The increase in parameter size of multimodal large language models (MLLMs) introduces significant capabilities, particularly multimodal in-context learning, where MLLMs enhance task performance without updating pre-trained parameters. However, this effectiveness hinges on the appropriate selection of in-context examples, a process currently biased towards visual data, overlooking textual information. More importantly, the area of supervised retrievers for retrieval of multimodal in-context learning, crucial for optimal in-context example selection, continues to be investigated. Our study provides an in-depth evaluation of the impact of textual information on the unsupervised selection of in-context examples in multimodal contexts, uncovering a notable sensitivity of retriever performance to the employed modalities. Based on the above finding, we introduce a novel supervised MLLM prompt retriever MSIER that leverages a trained retriever based on MLLM’s confidence to select examples, which enhances multimodal in-context learning efficiency. This approach is validated through extensive testing across three different tasks, demonstrating the method’s effectiveness. Additionally, we investigate the influence of modalities on our supervised retrieval method’s training and explore the transferability of the supervised prompt retriever. This exploration paves the way for future advancements, highlighting the potential for refined in-context learning in MLLMs through the strategic use of multimodal data. The public code is available at https://github.com/NUS-HPC-AI-Lab/Multimodal-ICL-Retriever.
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
Formality is Favored: Unraveling the Learning Preferences of Large Language Models on Data with Conflicting Knowledge
PDF ↗Having been trained on massive pretraining data, large language models have shown excellent performance on many knowledge-intensive tasks. However, pretraining data tends to contain misleading and even conflicting information, and it is intriguing to understand how LLMs handle these noisy data during training. In this study, we systematically analyze LLMs’ learning preferences for data with conflicting knowledge. We find that pretrained LLMs establish learning preferences similar to humans, i.e., preferences towards formal texts and texts with fewer spelling errors, resulting in faster learning and more favorable treatment of knowledge in data with such features when facing conflicts. This finding is generalizable across models and languages and is more evident in larger models. An in-depth analysis reveals that LLMs tend to trust data with features that signify consistency with the majority of data, and it is possible to instill new preferences and erase old ones by manipulating the degree of consistency with the majority data.
Private data, being larger and quality-higher than public data, can greatly improve large language models (LLM). However, due to privacy concerns, this data is often dispersed in multiple silos, making its secure utilization for LLM training a challenge. Federated learning (FL) is an ideal solution for training models with distributed private data, but traditional frameworks like FedAvg are unsuitable for LLM due to their high computational demands on clients. An alternative, split learning, offloads most training parameters to the server while training embedding and output layers locally, making it more suitable for LLM. Nonetheless, it faces significant challenges in security and efficiency. Firstly, the gradients of embeddings are prone to attacks, leading to potential reverse engineering of private data. Furthermore, the server’s limitation of handling only one client’s training request at a time hinders parallel training, severely impacting training efficiency. In this paper, we propose a Federated Learning framework for LLM, named FL-GLM, which prevents data leakage caused by both server-side and peer-client attacks while improving training efficiency. Specifically, we first place the input block and output block on local client to prevent embedding gradient attacks from server. Secondly, we employ key-encryption during client-server communication to prevent reverse engineering attacks from peer-clients. Lastly, we employ optimization methods like client-batching or server-hierarchical, adopting different acceleration methods based on the actual computational capabilities of the server. Experimental results on NLU and generation tasks demonstrate that FL-GLM achieves comparable metrics to centralized chatGLM model, validating the effectiveness of our federated learning framework.
Recent progress in Spoken Language Modeling has shown that learning language directly from speech is feasible. Generating speech through a pipeline that operates at the text level typically loses nuances, intonations, and non-verbal vocalizations. Modeling directly from speech opens up the path to more natural and expressive systems. On the other hand, speech-only systems require up to three orders of magnitude more data to catch up to their text-based counterparts in terms of their semantic abilities. We show that fine-tuning speech representation models on phoneme classification leads to more context-invariant representations, and language models trained on these units achieve comparable lexical comprehension to ones trained on hundred times more data.
Logical reasoning remains a challenge for natural language processing, but it can be improved by training language models to mimic theorem provers on procedurally generated problems. Previous work used domain-specific proof generation algorithms, which biases reasoning toward specific proof traces and limits auditability and extensibility. We present a simpler and more general declarative framework with flexible context-sensitive rules binding multiple languages (specifically, simplified English and the TPTP theorem-proving language). We construct first-order logic problems by selecting up to 32 premises and one hypothesis. We demonstrate that using semantic constraints during generation and careful English verbalization of predicates enhances logical reasoning without hurting natural English tasks. Using relatively small DeBERTa-v3 models, we achieve state-of-the-art accuracy on the FOLIO human-authored logic dataset, surpassing GPT-4 in accuracy with or without an external solver by 12%.
As the scale of training corpora for large language models (LLMs) grows, model developers become increasingly reluctant to disclose details on their data. This lack of transparency poses challenges to scientific evaluation and ethical deployment. Recently, pretraining data detection approaches, which infer whether a given text was part of an LLM’s training data through black-box access, have been explored. The Min-K% Prob method, which has achieved state-of-the-art results, assumes that a non-training example tends to contain a few outlier words with low token probabilities. However, the effectiveness may be limited as it tends to misclassify non-training texts that contain many common words with high probabilities predicted by LLMs. To address this issue, we introduce a divergence-based calibration method, inspired by the divergence-from-randomness concept, to calibrate token probabilities for pretraining data detection. We compute the cross-entropy (i.e., the divergence) between the token probability distribution and the token frequency distribution to derive a detection score.We have developed a Chinese-language benchmark, PatentMIA, to assess the performance of detection approaches for LLMs on Chinese text. Experimental results on English-language benchmarks and PatentMIA demonstrate that our proposed method significantly outperforms existing methods. Our code and PatentMIA benchmark are available at https://github.com/zhang-wei-chao/DC-PDD.
Large Language Models (LLMs) have demonstrated remarkable capability in a variety of NLP tasks. However, LLMs are also prone to generate nonfactual content. Uncertainty Quantification (UQ) is pivotal in enhancing our understanding of a model’s confidence on its generation, thereby aiding in the mitigation of nonfactual outputs. Existing research on UQ predominantly targets short text generation, typically yielding brief, word-limited responses. However, real-world applications frequently necessitate much longer responses. Our study first highlights the limitations of current UQ methods in handling long text generation. We then introduce Luq and its two variations, a series of novel sampling-based UQ approaches specifically designed for long text. Our findings reveal that Luq outperforms existing baseline methods in correlating with the model’s factuality scores (negative coefficient of -0.85 observed for Gemini Pro). To further improve the factuality of LLM responses, we propose Luq-Ensemble, a method that ensembles responses from multiple models and selects the response with the lowest uncertainty. The ensembling method greatly improves the response factuality upon the best standalone LLM.
Large Language Models (LLMs) have demonstrated capability in “instruction induction,” generating instructions from demonstrations (input-output pairs). However, existing methods often rely on large datasets or numerous examples, which is impractical and costly in real-world scenarios. In this work, we propose a low-cost, task-level framework called Induct-Learn. It induces pseudo instructions from a few demonstrations and a short phrase, adding a CoT process into existing demonstrations. When encountering new problems, the learned pseudo instructions and demonstrations with the pseudo CoT process can be combined into a prompt to guide the LLM’s problem-solving process. We validate our approach on the BBH-Induct and Evals-Induct datasets, and the results show that the Induct-Learn framework outperforms state-of-the-art methods. We also exhibit cross-model adaptability and achieve superior performance at a lower cost compared to existing methods.
SEACrowd: A Multilingual Multimodal Data Hub and Benchmark Suite for Southeast Asian Languages
PDF ↗Southeast Asia (SEA) is a region rich in linguistic diversity and cultural variety, with over 1,300 indigenous languages and a population of 671 million people. However, prevailing AI models suffer from a significant lack of representation of texts, images, and audio datasets from SEA, compromising the quality of AI models for SEA languages. Evaluating models for SEA languages is challenging due to the scarcity of high-quality datasets, compounded by the dominance of English training data, raising concerns about potential cultural misrepresentation. To address these challenges, through a collaborative movement, we introduce SEACrowd, a comprehensive resource center that fills the resource gap by providing standardized corpora in nearly 1,000 SEA languages across three modalities. Through our SEACrowd benchmarks, we assess the quality of AI models on 36 indigenous languages across 13 tasks, offering valuable insights into the current AI landscape in SEA. Furthermore, we propose strategies to facilitate greater AI advancements, maximizing potential utility and resource equity for the future of AI in Southeast Asia.
Large language models (LLMs) demonstrate exceptional instruct-following ability to complete various downstream tasks. Although this impressive ability makes LLMs flexible task solvers, their performance in solving tasks also heavily relies on instructions. In this paper, we reveal that LLMs are over-sensitive to lexical variations in task instructions, even when the variations are imperceptible to humans. By providing models with neighborhood instructions, which are closely situated in the latent representation space and differ by only one semantically similar word, the performance on downstream tasks can be vastly different. Following this property, we propose a black-box Combinatorial Optimization framework for Prompt Lexical Enhancement (COPLE). COPLE performs iterative lexical optimization according to the feedback from a batch of proxy tasks, using a search strategy related to word influence. Experiments show that even widely-used human-crafted prompts for current benchmarks suffer from the lexical sensitivity of models, and COPLE recovers the declined model ability in both instruct-following and solving downstream tasks.
In Machine Translation (MT) evaluations, the conventional approach is to compare a translated sentence against its human-created reference sentence. MT metrics provide an absolute score (e.g., from 0 to 1) to a candidate sentence based on the similarity with the reference sentence. Thus, existing MT metrics give the maximum score to the reference sentence. However, this approach overlooks the potential for a candidate sentence to exceed the reference sentence in terms of quality. In particular, recent advancements in Large Language Models (LLMs) have highlighted this issue, as LLM-generated sentences often exceed the quality of human-written sentences. To address the problem, we introduce the Residual score Metric (ResuMe), which evaluates the relative quality between reference and candidate sentences. ResuMe assigns a positive score to candidate sentences that outperform their reference sentences, and a negative score when they fall short. By adding the residual scores from ResuMe to the absolute scores from MT metrics, it can be possible to allocate higher scores to candidate sentences than what reference sentences are received from MT metrics. Experimental results demonstrate that ResuMe enhances the alignments between MT metrics and human judgments both at the segment-level and the system-level.
Commercially available models dominate academic leaderboards. While impressive, this has concentrated research on creating and adapting general-purpose models to improve NLP leaderboard standings for large language models. However, leaderboards collect many individual tasks and general-purpose models often underperform in specialized domains; domain-specific or adapted models yield superior results. This focus on large general-purpose models excludes many academics and draws attention away from areas where they can make important contributions. We advocate for a renewed focus on developing and evaluating domain- and task-specific models, and highlight the unique role of academics in this endeavor.
Claim: This work is not advocating the use of LLMs for paper (meta-)reviewing. Instead, wepresent a comparative analysis to identify and distinguish LLM activities from human activities. Two research goals: i) Enable better recognition of instances when someone implicitly uses LLMs for reviewing activities; ii) Increase community awareness that LLMs, and AI in general, are currently inadequate for performing tasks that require a high level of expertise and nuanced judgment.This work is motivated by two key trends. On one hand, large language models (LLMs) have shown remarkable versatility in various generative tasks such as writing, drawing, and question answering, significantly reducing the time required for many routine tasks. On the other hand, researchers, whose work is not only time-consuming but also highly expertise-demanding, face increasing challenges as they have to spend more time reading, writing, and reviewing papers. This raises the question: how can LLMs potentially assist researchers in alleviating their heavy workload?This study focuses on the topic of LLMs as NLP Researchers, particularly examining the effectiveness of LLMs in assisting paper (meta-)reviewing and its recognizability. To address this, we constructed the ReviewCritique dataset, which includes two types of information: (i) NLP papers (initial submissions rather than camera-ready) with both human-written and LLM-generated reviews, and (ii) each review comes with “deficiency” labels and corresponding explanations for individual segments, annotated by experts. Using ReviewCritique, this study explores two threads of research questions: (i) “LLMs as Reviewers”, how do reviews generated by LLMs compare with those written by humans in terms of quality and distinguishability? (ii) “LLMs as Metareviewers”, how effectively can LLMs identify potential issues, such as Deficient or unprofessional review segments, within individual paper reviews? To our knowledge, this is the first work to provide such a comprehensive analysis.
Instruction tuning, or supervised finetuning on extensive task-specific data, is necessary for Large Vision-Language Models (LVLMs) to generalize well across a broad range of vision-language (VL) tasks. However, training on large VL datasets can become prohibitively expensive. In this work, we introduce COINCIDE, an effective and scalable data selection technique that uses a small model as a reference model to select visual instruction tuning data for efficient finetuning of a target LVLM, focusing on diversity and transferability. Specifically, we cluster the training data using internal activations from a small model, which identifies VL concept-skill compositions needed by a target LVLM. We then sample data from these diverse clusters by considering their density and transferability, or the ability to transfer well to other concept-skill compositions. This approach ensures the diversity of these compositions, which is vital for LVLM generalization. Extensive experiments demonstrate that COINCIDE achieves superior performance and data selection efficiency against 8 strong baselines on two distinct datasets: LLaVA-1.5 and Vision-Flan. Using only 20% of the LLaVA-1.5 dataset, COINCIDE achieves performance comparable to the LVLM finetuned on the whole dataset, with 70% reduction of the wall-clock running time. On the Vision-Flan dataset, our method achieves superior results with only 16.7% of the training data.
FOOL ME IF YOU CAN! An Adversarial Dataset to Investigate the Robustness of LMs in Word Sense Disambiguation
PDF ↗Word sense disambiguation (WSD) is a key task in natural language processing and lexical semantics. Pre-trained language models with contextualized word embeddings have significantly improved performance in regular WSD tasks. However, these models still struggle with recognizing semantic boundaries and often misclassify homonyms in adversarial context. Therefore, we propose FOOL: FOur-fold Obscure Lexical, a new coarse-grained WSD dataset, which includes four different test sets designed to assess the robustness of language models in WSD tasks. Two sets feature typical WSD scenarios, while the other two include sentences with opposing contexts to challenge the models further.We tested two types of models on the proposed dataset: models with encoders, such as the BERT and T5 series of varying sizes by probing their embeddings, and state-of-the-art large decoder models like GPT-4o and the LlaMA3 family, using zero shot prompting. Across different state-of-the-art language models, we observed a decrease in performance in the latter two sets compared to the first two, with some models being affected more than others. We show interesting findings where small models like T5-large and BERT-large performed better than GPT-4o on Set 3 of the dataset. This indicates that, despite excelling in regular WSD tasks, these models still struggle to correctly disambiguate homonyms in artificial (Set 3) or realistic adversarial contexts (Set 4).
Large Language models (LLMs) have exhibited remarkable abilities in understanding complex texts, offering a promising path towards human-like translation performance. However, this study reveals the misalignment between the translation-specific understanding and the general understanding inside LLMs. This understanding misalignment leads to LLMs mistakenly or literally translating some complicated concepts that they accurately comprehend in the general scenarios (e.g., QA). To align the translation-specific understanding to the general one, we propose a novel translation process, DUAT (Difficult words Understanding Aligned Translation), explicitly incorporating the general understanding on the complicated content incurring inconsistent understandings to guide the translation. Specifically, DUAT performs cross-lingual interpretation for the difficult-to-translate words and enhances the translation with the generated interpretations. Furthermore, we reframe the external tools to improve DUAT in detecting difficult words and generating helpful interpretations. We conduct experiments on the self-constructed benchmark Challenge-WMT, consisting of samples that are prone to mistranslation. Human evaluation results on high-resource and low-resource language pairs indicate that DUAT significantly facilitates the understanding alignment, which improves the translation quality (up to +3.85 COMET) and reduces translation literalness by -25% ∼ -51%.
Hierarchical Deconstruction of LLM Reasoning: A Graph-Based Framework for Analyzing Knowledge Utilization
PDF ↗Despite the advances in large language models (LLMs), how they use their knowledge for reasoning is not yet well understood.In this study, we propose a method that deconstructs complex real-world questions into a graph, representing each question as a node with predecessors of background knowledge needed to solve the question. We develop the DepthQA dataset, deconstructing questions into three depths: (i) recalling conceptual knowledge, (ii) applying procedural knowledge, and (iii) analyzing strategic knowledge. Based on a hierarchical graph, we quantify forward discrepancy, a discrepancy in LLM performance on simpler sub-problems versus complex questions. We also measure backward discrepancy where LLMs answer complex questions but struggle with simpler ones. Our analysis shows that smaller models exhibit more discrepancies than larger models. Distinct patterns of discrepancies are observed across model capacity and possibility of training data memorization. Additionally, guiding models from simpler to complex questions through multi-turn interactions improves performance across model sizes, highlighting the importance of structured intermediate steps in knowledge reasoning. This work enhances our understanding of LLM reasoning and suggests ways to improve their problem-solving abilities.
MaPPER: Multimodal Prior-guided Parameter Efficient Tuning for Referring Expression Comprehension
PDF ↗Referring Expression Comprehension (REC), which aims to ground a local visual region via natural language, is a task that heavily relies on multimodal alignment. Most existing methods utilize powerful pre-trained models to transfer visual/linguistic knowledge by full fine-tuning. However, full fine-tuning the entire backbone not only breaks the rich prior knowledge embedded in the pre-training, but also incurs significant computational costs. Motivated by the recent emergence of Parameter-Efficient Transfer Learning (PETL) methods, we aim to solve the REC task in an effective and efficient manner. Directly applying these PETL methods to the REC task is inappropriate, as they lack the specific-domain abilities for precise local visual perception and visual-language alignment. Therefore, we propose a novel framework of Multimodal Prior-guided Parameter Efficient Tuning, namely MaPPER. Specifically, MaPPER comprises Dynamic Prior Adapters guided by a aligned prior, and Local Convolution Adapters to extract precise local semantics for better visual perception. Moreover, the Prior-Guided Text module is proposed to further utilize the prior for facilitating the cross-modal alignment. Experimental results on three widely-used benchmarks demonstrate that MaPPER achieves the best accuracy compared to the full fine-tuning and other PETL methods with only 1.41% tunable backbone parameters.
Audio deepfake detection (ADD) is essential for preventing the misuse of synthetic voices that may infringe on personal rights and privacy. Recent zero-shot text-to-speech (TTS) models pose higher risks as they can clone voices with a single utterance. However, the existing ADD datasets are outdated, leading to suboptimal generalization of detection models. In this paper, we construct a new cross-domain ADD dataset comprising over 300 hours of speech data that is generated by five advanced zero-shot TTS models. To simulate real-world scenarios, we employ diverse attack methods and audio prompts from different datasets. Experiments show that, through novel attack-augmented training, the Wav2Vec2-large and Whisper-medium models achieve equal error rates of 4.1% and 6.5% respectively. Additionally, we demonstrate our models’ outstanding few-shot ADD ability by fine-tuning with just one minute of target-domain data. Nonetheless, neural codec compressors greatly affect the detection accuracy, necessitating further research. Our dataset is publicly available (https://github.com/leolya/CD-ADD).