Instruction-tuned Code Large Language Models (Code LLMs) have excelled in diverse code-related tasks, such as program synthesis, automatic program repair, and code explanation. To collect training datasets for instruction-tuning, a popular method involves having models autonomously generate instructions and corresponding responses. However, the direct generation of responses does not ensure functional correctness, a crucial requirement for generating responses to code instructions. To overcome this, we present Verification-Based Self-Play (VERSE), aiming to enhance model proficiency in generating correct responses. VERSE establishes a robust verification framework that covers various code instructions. Employing VERSE, Code LLMs engage in self-play to generate instructions and corresponding verifications. They evaluate execution results and self-consistency as verification outcomes, using them as scores to rank generated data for self-training. Experiments show that VERSE improves multiple base Code LLMs (average 7.6%) across various languages and tasks on many benchmarks, affirming its effectiveness.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Can LLMs consistently improve their previous outputs for better results? For this to be true, LLMs would need to be better at discriminating among previously-generated alternatives, than generating initial responses. We explore the validity of this hypothesis in practice. We first formulate a unified framework that allows us to compare the generative and discriminative capability of any model on any task. In our resulting experimental analysis of several open-source and industrial LLMs, we observe that model’s are not reliably better at discriminating among previously-generated alternatives than generating initial responses. This finding challenges the notion that LLMs may be able to enhance their performance only through their own judgment.
Large language models (LLMs) have shown remarkable capability in numerous tasks and applications. However, fine-tuning LLMs using high-quality datasets under external supervision remains prohibitively expensive. In response, LLM self-improvement approaches have been vibrantly developed recently. The typical paradigm of LLM self-improvement involves training LLM on self-generated data, part of which may be detrimental and should be filtered out due to the unstable data quality. While current works primarily employs filtering strategies based on answer correctness, in this paper, we demonstrate that filtering out correct but with high distribution shift extent (DSE) samples could also benefit the results of self-improvement. Given that the actual sample distribution is usually inaccessible, we propose a new metric called DS weight to approximate DSE, inspired by the Importance Weighting methods. Consequently, we integrate DS weight with self-consistency to comprehensively filter the self-generated samples and fine-tune the language model. Experiments show that with only a tiny valid set (up to 5% size of the training set) to compute DS weight, our approach can notably promote the reasoning ability of current LLM self-improvement methods. The resulting performance is on par with methods that rely on external supervision from pre-trained reward models.
SimRP: Syntactic and Semantic Similarity Retrieval Prompting Enhances Aspect Sentiment Quad Prediction
PDF ↗Aspect Sentiment Quad Prediction (ASQP) is the most complex subtask of Aspect-based Sentiment Analysis (ABSA), aiming to predict all sentiment quadruples within the given sentence. Due to the complexity of sentence syntaxes and the diversity of sentiment expressions, generative methods gradually become the mainstream approach in ASQP. However, existing generative models are constrained in the effectiveness of demonstrations. Semantically similar demonstrations help in judging sentiment categories and polarities but may confuse the model in recognizing aspect and opinion terms, which are more related to sentence syntaxes. To this end, we first develop Syn2Vec, a method for calculating syntactic vectors to support the retrieval of syntactically similar demonstrations. Then, we propose Syntactic and Semantic Similarity Retrieval Prompting (SimRP) to construct effective prompts by retrieving the most related demonstrations that are syntactically and semantically similar. With these related demonstrations, pre-trained generative models, especially Large Language Models (LLMs), can fully release their potential to recognize sentiment quadruples. Extensive experiments in Supervised Fine-Tuning (SFT) and In-context Learning (ICL) paradigms demonstrate the effectiveness of SimRP. Furthermore, we find that LLMs' capabilities in ASQP are severely underestimated by biased data annotations and the exact matching metric. We propose a novel constituent subtree-based fuzzy metric for more accurate and rational quadruple recognition.
Speech watermarking techniques can proactively mitigate the potential harmful consequences of instant voice cloning techniques. These techniques involve the insertion of signals into speech that are imperceptible to humans but can be detected by algorithms. Previous approaches typically embed watermark messages into continuous space. However, intuitively, embedding watermark information into robust discrete latent space can significantly improve the robustness of watermarking systems. In this paper, we propose DiscreteWM, a novel speech watermarking framework that injects watermarks into the discrete intermediate representations of speech. Specifically, we map speech into discrete latent space with a vector-quantized autoencoder and inject watermarks by changing the modular arithmetic relation of discrete IDs. To ensure the imperceptibility of watermarks, we also propose a manipulator model to select the candidate tokens for watermark embedding. Experimental results demonstrate that our framework achieves state-of-the-art performance in robustness and imperceptibility, simultaneously. Moreover, our flexible frame-wise approach can serve as an efficient solution for both voice cloning detection and information hiding. Additionally, DiscreteWM can encode 1 to 150 bits of watermark information within a 1-second speech clip, indicating its encoding capacity.
Text generation with citations makes it easy to verify the factuality of Large Language Models’ (LLMs) generations. Existing one-step generation studies expose distinct shortages in answer refinement and in-context demonstration matching. In light of these challenges, we propose R2-MGA, a Retrieval and Reflection Memory-augmented Generative Agent. Specifically, it first retrieves the memory bank to obtain the best-matched memory snippet, then reflects the retrieved snippet as a reasoning rationale, next combines the snippet and the rationale as the best-matched in-context demonstration. Additionally, it is capable of in-depth answer refinement with two specifically designed modules. We evaluate R2-MGA across five LLMs on the ALCE benchmark. The results reveal R2-MGA’ exceptional capabilities in text generation with citations. In particular, compared to the selected baselines, it delivers up to +58.8% and +154.7% relative performance gains on answer correctness and citation quality, respectively. Extensive analyses strongly support the motivations of R2-MGA.
Deep Submodular Optimization and LLM for Multimodal Content Extraction and Automatic Poster Generation from Long Document
PDF ↗A poster from a long input document can be considered as a one-page easy-to-read multimodal (text and images) summary presented on a nice template with good design elements. Automatic transformation of a long document into a poster is a very less studied but challenging task. It involves content summarization of the input document followed by template generation and harmonization. In this work, we propose a novel deep submodular function which can be trained on ground truth summaries to extract multimodal content from the document and explicitly ensures good coverage, diversity and alignment of text and images. Then, we use an LLM based paraphraser and propose to generate a template with various design aspects conditioned on the input content. We show the merits of our approach through extensive automated and human evaluations.
LLM+AL: Bridging Large Language Models and Action Languages for Complex Reasoning About Actions
PDF ↗Large Language Models (LLMs) have made significant strides in various intelligent tasks but still struggle with complex action reasoning tasks that require systematic search. To address this limitation, we propose a method that bridges the natural language understanding capabilities of LLMs with the symbolic reasoning strengths of action languages. Our approach, termed LLM+AL, leverages the LLM's strengths in semantic parsing and commonsense knowledge generation alongside the action language's proficiency in automated reasoning based on encoded knowledge. We compare LLM+AL against state-of-the-art LLMs, including ChatGPT-4, Claude 3 Opus, Gemini Ultra 1.0, and o1-preview, using benchmarks for complex reasoning about actions. Our findings indicate that, although all methods exhibit errors, LLM+AL, with relatively minimal human corrections, consistently leads to correct answers, whereas standalone LLMs fail to improve even with human feedback. LLM+AL also contributes to automated generation of action languages.
This paper presents VDAct, a dataset for a Video-grounded Dialogue on Event-driven Activities, alongside VDEval, a session-based context evaluation metric specially designed for the task. Unlike existing datasets, VDAct includes longer and more complex video sequences that depict a variety of event-driven activities that require advanced contextual understanding for accurate response generation. The dataset comprises 3,000 dialogues with over 30,000 question-and-answer pairs, derived from 1,000 videos with diverse activity scenarios. VDAct displays a notably challenging characteristic due to its broad spectrum of activity scenarios and wide range of question types. Empirical studies on state-of-the-art vision foundation models highlight their limitations in addressing certain question types on our dataset. Furthermore, VDEval, which integrates dialogue session history and video content summaries extracted from our supplementary Knowledge Graphs to evaluate individual responses, demonstrates a significantly higher correlation with human assessments on the VDAct dataset than existing evaluation metrics that rely solely on the context of single dialogue turns.
In various academic and professional settings, such as mathematics lectures or research presentations, it is often necessary to convey mathematical expressions orally. However, reading mathematical expressions aloud without accompanying visuals can significantly hinder comprehension, especially for those who are hearing-impaired or rely on subtitles due to language barriers. For instance, when a presenter reads Euler's Formula, current Automatic Speech Recognition (ASR) models often produce a verbose and error-prone textual description (e.g., e to the power of i x equals cosine of x plus i 'side' of x), instead of the concise LaTeX format, which hampers clear understanding and communication. To address this issue, we introduce MathSpeech, a novel pipeline that integrates ASR models with small Language Models (sLMs) to correct errors in mathematical expressions and accurately convert spoken expressions into structured LaTeX representations. Evaluated on a new dataset derived from lecture recordings, MathSpeech demonstrates LaTeX generation capabilities comparable to leading commercial Large Language Models (LLMs), while leveraging fine-tuned small language models of only 120M parameters. Specifically, in terms of CER, BLEU, and ROUGE scores for LaTeX translation, MathSpeech demonstrated significantly superior capabilities compared to GPT-4o. We observed a decrease in CER from 0.390 to 0.298, and higher ROUGE/BLEU scores compared to GPT-4o.
Enhancing Non-English Capabilities of English-Centric Large Language Models Through Deep Supervision Fine-Tuning
PDF ↗Large language models (LLMs) have demonstrated significant progress in multilingual language understanding and generation. However, due to the imbalance in training data, their capabilities in non-English languages are limited. Recent studies revealed the English-pivot multilingual mechanism of LLMs, where LLMs implicitly convert non-English queries into English ones at the bottom layers and adopt English for thinking at the middle layers. However, due to the absence of explicit supervision for cross-lingual alignment in the intermediate layers of LLMs, the internal representations during these stages may become inaccurate. In this work, we introduce a deep supervision fine-tuning method (DFT) that incorporates additional supervision in the internal layers of the model to guide its workflow. Specifically, we introduce two training objectives on different layers of LLMs: one at the bottom layers to constrain the conversion of the target language into English, and another at the middle layers to constrain reasoning in English. To effectively achieve the guiding purpose, we designed two types of supervision signals: logits and feature, which represent a stricter constraint and a relatively more relaxed guidance. Our method guides the model to not only consider the final generated result when processing non-English inputs but also ensure the accuracy of internal representations. We conducted extensive experiments on typical English-centric large models, LLaMA-2 and Gemma-2, and the results on multiple multilingual datasets show that our method significantly outperforms traditional fine-tuning methods.
Large language models have shown great potential in complex reasoning tasks, yet their performance is often hampered by the scarcity of high-quality and reasoning-focused training datasets. Addressing this challenge, we propose Key-PointDriven Data Synthesis (KPDDS), a novel data synthesis framework that synthesizes question-answer pairs by leveraging key points and exemplar practices from authentic data sources. KPDDS ensures the generation of novel questions with rigorous quality control and substantial scalability. As a result, we present KPMath, an extensive synthetic dataset tailored for mathematical reasoning, comprising over 800K questionanswer pairs. Utilizing KPMath and augmenting it with additional reasoning-intensive corpora, we create the comprehensive KPMath-Plus dataset. Our experiments demonstrate that this dataset can enhance the mathematical reasoning performance of models across various architectures and sizes. The Qwen1.5-72B model, fine-tuned on KPMath-Plus, achieves 87.0% accuracy on GSM8K and 58.3% on MATH, surpassing competitors in the 7B to 72B range and best commercial models like GPT-4 across multiple math reasoning datasets.
The remarkable success of Large Language Models (LLMs) relies heavily on their substantial scale, which poses significant challenges during model deployment in terms of latency and memory consumption. Recently, numerous studies have attempted to compress LLMs using one-shot pruning methods. However, these methods often suffer from considerable performance degradation on complex language understanding tasks, raising concerns about the feasibility of pruning in LLMs. To address this issue, we propose Adaptive Sparse Trainer (AST), a novel and efficient retraining framework tailored for semi-structured sparse models. AST enables models to learn optimal masks during the weight update process without incurring additional computational overhead. Furthermore, we demonstrate that incorporating knowledge distillation significantly improves retraining efficiency and enhances model performance under fixed computational constraints. Additionally, a supplementary set of well-initialized parameters is integrated to further augment the model's efficacy. AST achieves state-of-the-art performance with minimal training cost. When applied to the LLaMA2-7B model, AST reduces the perplexity and zero-shot accuracy gap between dense and 2:4 semi-structured sparse models to 0.6 and 1.16%, respectively, utilizing less than 0.4% of the pretraining tokens and GPU hours. Our work demonstrates the feasibility of deploying semi-structured sparse LLMs and offers a promising alternative for achieving highly compressed models when combined with existing quantization techniques.
GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition
PDF ↗Large language models (LLMs) demonstrate impressive performance on downstream tasks through in-context learning(ICL). However, there is a significant gap between their performance in Named Entity Recognition (NER) and in fine-tuning methods. We believe this discrepancy is due to inconsistencies in labeling definitions in NER. In addition, recent research indicates that LLMs do not learn the specific input-label mappings from the demonstrations. Therefore, we argue that using examples to implicitly capture the mapping between inputs and labels in in-context learning is not suitable for NER. Instead, it requires explicitly informing the model of the range of entities contained in the labels, such as annotation guidelines. In this paper, we propose GuideNER, which uses LLMs to summarize concise annotation guidelines as contextual information in ICL. We have conducted experiments on widely used NER datasets, and the experimental results indicate that our method can consistently and significantly outperform state-of-the-art methods, while using shorter prompts. Especially on the GENIA dataset, our model outperforms the previous state-of-the-art model by 12.63 F1 scores.
Self-learning of Large Language Models (LLMs) facilitates their advancement towards super-intelligence by training with self-synthesized experiences. However, a critical challenge is the amplification of hallucinations in generated data during iterative self-learning, underscoring the need for reliable data selection. To address this, we investigate the mechanism of Inner Knowledge Explicitation, which involves explicitly extracting the inner knowledge from memory of LLMs, to concurrently improves reasoning, and enables reliable self-learning data selection. This paper introduces a Self Knowledge Explicitation Learning (SKE-Learn) framework, which equips the LLMs with meta-skills to explicitly extract, verify and utilize inner knowledge for reasoning. By leveraging these meta-skills, SKE-Learn establishes a self-learning approach that ensures reliable selection of self-synthetic data. This approach enhances performance through iterative self-learning while mitigating the problem of hallucinations. Empirical results from six benchmarks demonstrate that Inner Knowledge Explicitation improves reasoning by serving as a more effective prompting method. Additionally, SKE-Learn, based on the verifiability of explicit knowledge, shows consistent performance improvements over multiple self-training iterations, with an average performance increase from 52.79% to 56.54% across all benchmarks. Furthermore, Inner Knowledge Explicitation provides explanation and intervention space during LLM's generation process.
To understand a document with multiple events, event-event relation extraction (ERE) emerges as a crucial task, aiming to discern how natural events temporally or structurally associate with each other. To achieve this goal, our work addresses the problems of temporal event relation extraction (TRE) and subevent relation extraction (SRE). The latest methods for such problems have commonly built document-level event graphs for global reasoning across sentences. However, the edges between events are usually derived from external tools heuristically, which are not always reliable and may introduce noise. Moreover, they are not capable of preserving logical constraints among event relations, e.g., coreference constraint, symmetry constraint and conjunction constraint. These constraints guarantee coherence between different relation types, enabling the generation of a unified event evolution graph. In this work, we propose a novel method named LogicERE, which performs high-order event relation reasoning through modeling logic constraints. Specifically, different from conventional event graphs, we design a logic constraint induced graph (LCG) without any external tools. LCG involves event nodes where the interactions among them can model the coreference constraint, and event pairs nodes where the interactions among them can retain the symmetry constraint and conjunction constraint. Then we perform high-order reasoning on LCG with relational graph transformer to obtain enhanced event and event pair embeddings. Finally, we further incorporate logic constraint information via a joint logic learning module. Extensive experiments demonstrate the effectiveness of the proposed method with state-of-the-art performance on benchmark datasets.
Few-shot Named Entity Recognition (NER) spotlights the tag of novel entity types in data-limited scenarios or lower-resource settings. Advances with Pre-trained Language Models (PLMs), including BERT, GPT, and their variants, have driven tremendous strategies to leverage context-dependent representations and exploit predefined relational cues, yielding significant gains in witnessing unseen entities. Nevertheless, a fundamental issue exists in prior efforts regarding their susceptibility to adversarial attacks in the intricate semantic environment. This vulnerability undermines the robustness of semantic representations, exacerbating the challenge of accurate entity identification, especially when transitioning across domains. To this end, we propose an Adversity-aware Augment Learning (AAL) solution for the few-shot NER task, dedicated to retrieving and reinforcing entity prototypes resilient to adversarial inference, thereby enhancing cross-domain semantic coherence. In particular, AAL employs a two-stage paradigm consisting of training and fine-tuning. The process initiates with augmentation learning by leveraging two kinds of prompt learning schemes, then identifies prototypes under the guidance of a variational manner. Furthermore, we devise a domain-oriented prototype refinement to optimize prototype learning under conditions of uncertainty attack, facilitating the effective transfer of common knowledge from source to target domains. The experimental results, encompassing the few-shot NER datasets under both certainty and uncertainty conditions, affirm the superiority of the proposed AAL over several representative baselines, particularly its capability against adversarial attacks.
Large Language Models (LLMs) have shown remarkable performance across various tasks, but the escalating demands on computational resources pose significant challenges, particularly in the extensive utilization of full fine-tuning for downstream tasks. To address this, parameter-efficient fine-tuning (PEFT) methods have been developed, but they often underperform compared to full fine-tuning and struggle with memory efficiency. In this work, we introduce Gradient Weight-Normalized Low-Rank Projection (GradNormLoRP), a novel approach that enhances both parameter and memory efficiency while maintaining comparable performance to full fine-tuning. GradNormLoRP normalizes the weight matrix to improve gradient conditioning, facilitating better convergence during optimization. Additionally, it applies low-rank approximations to the weight and gradient matrices, significantly reducing memory usage during training. Extensive experiments demonstrate that our 8-bit GradNormLoRP reduces optimizer memory usage by up to 89.5\% and enables the pre-training of large LLMs, such as LLaMA 7B, on consumer-level GPUs like the NVIDIA RTX 4090, without additional inference costs. Moreover, GradNormLoRP outperforms existing low-rank methods in fine-tuning tasks. For instance, when fine-tuning the RoBERTa model on all GLUE tasks with a rank of 8, GradNormLoRP achieves an average score of 80.65, surpassing LoRA's score of 79.23. These results underscore GradNormLoRP as a promising alternative for efficient LLM pre-training and fine-tuning.
Beyond Prompt Engineering: A Reinforced Token-Level Input Refinement for Large Language Models
PDF ↗In the rapidly developing field of automatic text generation and understanding, the quality of input data has been shown to be a key factor affecting the efficiency and accuracy of large language model (LLM) output. With the advent of advanced tools such as ChatGPT, input refinement work has mainly focused on prompt engineering. However, existing methods are often too dependent on specific contexts and are easily affected by individual expert experience and potential biases, limiting their wide applicability in diverse real-world applications. To address this problem, this study develops an Reinforced Token-Level Input Refinement, called RTLIR. We choose to optimize the input data at the fine-grained level of tokens, cleverly preserving the original text structure. Operationally, each state is defined by the token set of the current text, and each action is a binary decision process to decide whether to retain a specific token information. The agent automatically calculates and determines the selection probability of each token based on the current state, thereby optimizing the entire decision process. Through continuous exploration and learning, the agent can autonomously learn to identify the key inputs that have the greatest impact on the generation results and achieve refinement of the input data. In addition, RTLIR is a plug-and-play, LLM-agnostic module that can be used for a wide range of tasks and models. Experimental results show that RTLIR improves the performance of LLM in various input scenarios and tasks, with an average accuracy increase of 6%.
Mathematical reasoning ability objectively reflects a language model's understanding of implicit knowledge in contexts, with logic being a prerequisite for exploring, articulating and establishing effective reasoning. Large language models (LLMs) have shown great potential in complex reasoning tasks represented by mathematical reasoning. However, existing mathematical datasets either focus on commonsense reasoning, assessing the model's knowledge application ability, or arithmetic problems with fixed calculation rules, evaluating the model's rapid learning capability. There is a lack of datasets that require solving problems solely through logical reasoning. As a result, the performance of LLMs in accurately understanding the implicit logical relationships in problems and deriving conclusions based solely on given conditions is hindered. To address this challenge, we construct a dataset specifically for multiple step reasoning tasks: Reasoning-Math (RMath). This dataset focuses on evaluating logical reasoning abilities with mathematical reasoning problems, covering typical problem types, including direct reasoning problems, hypothetical reasoning problems, and nested reasoning problems. Additionally, we design a standardized annotation scheme that transforms natural language descriptions of conditions into formal propositions. Other annotation contents include problem categories, proposition truth values, and proposition relationship types. This not only reduces biases caused by semantic misunderstandings during problem-solving, but also facilitates the incorporation of theoretically grounded logical reasoning methods to enhance reasoning abilities. Furthermore, we propose a normalization problem-solving framework based on propositional logic for RMath and design the problem-solving process for prompt tuning to guide LLMs to absorb mathematical logical theories and improving reasoning abilities. Finally, we evaluate RMath on several popular LLMs and present the corresponding results.