How well do language models deal with quantification? In this study, we focus on 'few'-type quantifiers, as in 'few children like toys', which might pose a particular challenge for language models because the sentence components with out the quantifier are likely to co-occur, and 'few'-type quantifiers are rare. We present 960 English sentence stimuli from two human neurolinguistic experiments to 22 autoregressive transformer models of differing sizes. Not only do all the models perform poorly on 'few'-type quantifiers, but overall the larger the model, the worse its performance. This inverse scaling is consistent with previous work suggesting that larger models increasingly reflect online rather than offline human processing, and we argue that the decreasing performance of larger models may challenge uses of language models as the basis for natural language systems.
相關內容
Large language models (LLMs) are trained to imitate humans to explain human decisions. However, do LLMs explain themselves? Can they help humans build mental models of how LLMs process different inputs? To answer these questions, we propose to evaluate $\textbf{counterfactual simulatability}$ of natural language explanations: whether an explanation can enable humans to precisely infer the model's outputs on diverse counterfactuals of the explained input. For example, if a model answers "yes" to the input question "Can eagles fly?" with the explanation "all birds can fly", then humans would infer from the explanation that it would also answer "yes" to the counterfactual input "Can penguins fly?". If the explanation is precise, then the model's answer should match humans' expectations. We implemented two metrics based on counterfactual simulatability: precision and generality. We generated diverse counterfactuals automatically using LLMs. We then used these metrics to evaluate state-of-the-art LLMs (e.g., GPT-4) on two tasks: multi-hop factual reasoning and reward modeling. We found that LLM's explanations have low precision and that precision does not correlate with plausibility. Therefore, naively optimizing human approvals (e.g., RLHF) may not be a sufficient solution.
Skills-based matching promises mobility of workers between different sectors and occupations in the labor market. In this case, job seekers can look for jobs they do not yet have experience in, but for which they do have relevant skills. Currently, there are multiple occupations with a skewed gender distribution. For skills-based matching, it is unclear if and how a shift in the gender distribution, which we call gender mobility, between occupations will be effected. It is expected that the skills-based matching approach will likely be data-driven, including computational language models and supervised learning methods. This work, first, shows the presence of gender segregation in language model-based skills representation of occupations. Second, we assess the use of these representations in a potential application based on simulated data, and show that the gender segregation is propagated by various data-driven skills-based matching models.These models are based on different language representations (bag of words, word2vec, and BERT), and distance metrics (static and machine learning-based). Accordingly, we show how skills-based matching approaches can be evaluated and compared on matching performance as well as on the risk of gender segregation. Making the gender segregation bias of models more explicit can help in generating healthy trust in the use of these models in practice.
Despite the superior performance, Large Language Models~(LLMs) require significant computational resources for deployment and use. To overcome this issue, quantization methods have been widely applied to reduce the memory footprint of LLMs as well as increasing the inference rate. However, a major challenge is that low-bit quantization methods often lead to performance degradation. It is important to understand how quantization impacts the capacity of LLMs. Different from previous studies focused on overall performance, this work aims to investigate the impact of quantization on \emph{emergent abilities}, which are important characteristics that distinguish LLMs from small language models. Specially, we examine the abilities of in-context learning, chain-of-thought reasoning, and instruction-following in quantized LLMs. Our empirical experiments show that these emergent abilities still exist in 4-bit quantization models, while 2-bit models encounter severe performance degradation on the test of these abilities. To improve the performance of low-bit models, we conduct two special experiments: (1) fine-gained impact analysis that studies which components (or substructures) are more sensitive to quantization, and (2) performance compensation through model fine-tuning. Our work derives a series of important findings to understand the impact of quantization on emergent abilities, and sheds lights on the possibilities of extremely low-bit quantization for LLMs.
Evaluating Natural Language Generation (NLG) outputs is crucial but laborious and expensive. While various automatic NLG assessment methods have been proposed, they often are quite task-specific and have to be engineered with a particular domain and attribute in mind. In this work, we propose a robust zero-shot approach to NLG evaluation using pairwise comparative judgment with open-source Large Language Models (LLMs). The motivation for this approach is that even as humans, it is easier to determine which of two options are better, than it is to independently objectively score each option. We use this insight and leverage the emergent abilities of LLMs, where we probe FlanT5 to determine which of two candidate responses is better, rather than assigning absolute scores. Our results demonstrate that comparative assessment is a more effective approach than absolute scoring, enabling smaller open-source LLMs to achieve comparable performance to larger public access APIs. We evaluate systems on both summary evaluation and dialogue response generation, and show that opensource LLMs can lead to good correlations with human scores for a range of different attributes.
Large language models (LLMs), such as GPT-3 and GPT-4, have demonstrated exceptional performance in various natural language processing tasks and have shown the ability to solve certain reasoning problems. However, their reasoning capabilities are limited and relatively shallow, despite the application of various prompting techniques. In contrast, formal logic is adept at handling complex reasoning, but translating natural language descriptions into formal logic is a challenging task that non-experts struggle with. This paper proposes a neuro-symbolic method that combines the strengths of large language models and answer set programming. Specifically, we employ an LLM to transform natural language descriptions of logic puzzles into answer set programs. We carefully design prompts for an LLM to convert natural language descriptions into answer set programs in a step by step manner. Surprisingly, with just a few in-context learning examples, LLMs can generate reasonably complex answer set programs. The majority of errors made are relatively simple and can be easily corrected by humans, thus enabling LLMs to effectively assist in the creation of answer set programs.
Molecular property prediction has gained significant attention due to its transformative potential in multiple scientific disciplines. Conventionally, a molecule graph can be represented either as a graph-structured data or a SMILES text. Recently, the rapid development of Large Language Models (LLMs) has revolutionized the field of NLP. Although it is natural to utilize LLMs to assist in understanding molecules represented by SMILES, the exploration of how LLMs will impact molecular property prediction is still in its early stage. In this work, we advance towards this objective through two perspectives: zero/few-shot molecular classification, and using the new explanations generated by LLMs as representations of molecules. To be specific, we first prompt LLMs to do in-context molecular classification and evaluate their performance. After that, we employ LLMs to generate semantically enriched explanations for the original SMILES and then leverage that to fine-tune a small-scale LM model for multiple downstream tasks. The experimental results highlight the superiority of text explanations as molecular representations across multiple benchmark datasets, and confirm the immense potential of LLMs in molecular property prediction tasks. Codes are available at \url{//github.com/ChnQ/LLM4Mol}.
There is abundant observational data in the software engineering domain, whereas running large-scale controlled experiments is often practically impossible. Thus, most empirical studies can only report statistical correlations -- instead of potentially more insightful and robust causal relations. To support analyzing purely observational data for causal relations, and to assess any differences between purely predictive and causal models of the same data, this paper discusses some novel techniques based on structural causal models (such as directed acyclic graphs of causal Bayesian networks). Using these techniques, one can rigorously express, and partially validate, causal hypotheses; and then use the causal information to guide the construction of a statistical model that captures genuine causal relations -- such that correlation does imply causation. We apply these ideas to analyzing public data about programmer performance in Code Jam, a large world-wide coding contest organized by Google every year. Specifically, we look at the impact of different programming languages on a participant's performance in the contest. While the overall effect associated with programming languages is weak compared to other variables -- regardless of whether we consider correlational or causal links -- we found considerable differences between a purely associational and a causal analysis of the very same data. The takeaway message is that even an imperfect causal analysis of observational data can help answer the salient research questions more precisely and more robustly than with just purely predictive techniques -- where genuine causal effects may be confounded.
In earlier work, we introduced the framework of language-based decisions, the core idea of which was to modify Savage's classical decision-theoretic framework by taking actions to be descriptions in some language, rather than functions from states to outcomes, as they are defined classically. Actions had the form "if psi then do(phi)", where psi and phi were formulas in some underlying language, specifying what effects would be brought about under what circumstances. The earlier work allowed only one-step actions. But, in practice, plans are typically composed of a sequence of steps. Here, we extend the earlier framework to sequential actions, making it much more broadly applicable. Our technical contribution is a representation theorem in the classical spirit: agents whose preferences over actions satisfy certain constraints can be modeled as if they are expected utility maximizers. As in the earlier work, due to the language-based specification of the actions, the representation theorem requires a construction not only of the probability and utility functions representing the agent's beliefs and preferences, but also the state and outcomes spaces over which these are defined, as well as a "selection function" which intuitively captures how agents disambiguate coarse descriptions. The (unbounded) depth of action sequencing adds substantial interest (and complexity!) to the proof.
Since the 1950s, machine translation (MT) has become one of the important tasks of AI and development, and has experienced several different periods and stages of development, including rule-based methods, statistical methods, and recently proposed neural network-based learning methods. Accompanying these staged leaps is the evaluation research and development of MT, especially the important role of evaluation methods in statistical translation and neural translation research. The evaluation task of MT is not only to evaluate the quality of machine translation, but also to give timely feedback to machine translation researchers on the problems existing in machine translation itself, how to improve and how to optimise. In some practical application fields, such as in the absence of reference translations, the quality estimation of machine translation plays an important role as an indicator to reveal the credibility of automatically translated target languages. This report mainly includes the following contents: a brief history of machine translation evaluation (MTE), the classification of research methods on MTE, and the the cutting-edge progress, including human evaluation, automatic evaluation, and evaluation of evaluation methods (meta-evaluation). Manual evaluation and automatic evaluation include reference-translation based and reference-translation independent participation; automatic evaluation methods include traditional n-gram string matching, models applying syntax and semantics, and deep learning models; evaluation of evaluation methods includes estimating the credibility of human evaluations, the reliability of the automatic evaluation, the reliability of the test set, etc. Advances in cutting-edge evaluation methods include task-based evaluation, using pre-trained language models based on big data, and lightweight optimisation models using distillation techniques.
Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy is to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191