Large Language Models (LLMs) do not differentially represent numbers, which are pervasive in text. In contrast, neuroscience research has identified distinct neural representations for numbers and words. In this work, we investigate how well popular LLMs capture the magnitudes of numbers (e.g., that $4 < 5$) from a behavioral lens. Prior research on the representational capabilities of LLMs evaluates whether they show human-level performance, for instance, high overall accuracy on standard benchmarks. Here, we ask a different question, one inspired by cognitive science: How closely do the number representations of LLMscorrespond to those of human language users, who typically demonstrate the distance, size, and ratio effects? We depend on a linking hypothesis to map the similarities among the model embeddings of number words and digits to human response times. The results reveal surprisingly human-like representations across language models of different architectures, despite the absence of the neural circuitry that directly supports these representations in the human brain. This research shows the utility of understanding LLMs using behavioral benchmarks and points the way to future work on the number representations of LLMs and their cognitive plausibility.
相關內容
Today, large language models (LLMs) are taught to use new tools by providing a few demonstrations of the tool's usage. Unfortunately, demonstrations are hard to acquire, and can result in undesirable biased usage if the wrong demonstration is chosen. Even in the rare scenario that demonstrations are readily available, there is no principled selection protocol to determine how many and which ones to provide. As tasks grow more complex, the selection search grows combinatorially and invariably becomes intractable. Our work provides an alternative to demonstrations: tool documentation. We advocate the use of tool documentation, descriptions for the individual tool usage, over demonstrations. We substantiate our claim through three main empirical findings on 6 tasks across both vision and language modalities. First, on existing benchmarks, zero-shot prompts with only tool documentation are sufficient for eliciting proper tool usage, achieving performance on par with few-shot prompts. Second, on a newly collected realistic tool-use dataset with hundreds of available tool APIs, we show that tool documentation is significantly more valuable than demonstrations, with zero-shot documentation significantly outperforming few-shot without documentation. Third, we highlight the benefits of tool documentations by tackling image generation and video tracking using just-released unseen state-of-the-art models as tools. Finally, we highlight the possibility of using tool documentation to automatically enable new applications: by using nothing more than the documentation of GroundingDino, Stable Diffusion, XMem, and SAM, LLMs can re-invent the functionalities of the just-released Grounded-SAM and Track Anything models.
Large language models (LLMs) have shown remarkable capacity for in-context learning (ICL), where learning a new task from just a few training examples is done without being explicitly pre-trained. However, despite the success of LLMs, there has been little understanding of how ICL learns the knowledge from the given prompts. In this paper, to make progress toward understanding the learning behaviour of ICL, we train the same LLMs with the same demonstration examples via ICL and supervised learning (SL), respectively, and investigate their performance under label perturbations (i.e., noisy labels and label imbalance) on a range of classification tasks. First, via extensive experiments, we find that gold labels have significant impacts on the downstream in-context performance, especially for large language models; however, imbalanced labels matter little to ICL across all model sizes. Second, when comparing with SL, we show empirically that ICL is less sensitive to label perturbations than SL, and ICL gradually attains comparable performance to SL as the model size increases.
One advantage of neural ranking models is that they are meant to generalise well in situations of synonymity i.e. where two words have similar or identical meanings. In this paper, we investigate and quantify how well various ranking models perform in a clear-cut case of synonymity: when words are simply expressed in different surface forms due to regional differences in spelling conventions (e.g., color vs colour). We first explore the prevalence of American and British English spelling conventions in datasets used for the pre-training, training and evaluation of neural retrieval methods, and find that American spelling conventions are far more prevalent. Despite these biases in the training data, we find that retrieval models often generalise well in this case of synonymity. We explore the effect of document spelling normalisation in retrieval and observe that all models are affected by normalising the document's spelling. While they all experience a drop in performance when normalised to a different spelling convention than that of the query, we observe varied behaviour when the document is normalised to share the query spelling convention: lexical models show improvements, dense retrievers remain unaffected, and re-rankers exhibit contradictory behaviour.
It is evident that the current state of Large Language Models (LLMs) necessitates the incorporation of external tools. The lack of straightforward algebraic and logical reasoning is well documented and prompted researchers to develop frameworks which allow LLMs to operate via external tools. The ontological nature of tool utilization for a specific task can be well formulated with a Directed Acyclic Graph (DAG). The central aim of the paper is to highlight the importance of graph based approaches to LLM-tool interaction in near future. We propose an exemplary framework to guide the orchestration of exponentially increasing numbers of external tools with LLMs,where objectives and functionalities of tools are graph encoded hierarchically. Assuming that textual segments of a Chain-of-Thought (CoT) can be imagined as a tool as defined here, the graph based framework can pave new avenues in that particular direction as well.
An important aspect in developing language models that interact with humans is aligning their behavior to be useful and unharmful for their human users. This is usually achieved by tuning the model in a way that enhances desired behaviors and inhibits undesired ones, a process referred to as alignment. In this paper, we propose a theoretical approach called Behavior Expectation Bounds (BEB) which allows us to formally investigate several inherent characteristics and limitations of alignment in large language models. Importantly, we prove that for any behavior that has a finite probability of being exhibited by the model, there exist prompts that can trigger the model into outputting this behavior, with probability that increases with the length of the prompt. This implies that any alignment process that attenuates undesired behavior but does not remove it altogether, is not safe against adversarial prompting attacks. Furthermore, our framework hints at the mechanism by which leading alignment approaches such as reinforcement learning from human feedback increase the LLM's proneness to being prompted into the undesired behaviors. Moreover, we include the notion of personas in our BEB framework, and find that behaviors which are generally very unlikely to be exhibited by the model can be brought to the front by prompting the model to behave as specific persona. This theoretical result is being experimentally demonstrated in large scale by the so called contemporary "chatGPT jailbreaks", where adversarial users trick the LLM into breaking its alignment guardrails by triggering it into acting as a malicious persona. Our results expose fundamental limitations in alignment of LLMs and bring to the forefront the need to devise reliable mechanisms for ensuring AI safety.
Recently, there has been growing interest in extending the context length of instruction-following models in order to effectively process single-turn long input (e.g. summarizing a paper) and conversations with more extensive histories. While proprietary models such as GPT-4 and Claude have shown significant strides in handling extremely lengthy input, open-sourced models are still in the early stages of experimentation. It also remains unclear whether extending the context can offer substantial gains over traditional methods such as retrieval, and to what extent it improves upon their regular counterparts in practical downstream tasks. To address this challenge, we propose instituting standardized evaluation for long context language models. Concretely, we develop L-Eval which contains 411 long documents and over 2,000 human-labeled query-response pairs encompassing areas such as law, finance, school lectures, lengthy conversations, news, long-form novels, and meetings. L-Eval also adopts diverse evaluation methods and instruction styles, enabling a more reliable assessment of Long Context Language Models (LCLMs). Our findings indicate that while open-source models typically lag behind commercial models, they still exhibit impressive performance compared with their regular versions. LLaMA2-13B achieves the best results on both open-ended tasks (win \textbf{42}\% vs turbo-16k-0613) and closed-ended tasks with only 4k context length. We release our new evaluation suite, code, and all generation results including predictions from all open-sourced LCLMs, GPT4-32k, Cluade-100k at {\url{//github.com/OpenLMLab/LEval}}.
The field of large language models (LLMs) has made significant progress, and their knowledge storage capacity is approaching that of human beings. Furthermore, advanced techniques, such as prompt learning and reinforcement learning, are being employed to address ethical concerns and hallucination problems associated with LLMs, bringing them closer to aligning with human values. This situation naturally raises the question of whether LLMs with human-like abilities possess a human-like personality? In this paper, we aim to investigate the feasibility of using the Myers-Briggs Type Indicator (MBTI), a widespread human personality assessment tool, as an evaluation metric for LLMs. Specifically, extensive experiments will be conducted to explore: 1) the personality types of different LLMs, 2) the possibility of changing the personality types by prompt engineering, and 3) How does the training dataset affect the model's personality. Although the MBTI is not a rigorous assessment, it can still reflect the similarity between LLMs and human personality. In practice, the MBTI has the potential to serve as a rough indicator. Our codes are available at //github.com/HarderThenHarder/transformers_tasks/tree/main/LLM/llms_mbti.
The pre-training and fine-tuning paradigm has contributed to a number of breakthroughs in Natural Language Processing (NLP). Instead of directly training on a downstream task, language models are first pre-trained on large datasets with cross-domain knowledge (e.g., Pile, MassiveText, etc.) and then fine-tuned on task-specific data (e.g., natural language generation, text summarization, etc.). Scaling the model and dataset size has helped improve the performance of LLMs, but unfortunately, this also lead to highly prohibitive computational costs. Pre-training LLMs often require orders of magnitude more FLOPs than fine-tuning and the model capacity often remains the same between the two phases. To achieve training efficiency w.r.t training FLOPs, we propose to decouple the model capacity between the two phases and introduce Sparse Pre-training and Dense Fine-tuning (SPDF). In this work, we show the benefits of using unstructured weight sparsity to train only a subset of weights during pre-training (Sparse Pre-training) and then recover the representational capacity by allowing the zeroed weights to learn (Dense Fine-tuning). We demonstrate that we can induce up to 75% sparsity into a 1.3B parameter GPT-3 XL model resulting in a 2.5x reduction in pre-training FLOPs, without a significant loss in accuracy on the downstream tasks relative to the dense baseline. By rigorously evaluating multiple downstream tasks, we also establish a relationship between sparsity, task complexity and dataset size. Our work presents a promising direction to train large GPT models at a fraction of the training FLOPs using weight sparsity, while retaining the benefits of pre-trained textual representations for downstream tasks.
Large Language Models (LLMs) have demonstrated remarkable abilities across numerous disciplines, primarily assessed through tasks in language generation, knowledge utilization, and complex reasoning. However, their alignment with human emotions and values, which is critical for real-world applications, has not been systematically evaluated. Here, we assessed LLMs' Emotional Intelligence (EI), encompassing emotion recognition, interpretation, and understanding, which is necessary for effective communication and social interactions. Specifically, we first developed a novel psychometric assessment focusing on Emotion Understanding (EU), a core component of EI, suitable for both humans and LLMs. This test requires evaluating complex emotions (e.g., surprised, joyful, puzzled, proud) in realistic scenarios (e.g., despite feeling underperformed, John surprisingly achieved a top score). With a reference frame constructed from over 500 adults, we tested a variety of mainstream LLMs. Most achieved above-average EQ scores, with GPT-4 exceeding 89% of human participants with an EQ of 117. Interestingly, a multivariate pattern analysis revealed that some LLMs apparently did not reply on the human-like mechanism to achieve human-level performance, as their representational patterns were qualitatively distinct from humans. In addition, we discussed the impact of factors such as model size, training method, and architecture on LLMs' EQ. In summary, our study presents one of the first psychometric evaluations of the human-like characteristics of LLMs, which may shed light on the future development of LLMs aiming for both high intellectual and emotional intelligence. Project website: //emotional-intelligence.github.io/
Large Language Models (LLMs) have shown excellent generalization capabilities that have led to the development of numerous models. These models propose various new architectures, tweaking existing architectures with refined training strategies, increasing context length, using high-quality training data, and increasing training time to outperform baselines. Analyzing new developments is crucial for identifying changes that enhance training stability and improve generalization in LLMs. This survey paper comprehensively analyses the LLMs architectures and their categorization, training strategies, training datasets, and performance evaluations and discusses future research directions. Moreover, the paper also discusses the basic building blocks and concepts behind LLMs, followed by a complete overview of LLMs, including their important features and functions. Finally, the paper summarizes significant findings from LLM research and consolidates essential architectural and training strategies for developing advanced LLMs. Given the continuous advancements in LLMs, we intend to regularly update this paper by incorporating new sections and featuring the latest LLM models.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191