Multiple Choice Questions: Reasoning Makes Large Language Models (LLMs) More Self-Confident Even When They Are Wrong

One of the most widely used methods to evaluate LLMs are Multiple Choice Question (MCQ) tests. MCQ benchmarks enable the testing of LLM knowledge on almost any topic at scale as the results can be processed automatically. To help the LLM answer, a few examples called few shots can be included in the prompt. Moreover, the LLM can be asked to answer the question directly with the selected option or to first provide the reasoning and then the selected answer, which is known as chain of thought. In addition to checking whether the selected answer is correct, the evaluation can look at the LLM-estimated probability of its response as an indication of the confidence of the LLM in the response. In this paper, we study how the LLM confidence in its answer depends on whether the model has been asked to answer directly or to provide the reasoning before answering. The results of the evaluation of questions on a wide range of topics in seven different models show that LLMs are more confident in their answers when they provide reasoning before the answer. This occurs regardless of whether the selected answer is correct. Our hypothesis is that this behavior is due to the reasoning that modifies the probability of the selected answer, as the LLM predicts the answer based on the input question and the reasoning that supports the selection made. Therefore, LLM estimated probabilities seem to have intrinsic limitations that should be understood in order to use them in evaluation procedures. Interestingly, the same behavior has been observed in humans, for whom explaining an answer increases confidence in its correctness.

Why Do Large Language Models (LLMs) Struggle to Count Letters?

Large Language Models (LLMs) have achieved unprecedented performance on many complex tasks, being able, for example, to answer questions on almost any topic. However, they struggle with other simple tasks, such as counting the occurrences of letters in a word, as illustrated by the inability of many LLMs to count the number of "r" letters in "strawberry". Several works have studied this problem and linked it to the tokenization used by LLMs, to the intrinsic limitations of the attention mechanism, or to the lack of character-level training data. In this paper, we conduct an experimental study to evaluate the relations between the LLM errors when counting letters with 1) the frequency of the word and its components in the training dataset and 2) the complexity of the counting operation. We present a comprehensive analysis of the errors of LLMs when counting letter occurrences by evaluating a representative group of models over a large number of words. The results show a number of consistent trends in the models evaluated: 1) models are capable of recognizing the letters but not counting them; 2) the frequency of the word and tokens in the word does not have a significant impact on the LLM errors; 3) there is a positive correlation of letter frequency with errors, more frequent letters tend to have more counting errors, 4) the errors show a strong correlation with the number of letters or tokens in a word and 5) the strongest correlation occurs with the number of letters with counts larger than one, with most models being unable to correctly count words in which letters appear more than twice.

Adaptive Resolution Inference (ARI): Energy-Efficient Machine Learning for Internet of Things

The implementation of machine learning in Internet of Things devices poses significant operational challenges due to limited energy and computation resources. In recent years, significant efforts have been made to implement simplified ML models that can achieve reasonable performance while reducing computation and energy, for example by pruning weights in neural networks, or using reduced precision for the parameters and arithmetic operations. However, this type of approach is limited by the performance of the ML implementation, i.e., by the loss for example in accuracy due to the model simplification. In this article, we present adaptive resolution inference (ARI), a novel approach that enables to evaluate new tradeoffs between energy dissipation and model performance in ML implementations. The main principle of the proposed approach is to run inferences with reduced precision (quantization) and use the margin over the decision threshold to determine if either the result is reliable, or the inference must run with the full model. The rationale is that quantization only introduces small deviations in the inference scores, such that if the scores have a sufficient margin over the decision threshold, it is unlikely that the full model would have a different result. Therefore, we can run the quantized model first, and only when the scores do not have a sufficient margin, the full model is run. This enables most inferences to run with the reduced precision model and only a small fraction requires the full model, so significantly reducing computation and energy while not affecting model performance. The proposed ARI approach is presented, analyzed in detail, and evaluated using different data sets for floating-point and stochastic computing implementations. The results show that ARI can significantly reduce the energy for inference in different configurations with savings between 40% and 85%.

Dependability Evaluation of Stable Diffusion with Soft Errors on the Model Parameters

Stable Diffusion is a popular Transformer-based model for image generation from text; it applies an image information creator to the input text and the visual knowledge is added in a step-by-step fashion to create an image that corresponds to the input text. However, this diffusion process can be corrupted by errors from the underlying hardware, which are especially relevant for implementations at the nanoscales. In this paper, the dependability of Stable Diffusion is studied focusing on soft errors in the memory that stores the model parameters; specifically, errors are injected into some critical layers of the Transformer in different blocks of the image information creator, to evaluate their impact on model performance. The simulations results reveal several conclusions: 1) errors on the down blocks of the creator have a larger impact on the quality of the generated images than those on the up blocks, while the errors on middle block have negligible effect; 2) errors on the self-attention (SA) layers have larger impact on the results than those on the cross-attention (CA) layers; 3) for CA layers, errors on deeper levels result in a larger impact; 4) errors on blocks at the first levels tend to introduce noise in the image, and those on deep layers tend to introduce large colored blocks. These results provide an initial understanding of the impact of errors on Stable Diffusion.

Concurrent Linguistic Error Detection (CLED) for Large Language Models

The wide adoption of Large language models (LLMs) makes their dependability a pressing concern. Detection of errors is the first step to mitigating their impact on a system and thus, efficient error detection for LLMs is an important issue. In many settings, the LLM is considered as a black box with no access to the internal nodes; this prevents the use of many error detection schemes that need access to the model's internal nodes. An interesting observation is that the output of LLMs in error-free operation should be valid and normal text. Therefore, when the text is not valid or differs significantly from normal text, it is likely that there is an error. Based on this observation we propose to perform Concurrent Linguistic Error Detection (CLED); this scheme extracts some linguistic features of the text generated by the LLM and feeds them to a concurrent classifier that detects errors. Since the proposed error detection mechanism only relies on the outputs of the model, then it can be used on LLMs in which there is no access to the internal nodes. The proposed CLED scheme has been evaluated on the T5 model when used for news summarization and on the OPUS-MT model when used for translation. In both cases, the same set of linguistic features has been used for error detection to illustrate the applicability of the proposed scheme beyond a specific case. The results show that CLED can detect most of the errors at a low overhead penalty. The use of the concurrent classifier also enables a trade-off between error detection effectiveness and its associated overhead, so providing flexibility to a designer.

Open Source Conversational LLMs do not know most Spanish words

The growing interest in Large Language Models (LLMs) and in particular in conversational models with which users can interact has led to the development of a large number of open-source chat LLMs. These models are evaluated on a wide range of benchmarks to assess their capabilities in answering questions or solving problems on almost any possible topic or to test their ability to reason or interpret texts. Instead, the evaluation of the knowledge that these models have of the languages has received much less attention. For example, the words that they can recognize and use in different languages. In this paper, we evaluate the knowledge that open-source chat LLMs have of Spanish words by testing a sample of words in a reference dictionary. The results show that open-source chat LLMs produce incorrect meanings for an important fraction of the words and are not able to use most of the words correctly to write sentences with context. These results show how Spanish is left behind in the open-source LLM race and highlight the need to push for linguistic fairness in conversational LLMs ensuring that they provide similar performance across languages.

Beware of Words: Evaluating the Lexical Richness of Conversational Large Language Models

The performance of conversational Large Language Models (LLMs) in general, and of ChatGPT in particular, is currently being evaluated on many different tasks, from logical reasoning or maths to answering questions on a myriad of topics. Instead, much less attention is being devoted to the study of the linguistic features of the texts generated by these LLMs. This is surprising since LLMs are models for language, and understanding how they use the language is important. Indeed, conversational LLMs are poised to have a significant impact on the evolution of languages as they may eventually dominate the creation of new text. This means that for example, if conversational LLMs do not use a word it may become less and less frequent and eventually stop being used altogether. Therefore, evaluating the linguistic features of the text they produce and how those depend on the model parameters is the first step toward understanding the potential impact of conversational LLMs on the evolution of languages. In this paper, we consider the evaluation of the lexical richness of the text generated by LLMs and how it depends on the model parameters. A methodology is presented and used to conduct a comprehensive evaluation of lexical richness using ChatGPT as a case study. The results show how lexical richness depends on the version of ChatGPT and some of its parameters, such as the presence penalty, or on the role assigned to the model. The dataset and tools used in our analysis are released under open licenses with the goal of drawing the much-needed attention to the evaluation of the linguistic features of LLM-generated text.

The continued usefulness of vocabulary tests for evaluating large language models

In their seminal article on semantic vectors, Landauer and Dumain (1997) proposed testing the quality of AI language models with a challenging vocabulary test. We show that their Test of English as a Foreign Language (TOEFL) test remains informative for contemporary major language models, since none of the models was perfect and made errors on divergent items. The TOEFL test consists of target words with four alternatives to choose from. We further tested the models on a Yes/No test that requires distinguishing between existing words and made-up nonwords. The models performed significantly worse on the nonword items, in line with other observations that current major language models provide non-existent information. The situation was worse when we generalized the tests to Spanish. Here, most models gave meanings/translations for the majority of random letter sequences. On the plus side, the best models began to perform quite well, and they also pointed to nonwords that were unknown to the test participants but can be found in dictionaries.

How many words does ChatGPT know? The answer is ChatWords

The introduction of ChatGPT has put Artificial Intelligence (AI) Natural Language Processing (NLP) in the spotlight. ChatGPT adoption has been exponential with millions of users experimenting with it in a myriad of tasks and application domains with impressive results. However, ChatGPT has limitations and suffers hallucinations, for example producing answers that look plausible but they are completely wrong. Evaluating the performance of ChatGPT and similar AI tools is a complex issue that is being explored from different perspectives. In this work, we contribute to those efforts with ChatWords, an automated test system, to evaluate ChatGPT knowledge of an arbitrary set of words. ChatWords is designed to be extensible, easy to use, and adaptable to evaluate also other NLP AI tools. ChatWords is publicly available and its main goal is to facilitate research on the lexical knowledge of AI tools. The benefits of ChatWords are illustrated with two case studies: evaluating the knowledge that ChatGPT has of the Spanish lexicon (taken from the official dictionary of the "Real Academia Espa\~nola") and of the words that appear in the Quixote, the well-known novel written by Miguel de Cervantes. The results show that ChatGPT is only able to recognize approximately 80% of the words in the dictionary and 90% of the words in the Quixote, in some cases with an incorrect meaning. The implications of the lexical knowledge of NLP AI tools and potential applications of ChatWords are also discussed providing directions for further work on the study of the lexical knowledge of AI tools.

Playing with Words: Comparing the Vocabulary and Lexical Richness of ChatGPT and Humans

The introduction of Artificial Intelligence (AI) generative language models such as GPT (Generative Pre-trained Transformer) and tools such as ChatGPT has triggered a revolution that can transform how text is generated. This has many implications, for example, as AI-generated text becomes a significant fraction of the text, would this have an effect on the language capabilities of readers and also on the training of newer AI tools? Would it affect the evolution of languages? Focusing on one specific aspect of the language: words; will the use of tools such as ChatGPT increase or reduce the vocabulary used or the lexical richness? This has implications for words, as those not included in AI-generated content will tend to be less and less popular and may eventually be lost. In this work, we perform an initial comparison of the vocabulary and lexical richness of ChatGPT and humans when performing the same tasks. In more detail, two datasets containing the answers to different types of questions answered by ChatGPT and humans, and a third dataset in which ChatGPT paraphrases sentences and questions are used. The analysis shows that ChatGPT tends to use fewer distinct words and lower lexical richness than humans. These results are very preliminary and additional datasets and ChatGPT configurations have to be evaluated to extract more general conclusions. Therefore, further research is needed to understand how the use of ChatGPT and more broadly generative AI tools will affect the vocabulary and lexical richness in different types of text and languages.