Large language models (LMs) are capable of in-context learning from a few demonstrations (example-label pairs) to solve new tasks during inference. Despite the intuitive importance of high-quality demonstrations, previous work has observed that, in some settings, ICL performance is minimally affected by irrelevant labels (Min et al., 2022). We hypothesize that LMs perform ICL with irrelevant labels via two sequential processes: an inference function that solves the task, followed by a verbalization function that maps the inferred answer to the label space. Importantly, we hypothesize that the inference function is invariant to remappings of the label space (e.g., “true”/“false” to “cat”/“dog”), enabling LMs to share the same inference function across settings with different label words. We empirically validate this hypothesis with controlled layer-wise interchange intervention experiments. Our findings confirm the hypotheses on multiple datasets and tasks (natural language inference, sentiment analysis, and topic classification) and further suggest that the two functions can be localized in specific layers across various open-sourced models, including GEMMA-7B, MISTRAL-7B-V0.3, GEMMA-2-27B, and LLAMA-3.1-70B.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Adapting pre-trained language models (PLMs) for cross-task generalization is a crucial research area within the field of NLP. While fine-tuning and in-context learning are effective approaches for adapting LMs to emerging tasks, they can be costly and inefficient. Recently, some researchers have focused on achieving efficient task adaptation via hypernetwork, which is a meta network that generates task-specific weights based on task-oriented information without any optimization. However, the training of hypernetworks often lacks stability since the optimization signal is not straightforward, and the task information is not adequately representative. Moreover, previous works train hypenetworks with the general corpus, which is struggling with few-shot adaptation. To address these issues, we introduce HyperLoRA, a hypernetwork for LoRA parameters generation involving hypernetwork pre-training on instruction-following data and generalization fine-tuning on sparse task data. Furthermore, we utilize a constrained training loss and a gradient-based demonstration selection strategy to enhance the training stability and performance. Experimental results and analysis across four benchmark datasets (P3, S-NI, BBH, and SuperGLUE) demonstrate the proposed approach has flexible generalization ability and superior performance.
Large language models have demonstrated significant potential as the next-generation information access engines. However, their reliability is hindered by issues of hallucination and generating non-factual content. This is particularly problematic in long-form responses, where assessing and ensuring factual accuracy is complex. In this paper, we address this gap by proposing FactAlign, a novel alignment framework designed to enhance the factuality of LLMs’ long-form responses while maintaining their helpfulness. We introduce fKTO, a fine-grained, sentence-level alignment algorithm that extends the Kahneman-Tversky Optimization (KTO) alignment method. Leveraging recent advances in automatic factuality evaluation, FactAlign utilizes fine-grained factuality assessments to guide the alignment process. Our experiments on open-domain prompts and information-seeking questions demonstrate that FactAlign significantly improves the factual accuracy of LLM responses while also improving their helpfulness. Further analyses identify that FactAlign is capable of training LLMs to provide more information without losing factual precision, thus improving the factual F1 score. Our source code, datasets, and trained models are publicly available at https://github.com/MiuLab/FactAlign
Hop, skip, jump to Convergence: Dynamics of Learning Rate Transitions for Improved Training of Large Language Models
PDF ↗Various types of learning rate (LR) schedulers are being used for training or fine tuning of Large Language Models today. In practice, several mid-flight changes are required in the LR schedule either manually, or with careful choices around warmup steps, peak LR, type of decay and restarts. To study this further, we consider the effect of switching the learning rate at a predetermined time during training, which we refer to as “SkipLR”. We model SGD as a stochastic gradient flow and show that when starting from the same initial parameters, switching the learning rate causes the loss curves to contract towards each other. We demonstrate this theoretically for some simple cases, and empirically on large language models. Our analysis provides insight into how learning rate schedules affect the training dynamics, and could inform the design of new schedules to accelerate convergence.
Remember This Event That Year? Assessing Temporal Information and Understanding in Large Language Models
PDF ↗Large Language Models (LLMs) are increasingly ubiquitous, yet their ability to retain and reason about temporal information remains limited, hindering their application in real-world scenarios where understanding the sequential nature of events is crucial. Our study experiments with 12 state-of-the-art models (ranging from 2B to 70B+ parameters) on a novel numerical-temporal dataset, TempUN, spanning from 10,000 BCE to 2100 CE, to uncover significant temporal retention and comprehension limitations. We propose six metrics to assess three learning paradigms to enhance temporal knowledge acquisition. Our findings reveal that open-source models exhibit knowledge gaps more frequently, suggesting a trade-off between limited knowledge and incorrect responses. Additionally, various fine-tuning approaches significantly improved performance, reducing incorrect outputs and impacting the identification of ‘information not available’ in the generations. The associated dataset and code are available at the [URL](https://anonymous.4open.science/r/TempUN-ARR/).
Individuals engaging in online communication frequently express personal opinions with informal styles (e.g., memes and emojis). While Language Models (LMs) with informal communications have been widely discussed, a unique and emphatic style, the Repetitive Lengthening Form (RLF), has been overlooked for years. In this paper, we explore answers to two research questions: 1) Is RLF important for SA? 2) Can LMs understand RLF? Inspired by previous linguistic research, we curate **Lengthening**, the first multi-domain dataset with 850k samples focused on RLF for sentiment analysis. Moreover, we introduce **Explnstruct**, a two-stage Explainable Instruction Tuning framework aimed at improving both the performance and explainability of LLMs for RLF. We further propose a novel unified approach to quantify LMs’ understanding of informal expressions. We show that RLF sentences are expressive expressions and can serve as signatures of document-level sentiment. Additionally, RLF has potential value for online content analysis. Our comprehensive results show that fine-tuned Pre-trained Language Models (PLMs) can surpass zero-shot GPT-4 in performance but not in explanation for RLF. Finally, we show ExpInstruct can improve the open-sourced LLMs to match zero-shot GPT-4 in performance and explainability for RLF with limited samples. Code and sample data are available at https://github.com/Tom-Owl/OverlookedRLF
Existing speech-to-speech translation (S2ST) models fall into two camps: they either leverage text as an intermediate step or require hundreds of hours of parallel speech data. Both approaches are incompatible with textless languages or language pairs with limited parallel data. We present PFB, a framework for training textless S2ST models that require just dozens of hours of parallel speech data. We first pretrain a model on large-scale monolingual speech data, finetune it with a small amount of parallel speech data (20-60 hours), and lastly train with an unsupervised backtranslation objective. We train and evaluate our models for English-to-German, German-to-English and Marathi-to-English translation on three different domains (European Parliament, Common Voice, and All India Radio) with single-speaker synthesized speech. Evaluated using the ASR-BLEU metric, our models achieve reasonable performance on all three domains, with some being within 1-2 points of our higher-resourced topline.
AI systems’ ability to interpret human emotions and adapt to variations is becoming more crucial as AI gets embedded into everyone’s daily lives. Emotion Recognition in Conversations (ERC) is based on this fundamental challenge. Current state-of-the-art technologies in ERC are limited due to the need for future information. We introduce High-Dimensional Temporal Fusion Transformer (HiTFT), a time-series forecasting transformer that predicts pseudo-future information to overcome this constraint. This retains the models’ dynamic nature and provides future information more efficiently than other methods. Our proposed method combines pseudo future embeddings with an encoder that models the speaker’s emotional state using past and pseudo-future information as well as inter and intra speaker interactions; these speaker states are then passed through a decoder block that predicts the inferred emotion of that utterance. We further evaluate our method and show that it achieves state of the art performance on three ERC datasets - MELD, EmoryNLP, and IEMOCap.
Large Language Models (LLMs) trained on web-scale text corpora have been shown to capture world knowledge in their parameters. However, the mechanism by which language models store different types of knowledge is poorly understood. In this work, we examine two types of knowledge relating to temporally sensitive entities and demonstrate that each type is localized to different sets of parameters within the LLMs. We hypothesize that the lack of consideration of the locality of knowledge in existing continual learning methods contributes to both: the failed uptake of new information, and catastrophic forgetting of previously learned information. We observe that sequences containing references to updated and newly mentioned entities exhibit larger gradient norms in a subset of layers. We demonstrate that targeting parameter updates to these relevant layers can improve the performance of continually pretraining on language containing temporal drift.
Structured Chain-of-Thought Prompting for Few-Shot Generation of Content-Grounded QA Conversations
PDF ↗We introduce a structured chain-of-thought (SCoT) prompting approach to generating content-grounded multi-turn question-answer conversations with a pre-trained large language model (LLM). At the core of our proposal is a structured breakdown of the complex task into a number of states in a state machine, so that actions corresponding to various subtasks, e.g., content reading and utterance generation, can be executed in their own dedicated states. Each state leverages a unique set of resources, including prompts and (optionally) additional tools, to augment the generation process. Automatic evaluation shows that SCoT prompting with designated states for hallucination mitigation can increase agent faithfulness to grounding documents by up to 16.8%. When used as training data, our open-domain conversations synthesized from only 6 Wikipedia-based seed demonstrations train strong conversational QA agents. In out-of-domain evaluation, for example, we observe improvements of up to 13.9% in F1-score against ground truth over target domain gold data when the latter is augmented with our generated examples.
Sign language translation from video to spoken text presents unique challenges owing to the distinct grammar, expression nuances, and high variation of visual appearance across different speakers and contexts. Gloss annotations serve as an intermediary to guide the translation process. In our work, we focus on Gloss2Text translation stage and propose several advances by leveraging pre-trained large language models (LLMs), data augmentation, and novel label-smoothing loss function exploiting gloss translation ambiguities improving significantly the performance of state-of-the-art approaches. Through extensive experiments and ablation studies on the PHOENIX Weather 2014T dataset, our approach surpasses state-of-the-art performance in Gloss2Text translation, indicating its efficacy in addressing sign language translation and suggesting promising avenues for future research and development.
As large language models achieve impressive scores on traditional benchmarks, an increasing number of researchers are becoming concerned about benchmark data leakage during pre-training, commonly known as the data contamination problem. To ensure fair evaluation, recent benchmarks release only the training and validation sets, keeping the test set labels closed-source. They require anyone wishing to evaluate his language model to submit the model’s predictions for centralized processing and then publish the model’s result on their leaderboard. However, this submission process is inefficient and prevents effective error analysis. To address this issue, we propose to variabilize benchmarks and evaluate language models dynamically. Specifically, we extract variables from each test case and define a value range for each variable. For each evaluation, we sample new values from these value ranges to create unique test cases, thus ensuring a fresh evaluation each time. We applied this variable perturbation method to four datasets: GSM8K, ARC, CommonsenseQA, and TruthfulQA, which cover mathematical generation and multiple-choice tasks. Our experimental results demonstrate that this approach provides a more accurate assessment of the true capabilities of language models, effectively mitigating the contamination problem.
Conversational systems must be robust to user interactions that naturally exhibit diverse conversational traits. Capturing and simulating these diverse traits coherently and efficiently presents a complex challenge. This paper introduces Multi-Trait Adaptive Decoding (mTAD), a method that generates diverse user profiles at decoding-time by sampling from various trait-specific Language Models (LMs). mTAD provides an adaptive and scalable approach to user simulation, enabling the creation of multiple user profiles without the need for additional fine-tuning. By analyzing real-world dialogues from the Conversational Task Assistant (CTA) domain, we identify key conversational traits and developed a framework to generate profile-aware dialogues that enhance conversational diversity. Experimental results validate the effectiveness of our approach in modeling single-traits using specialized LMs, which can capture less common patterns, even in out-of-domain tasks. Furthermore, the results demonstrate that mTAD is a robust and flexible framework for combining diverse user simulators.
Sequential sentence classification (SSC) in scientific publications is crucial for supporting downstream tasks such as fine-grained information retrieval and extractive summarization. However, current SSC methods are constrained by model size, sequence length, and single-label setting. To address these limitations, this paper proposes LLM-SSC, a large language model (LLM)-based framework for both single- and multi-label SSC tasks. Unlike previous approaches that employ small- or medium-sized language models, the proposed framework utilizes LLMs to generate SSC labels through designed prompts, which enhance task understanding by incorporating demonstrations and a query to describe the prediction target. We also present a multi-label contrastive learning loss with auto-weighting scheme, enabling the multi-label classification task. To support our multi-label SSC analysis, we introduce and release a new dataset, biorc800, which mainly contains unstructured abstracts in the biomedical domain with manual annotations. Experiments demonstrate LLM-SSC’s strong performance in SSC under both in-context learning and task-specific tuning settings. We release biorc800 and our code at: https://github.com/ScienceNLP-Lab/LLM-SSC.
BLASER 2.0: a metric for evaluation and quality estimation of massively multilingual speech and text translation
PDF ↗We present BLASER 2.0, an automatic metric of machine translation quality which supports both speech and text modalities. Compared to its predecessor BLASER (Chen et al., 2023), BLASER 2.0 is based on better underlying text and speech representations that cover 202 text languages and 57 speech ones and extends the training data. BLASER 2.0 comes in two varieties: a reference-based and a reference-free (quality estimation) model. We demonstrate that the reference-free version is applicable not only at the dataset level, for evaluating the overall model performance, but also at the sentence level, for scoring individual translations. In particular, we show its applicability for detecting translation hallucinations and filtering training datasets to obtain more reliable translation models. The BLASER 2.0 models are publicly available at https://github.com/facebookresearch/sonar.
Productive interactions between diverse users and language technologies require outputs from the latter to be culturally relevant and sensitive. Prior works have evaluated models’ knowledge of cultural norms, values, and artefacts, without considering how this knowledge manifests in downstream applications. In this work, we focus on extrinsic evaluation of cultural competence in two text generation tasks, open-ended question answering and story generation. We quantitatively and qualitatively evaluate model outputs when an explicit cue of culture, specifically nationality, is perturbed in the prompts. Although we find that model outputs do vary when varying nationalities and feature culturally relevant words, we also find weak correlations between text similarity of outputs for different countries and the cultural values of these countries. Finally, we discuss important considerations in designing comprehensive evaluation of cultural competence in user-facing tasks.
Writing has long been considered a hallmark of human intelligence and remains a pinnacle task for artificial intelligence (AI) due to the intricate cognitive processes involved. Recently, rapid advancements in generative AI, particularly through the development of Large Language Models (LLMs), have significantly transformed the landscape of writing assistance. However, underrepresented languages like Arabic encounter significant challenges in the development of advanced AI writing tools, largely due to the limited availability of data. This scarcity constrains the training of effective models, impeding the creation of sophisticated writing assistance technologies. To address these issues, we present *Gazelle*, a comprehensive dataset for Arabic writing assistance. In addition, we offer an evaluation framework designed to enhance Arabic writing assistance tools. Our human evaluation of leading LLMs, including GPT-**4**, GPT-**4o**, Cohere Command R+, and Gemini **1.5** Pro, highlights their respective strengths and limitations in addressing the challenges of Arabic writing. Our findings underscore the need for continuous model training and dataset enrichment to manage the complexities of Arabic language processing, paving the way for more effective AI-powered Arabic writing tools
On the Limited Generalization Capability of the Implicit Reward Model Induced by Direct Preference Optimization
PDF ↗Reinforcement Learning from Human Feedback (RLHF) is an effective approach for aligning language models to human preferences. Central to RLHF is learning a reward function for scoring human preferences. Two main approaches for learning a reward model are 1) training an EXplicit Reward Model (EXRM) as in RLHF, and 2) using an implicit reward learned from preference data through methods such as Direct Preference Optimization (DPO). Prior work has shown that the implicit reward model of DPO (denoted as DPORM) can approximate an EXRM on the limit infinite samples. However, it is unclear how effective is DPORM in practice. DPORM’s effectiveness directly implies the optimality of learned policy of DPO and also has practical implication for more advanced alignment methods, such as iterative DPO. We compare the accuracy at distinguishing preferred and rejected answers using both DPORM and EXRM. Our findings indicate that even though DPORM can fit the training dataset, it generalizes less effective than EXRM, especially when the validation datasets contain distributional shifts. Across five out-of-distribution settings, DPORM has a mean drop in accuracy of 3% and a maximum drop of 7%. These findings highlight that DPORM has limited generalization ability and substantiates the integration of an explicit reward model in iterative DPO approaches.
The correct specification of reward models is a well-known challenge in reinforcement learning.Hand-crafted reward functions often lead to inefficient or suboptimal policies and may not be aligned with user values.Reinforcement learning from human feedback is a successful technique that can mitigate such issues, however, the collection of human feedback can be laborious.Recent works have solicited feedback from pre-trained large language models rather than humans to reduce or eliminate human effort, however, these approaches yield poor performance in the presence of hallucination and other errors.This paper studies the advantages and limitations of reinforcement learning from large language model feedback and proposes a simple yet effective method for soliciting and applying feedback as a potential-based shaping function.We theoretically show that inconsistent rankings – which approximate ranking errors – lead to uninformative rewards with our approach. Our method empirically improves convergence speed and policy returns over commonly used baselines even with significant ranking errors, and eliminates the need for complex post-processing of reward functions.
Large Language Models (LLMs) have led to significant improvements in the Knowledge Base Question Answering (KBQA) task. However, datasets used in KBQA studies do not capture the true complexity of KBQA tasks. They either have simple questions, use synthetically generated logical forms, or are based on small knowledge base (KB) schemas.We introduce the SPINACH dataset, an expert-annotated KBQA dataset collected from discussions on Wikidata’s “Request a Query” forum with 320 decontextualized question-SPARQL pairs. The complexity of these in-the-wild queries calls for a KBQA system that can dynamically explore large and often incomplete schemas and reason about them, as it is infeasible to create a comprehensive training dataset. We also introduce an in-context learning KBQA agent, also called SPINACH, that mimics how a human expert would write SPARQLs to handle challenging questions. SPINACH achieves a new state of the art on the QALD-7, QALD-9 Plus and QALD-10 datasets by 31.0%, 27.0%, and 10.0% in F_1, respectively, and coming within 1.6% of the fine-tuned LLaMA SOTA model on WikiWebQuestions.On our new SPINACH dataset, the SPINACH agent outperforms all baselines, including the best GPT-4-based KBQA agent, by at least 38.1% in F_1.