapse (NC) in classification tasks to derive optimal intent center structure. Meanwhile, we devise a dual pseudo-labeling strategy based on optimal transport assignments and semi-supervised clustering, ensuring proper utterances-to-center arrangement.Extensive results show that our GeoID method establishes a new state-of-the-art performance, achieving a +3.49% average accuracy improvement on three standardized benchmarking datasets. We also verify its usefulness in assisting large language models for improved in-context performance.
Recent advancements in large language models (LLMs) have brought significant changes to various domains, especially through LLM-driven autonomous agents. A representative scenario is in software development, where LLM agents demonstrate efficient collaboration, task division, and assurance of software quality, markedly reducing the need for manual involvement. However, these agents frequently perform a variety of tasks independently, without benefiting from past experiences, which leads to repeated mistakes and inefficient attempts in multi-step task execution. To this end, we introduce Experiential Co-Learning, a novel LLM-agent learning framework in which instructor and assistant agents gather shortcut-oriented experiences from their historical trajectories and use these past experiences for future task execution. The extensive experiments demonstrate that the framework enables agents to tackle unseen software-developing tasks more effectively. We anticipate that our insights will guide LLM agents towards enhanced autonomy and contribute to their evolutionary growth in cooperative learning. The code and data are available at https://github.com/OpenBMB/ChatDev.
Large language models (LLMs) are typically prompted to follow a single instruction per inference call. In this work, we analyze whether LLMs also hold the capability to handle multiple instructions simultaneously, denoted as Multi-Task Inference. For this purpose, we introduce the MTI Bench (Multi-Task Inference Benchmark), a comprehensive evaluation benchmark encompassing 5,000 instances across 25 tasks. Each task in the MTI Bench involves 2 to 3 sub-tasks. As expected, we first demonstrate that Multi-Task Inference reduces the total inference time by \times 1.46 times in average since it does not require multiple inference calls. Interestingly, contrary to the expectation that LLMs would perform better when tasks are divided, we find that state-of-the-art LLMs, such as Llama-2-Chat-70B and GPT-4, show up to 7.3% and 12.4% improved performance with Multi-Task Inference compared to Single-Task Inference on the MTI Bench. We release the MTI Bench dataset and our code at this [link](https://anonymous.4open.science/r/MTI-Bench-6F01).
As large language models (LLMs) become increasingly integrated into real-world applications such as code generation and chatbot assistance, extensive efforts have been made to align LLM behavior with human values, including safety. Jailbreak attacks, which aim to provoke unintended and unsafe behaviors from LLMs, remain a significant LLM safety threat. We analyze tokens, which are the smallest unit of text that can be processed by LLMs and make the following observations: (1) probabilities of tokens representing harmful responses are higher than those of harmless responses, and (2) responses containing safety disclaimers appear among the top tokens when token probabilities are sorted in descending order. In this paper, we leverage (1) and (2) to develop SafeDecoding, a safety-aware decoding strategy for LLMs, to defend against jailbreak attacks. We perform extensive experiments to evaluate SafeDecoding against six SOTA jailbreak attacks (GCG, AutoDAN, PAIR, DeepInception, SAP30, and template based attack) on five LLMs (Vicuna, Llama2, Guanaco, falcon, and Dolphin) using four benchmark datasets (AdvBench, HEx-PHI, MT-Bench, and Just-Eval). Our results show that SafeDecoding significantly reduces attack success rate and harmfulness of jailbreak attacks without compromising the helpfulness of responses to benign user queries while outperforming six defense methods (Perpelexity, Paraphrase, Retokenization, Self-Reminder, ICD, and Self-Examination).
While LLMs can provide reasoned explanations along with their answers, the nature and quality of those explanations are still poorly understood. In response, our goal is to define a detailed way of characterizing the explanation capabilities of modern models and to create a nuanced, interpretable explanation evaluation tool that can generate such characterizations automatically, without relying on expensive API calls or human annotations. Our approach is to (a) define the new task of explanation critiquing - identifying and categorizing any main flaw in an explanation and providing suggestions to address the flaw, (b) create a sizeable, human-verified dataset for this task, and (c) train an open-source, automatic critique model (called Digital Socrates) using this data. Through quantitative and qualitative analysis, we demonstrate how Digital Socrates is useful for revealing insights about student models by examining their reasoning chains, and how it can provide high-quality, nuanced, automatic evaluation of those model explanations for the first time. Digital Socrates thus fills an important gap in evaluation tools for understanding and improving the explanation behavior of models.
CodeScope: An Execution-based Multilingual Multitask Multidimensional Benchmark for Evaluating LLMs on Code Understanding and Generation
PDF ↗Large Language Models (LLMs) have demonstrated remarkable performance on assisting humans in programming and facilitating programming automation. However, existing benchmarks for evaluating the code understanding and generation capacities of LLMs suffer from severe limitations. First, most benchmarks are insufficient as they focus on a narrow range of popular programming languages and specific tasks, whereas real-world software development scenarios show a critical need to implement systems with multilingual and multitask programming environments to satisfy diverse requirements. Second, most benchmarks fail to consider the actual executability and the consistency of execution results of the generated code. To bridge these gaps between existing benchmarks and expectations from practical applications, we introduce **CodeScope**, an execution-based, multilingual, multitask, multidimensional evaluation benchmark for comprehensively measuring LLM capabilities on coding tasks. CodeScope covers **43 programming languages** and **eight coding tasks**. It evaluates the coding performance of LLMs from three dimensions (perspectives): **length**, **difficulty**, and **efficiency**. To facilitate execution-based evaluations of code generation, we develop **MultiCodeEngine**, an automated code execution engine that supports 14 programming languages. Finally, we systematically evaluate and analyze eight mainstream LLMs and demonstrate the superior breadth and challenges of CodeScope for evaluating LLMs on code understanding and generation tasks compared to other benchmarks. The CodeScope benchmark and code are publicly available at https://github.com/WeixiangYAN/CodeScope.
Recent work shows that in-context learning and optimization of in-context examples (ICE) can significantly improve the accuracy of large language models (LLMs) on a wide range of tasks, leading to an apparent consensus that ICE optimization is crucial for better performance. However, most of these studies assume a fixed or no instruction provided in the prompt. We challenge this consensus by investigating the necessity of optimizing ICE when task-specific instructions are provided and find that there are many tasks for which it yields diminishing returns. In particular, using a diverse set of tasks and a systematically created instruction set with gradually added details, we find that as the prompt instruction becomes more detailed, the returns on ICE optimization diminish. To characterize this behavior, we introduce a task-specific metric called Normalized Invariability to Choice of Examples (NICE) that quantifies the learnability of tasks from a given instruction, and provides a heuristic to help decide whether to optimize instructions or ICE for a new task. Given a task, the proposed metric can reliably predict the utility of optimizing ICE compared to using random ICE. Our code is available at [https://github.com/microsoft/nice-icl](https://github.com/microsoft/nice-icl).
Large Language Models (LLMs) have achieved unprecedented performance in Natural Language Generation (NLG) tasks. However, many existing studies have shown that they could be misused to generate undesired content. In response, before releasing LLMs for public access, model developers usually align those language models through Supervised Fine-Tuning (SFT) or Reinforcement Learning with Human Feedback (RLHF). Consequently, those aligned large language models refuse to generate undesired content when facing potentially harmful/unethical requests. A natural question is “could alignment really prevent those open-sourced large language models from being misused to generate undesired content?”. In this work, we provide a negative answer to this question. In particular, we show those open-sourced, aligned large language models could be easily misguided to generate undesired content without heavy computations or careful prompt designs. Our key idea is to directly manipulate the generation process of open-sourced LLMs to misguide it to generate undesired content including harmful or biased information and even private data. We evaluate our method on 4 open-sourced LLMs accessible publicly and our finding highlights the need for more advanced mitigation strategies for open-sourced LLMs.
With the advent of large language models (LLM), the line between human-crafted and machine-generated texts has become increasingly blurred. This paper delves into the inquiry of identifying discernible and unique linguistic properties in texts that were written by humans, particularly uncovering the underlying discourse structures of texts beyond their surface structures. Introducing a novel methodology, we leverage hierarchical parse trees and recursive hypergraphs to unveil distinctive discourse patterns in texts produced by both LLMs and humans. Empirical findings demonstrate that, although both LLMs and humans generate distinct discourse patterns influenced by specific domains, human-written texts exhibit more structural variability, reflecting the nuanced nature of human writing in different domains. Notably, incorporating hierarchical discourse features enhances binary classifiers’ overall performance in distinguishing between human-written and machine-generated texts, even on out-of-distribution and paraphrased samples. This underscores the significance of incorporating hierarchical discourse features in the analysis of text patterns. The code and dataset will be available at [TBA].
This work explores whether language models encode meaningfully grounded representations of sounds of objects. We learn a linear probe that retrieves the correct text representation of an object given a snippet of audio related to that object, where the sound representation is given by a pretrained audio model. This probe is trained via a contrastive loss that pushes the language representations and sound representations of an object to be close to one another. After training, the probe is tested on its ability to generalize to objects that were not seen during training. Across different language models and audio models, we find that the probe generalization is above chance in many cases, indicating that despite being trained only on raw text, language models encode grounded knowledge of sounds for some objects.
In order for large language model (LLM)-based assistants to effectively adapt to evolving information needs, it must be possible to update their factual knowledge through continued training on new data. The standard recipe for doing so involves continued pre-training on new documents followed by instruction-tuning on question-answer (QA) pairs. However, we find that LLMs trained with this recipe struggle to answer questions, even though the perplexity of documents is minimized. We found that QA pairs are generally straightforward, while documents are more complex, weaving many factual statements together in an intricate manner. Therefore, we hypothesize that it is beneficial to expose LLMs to QA pairs before continued pre-training on documents so that the process of encoding knowledge from complex documents takes into account how this knowledge is accessed through questions. Based on this, we propose pre-instruction-tuning (PIT), a method that instruction-tunes on questions prior to training on documents. This contrasts with standard instruction-tuning, which learns how to extract knowledge after training on documents. Extensive experiments and ablation studies demonstrate that pre-instruction-tuning significantly enhances the ability of LLMs to absorb knowledge from new documents, outperforming standard instruction-tuning by 17.8%.