Large-language models are notoriously famous for their impressive performance across a wide range of tasks. One surprising example of such impressive performance is a recently identified capacity of LLMs to understand the governing principles of dynamical systems satisfying the Markovian property. In this paper, we seek to explore this direction further by studying the dynamics of stochastic gradient descent in convex and non-convex optimization. By leveraging the theoretical link between the SGD and Markov chains, we show a remarkable zero-shot performance of LLMs in predicting the local minima to which SGD converges for previously unseen starting points. On a more general level, we inquire about the possibility of using LLMs to perform zero-shot randomized trials for larger deep learning models used in practice.
相關內容
Neural language models (LMs) are arguably less data-efficient than humans from a language acquisition perspective. One fundamental question is why this human-LM gap arises. This study explores the advantage of grounded language acquisition, specifically the impact of visual information -- which humans can usually rely on but LMs largely do not have access to during language acquisition -- on syntactic generalization in LMs. Our experiments, following the poverty of stimulus paradigm under two scenarios (using artificial vs. naturalistic images), demonstrate that if the alignments between the linguistic and visual components are clear in the input, access to vision data does help with the syntactic generalization of LMs, but if not, visual input does not help. This highlights the need for additional biases or signals, such as mutual gaze, to enhance cross-modal alignment and enable efficient syntactic generalization in multimodal LMs.
When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge. We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model's knowledge. However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model's tendency to hallucinate. Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently.
Vision-Language Models (VLMs) have recently demonstrated incredible strides on diverse vision language tasks. We dig into vision-based deductive reasoning, a more sophisticated but less explored realm, and find previously unexposed blindspots in the current SOTA VLMs. Specifically, we leverage Raven's Progressive Matrices (RPMs), to assess VLMs' abilities to perform multi-hop relational and deductive reasoning relying solely on visual clues. We perform comprehensive evaluations of several popular VLMs employing standard strategies such as in-context learning, self-consistency, and Chain-of-thoughts (CoT) on three diverse datasets, including the Mensa IQ test, IntelligenceTest, and RAVEN. The results reveal that despite the impressive capabilities of LLMs in text-based reasoning, we are still far from achieving comparable proficiency in visual deductive reasoning. We found that certain standard strategies that are effective when applied to LLMs do not seamlessly translate to the challenges presented by visual reasoning tasks. A detailed analysis reveals that VLMs struggle to solve these tasks mainly because they are unable to perceive and comprehend multiple, confounding abstract patterns in RPM examples.
ChatGPT and other state-of-the-art large language models (LLMs) are rapidly transforming multiple fields, offering powerful tools for a wide range of applications. These models, commonly trained on vast datasets, exhibit human-like text generation capabilities, making them useful for research tasks such as ideation, literature review, coding, drafting, and outreach. We conducted a study involving 13 astronomers at different career stages and research fields to explore LLM applications across diverse tasks over several months and to evaluate their performance in research-related activities. This work was accompanied by an anonymous survey assessing participants' experiences and attitudes towards LLMs. We provide a detailed analysis of the tasks attempted and the survey answers, along with specific output examples. Our findings highlight both the potential and limitations of LLMs in supporting research while also addressing general and research-specific ethical considerations. We conclude with a series of recommendations, emphasizing the need for researchers to complement LLMs with critical thinking and domain expertise, ensuring these tools serve as aids rather than substitutes for rigorous scientific inquiry.
Influence functions aim to quantify the impact of individual training data points on a model's predictions. While extensive research has been conducted on influence functions in traditional machine learning models, their application to large language models (LLMs) has been limited. In this work, we conduct a systematic study to address a key question: do influence functions work on LLMs? Specifically, we evaluate influence functions across multiple tasks and find that they consistently perform poorly in most settings. Our further investigation reveals that their poor performance can be attributed to: (1) inevitable approximation errors when estimating the iHVP component due to the scale of LLMs, (2) uncertain convergence during fine-tuning, and, more fundamentally, (3) the definition itself, as changes in model parameters do not necessarily correlate with changes in LLM behavior. Our study thus suggests the need for alternative approaches for identifying influential samples. To support future work, our code is made available at //github.com/plumprc/Failures-of-Influence-Functions-in-LLMs.
Creativity is a fundamental pillar of human expression and a driving force behind innovation, yet it now stands at a crossroads. As artificial intelligence advances at an astonishing pace, the question arises: can machines match and potentially surpass human creativity? This study investigates the creative performance of artificial intelligence (AI) compared to humans by analyzing the effects of two distinct prompting strategies (a Naive and an Expert AI) on AI and across three different tasks (Text, Draw and Alternative Uses tasks). Human external evaluators have scored creative outputs generated by humans and AI, and these subjective creative scores were complemented with objective measures based on quantitative measurements and NLP tools. The results reveal that AI generally outperforms humans in creative tasks, though this advantage is nuanced by the specific nature of each task and the chosen creativity criteria. Ultimately, while AI demonstrates superior performance in certain creative domains, our results suggest that integrating human feedback is crucial for maximizing AI's creative potential.
Co-speech gestures are fundamental for communication. The advent of recent deep learning techniques has facilitated the creation of lifelike, synchronous co-speech gestures for Embodied Conversational Agents. "In-the-wild" datasets, aggregating video content from platforms like YouTube via human pose detection technologies, provide a feasible solution by offering 2D skeletal sequences aligned with speech. Concurrent developments in lifting models enable the conversion of these 2D sequences into 3D gesture databases. However, it is important to note that the 3D poses estimated from the 2D extracted poses are, in essence, approximations of the ground-truth, which remains in the 2D domain. This distinction raises questions about the impact of gesture representation dimensionality on the quality of generated motions - a topic that, to our knowledge, remains largely unexplored. Our study examines the effect of using either 2D or 3D joint coordinates as training data on the performance of speech-to-gesture deep generative models. We employ a lifting model for converting generated 2D pose sequences into 3D and assess how gestures created directly in 3D stack up against those initially generated in 2D and then converted to 3D. We perform an objective evaluation using widely used metrics in the gesture generation field as well as a user study to qualitatively evaluate the different approaches.
Although recent advancements in large language models (LLMs) have significantly improved their performance on various tasks, they still face challenges with complex and symbolic multi-step reasoning, particularly in mathematical reasoning. To bolster the mathematical reasoning capabilities of LLMs, most existing efforts concentrate on seeking assistance from either domain experts or GPT-4 for high-quality process-supervised data, which is not only expensive but also labor-intensive. In our study, we propose an innovative framework, AlphaMath, that bypasses the need for process annotations (from humans or GPTs) by leveraging Monte Carlo Tree Search (MCTS). This framework focuses on unleashing the potential of a well-pretrained LLM to autonomously enhance its mathematical reasoning. Specifically, we integrate a value model with the LLM, automatically generating both process supervision and step-level evaluation signals in MCTS. Furthermore, we propose an efficient inference strategy, step-level beam search, where the value model is crafted to assist the policy model (i.e., LLM) in navigating more effective reasoning paths, rather than solely relying on prior probabilities. The experimental results on both in-domain and out-of-domain datasets demonstrate that even without GPT-4 or human-annotated process supervision, our AlphaMath framework achieves comparable or superior results to previous state-of-the-art methods.
Self-supervised pre-training has proven highly effective for many computer vision tasks, particularly when labelled data are scarce. In the context of Earth Observation (EO), foundation models and various other Vision Transformer (ViT)-based approaches have been successfully applied for transfer learning to downstream tasks. However, it remains unclear under which conditions pre-trained models offer significant advantages over training from scratch. In this study, we investigate the effectiveness of pre-training ViT-based Masked Autoencoders (MAE) for downstream EO tasks, focusing on reconstruction, segmentation, and classification. We consider two large ViT-based MAE pre-trained models: a foundation model (Prithvi) and SatMAE. We evaluate Prithvi on reconstruction and segmentation-based downstream tasks, and for SatMAE we assess its performance on a classification downstream task. Our findings suggest that pre-training is particularly beneficial when the fine-tuning task closely resembles the pre-training task, e.g. reconstruction. In contrast, for tasks such as segmentation or classification, training from scratch with specific hyperparameter adjustments proved to be equally or more effective.
Language model pre-training has proven to be useful in learning universal language representations. As a state-of-the-art language model pre-training model, BERT (Bidirectional Encoder Representations from Transformers) has achieved amazing results in many language understanding tasks. In this paper, we conduct exhaustive experiments to investigate different fine-tuning methods of BERT on text classification task and provide a general solution for BERT fine-tuning. Finally, the proposed solution obtains new state-of-the-art results on eight widely-studied text classification datasets.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191