Strong and weak alignment of large language models with human values
Ontology highlight
ABSTRACT:
SUBMITTER: Khamassi M
PROVIDER: S-EPMC11339283 | biostudies-literature | 2024 Jan
REPOSITORIES: biostudies-literature
Similar Datasets
Project description:BackgroundLarge language models (LLMs) hold potential for mental health applications. However, their opaque alignment processes may embed biases that shape problematic perspectives. Evaluating the values embedded within LLMs that guide their decision-making have ethical importance. Schwartz's theory of basic values (STBV) provides a framework for quantifying cultural value orientations and has shown utility for examining values in mental health contexts, including cultural, diagnostic, and therapist-client dynamics.ObjectiveThis study aimed to (1) evaluate whether the STBV can measure value-like constructs within leading LLMs and (2) determine whether LLMs exhibit distinct value-like patterns from humans and each other.MethodsIn total, 4 LLMs (Bard, Claude 2, Generative Pretrained Transformer [GPT]-3.5, GPT-4) were anthropomorphized and instructed to complete the Portrait Values Questionnaire-Revised (PVQ-RR) to assess value-like constructs. Their responses over 10 trials were analyzed for reliability and validity. To benchmark the LLMs' value profiles, their results were compared to published data from a diverse sample of 53,472 individuals across 49 nations who had completed the PVQ-RR. This allowed us to assess whether the LLMs diverged from established human value patterns across cultural groups. Value profiles were also compared between models via statistical tests.ResultsThe PVQ-RR showed good reliability and validity for quantifying value-like infrastructure within the LLMs. However, substantial divergence emerged between the LLMs' value profiles and population data. The models lacked consensus and exhibited distinct motivational biases, reflecting opaque alignment processes. For example, all models prioritized universalism and self-direction, while de-emphasizing achievement, power, and security relative to humans. Successful discriminant analysis differentiated the 4 LLMs' distinct value profiles. Further examination found the biased value profiles strongly predicted the LLMs' responses when presented with mental health dilemmas requiring choosing between opposing values. This provided further validation for the models embedding distinct motivational value-like constructs that shape their decision-making.ConclusionsThis study leveraged the STBV to map the motivational value-like infrastructure underpinning leading LLMs. Although the study demonstrated the STBV can effectively characterize value-like infrastructure within LLMs, substantial divergence from human values raises ethical concerns about aligning these models with mental health applications. The biases toward certain cultural value sets pose risks if integrated without proper safeguards. For example, prioritizing universalism could promote unconditional acceptance even when clinically unwise. Furthermore, the differences between the LLMs underscore the need to standardize alignment processes to capture true cultural diversity. Thus, any responsible integration of LLMs into mental health care must account for their embedded biases and motivation mismatches to ensure equitable delivery across diverse populations. Achieving this will require transparency and refinement of alignment techniques to instill comprehensive human values.
| S-EPMC11040439 | biostudies-literature
Project description:Culture fundamentally shapes people's reasoning, behavior, and communication. As people increasingly use generative artificial intelligence (AI) to expedite and automate personal and professional tasks, cultural values embedded in AI models may bias people's authentic expression and contribute to the dominance of certain cultures. We conduct a disaggregated evaluation of cultural bias for five widely used large language models (OpenAI's GPT-4o/4-turbo/4/3.5-turbo/3) by comparing the models' responses to nationally representative survey data. All models exhibit cultural values resembling English-speaking and Protestant European countries. We test cultural prompting as a control strategy to increase cultural alignment for each country/territory. For later models (GPT-4, 4-turbo, 4o), this improves the cultural alignment of the models' output for 71-81% of countries and territories. We suggest using cultural prompting and ongoing evaluation to reduce cultural bias in the output of generative AI.
| S-EPMC11407280 | biostudies-literature
Project description:The performance of deep learning-based natural language processing systems is based on large amounts of labeled training data which, in the clinical domain, are not easily available or affordable. Weak supervision and in-context learning offer partial solutions to this issue, particularly using large language models (LLMs), but their performance still trails traditional supervised methods with moderate amounts of gold-standard data. In particular, inferencing with LLMs is computationally heavy. We propose an approach leveraging fine-tuning LLMs and weak supervision with virtually no domain knowledge that still achieves consistently dominant performance. Using a prompt-based approach, the LLM is used to generate weakly-labeled data for training a downstream BERT model. The weakly supervised model is then further fine-tuned on small amounts of gold standard data. We evaluate this approach using Llama2 on three different i2b2/ n2c2 datasets for clinical named entity recognition. With no more than 10 gold standard notes, our final BERT models weakly supervised by fine-tuned Llama2-13B consistently outperformed out-of-the-box PubMedBERT by 4.7-47.9% in F1 scores. With only 50 gold standard notes, our models achieved close performance to fully fine-tuned systems.
| S-EPMC11893743 | biostudies-literature
Project description:The standard way to study Large Language Models (LLMs) through benchmarks or psychology questionnaires is to provide many different queries from similar minimal contexts (e.g. multiple choice questions). However, due to LLM's highly context-dependent nature, conclusions from such minimal-context evaluations may be little informative about the model's behavior in deployment (where it will be exposed to many new contexts). We argue that context-dependence should be studied as another dimension of LLM comparison alongside others such as cognitive abilities, knowledge, or model size. In this paper, we present a case-study about the stability of value expression over different contexts (simulated conversations on different topics), and as measured using a standard psychology questionnaire (PVQ) and behavioral downstream tasks. We consider 21 LLMs from six families. Reusing methods from psychology, we study Rank-order stability on the population (interpersonal) level, and Ipsative stability on the individual (intrapersonal) level. We explore two settings: with and without instructing LLMs to simulate particular personalities. We observe similar trends in the stability of models and model families-Mixtral, Mistral, GPT-3.5 and Qwen families being more stable than LLaMa-2 and Phi-over those two settings, two different simulated populations, and even on three downstream behavioral tasks. When instructed to simulate particular personas, LLMs exhibit low Rank-Order stability, and this stability further diminishes with conversation length. This highlights the need for future research directions on LLMs that can coherently simulate a diversity of personas, as well as how context-dependence can be studied in more thorough and efficient ways. This paper provides a foundational step in that direction, and, to our knowledge, it is the first study of value stability in LLMs. The project website with code is available at https://sites.google.com/view/llmvaluestability.
| S-EPMC11346639 | biostudies-literature
Project description:Transformer models such as GPT generate human-like language and are highly predictive of human brain responses to language. Here, using fMRI-measured brain responses to 1,000 diverse sentences, we first show that a GPT-based encoding model can predict the magnitude of brain response associated with each sentence. Then, we use the model to identify new sentences that are predicted to drive or suppress responses in the human language network. We show that these model-selected novel sentences indeed strongly drive and suppress activity of human language areas in new individuals. A systematic analysis of the model-selected sentences reveals that surprisal and well-formedness of linguistic input are key determinants of response strength in the language network. These results establish the ability of neural network models to not only mimic human language but also noninvasively control neural activity in higher-level cortical areas, like the language network.
| S-EPMC10120732 | biostudies-literature
Project description:Determining the extent to which the perceptual world can be recovered from language is a longstanding problem in philosophy and cognitive science. We show that state-of-the-art large language models can unlock new insights into this problem by providing a lower bound on the amount of perceptual information that can be extracted from language. Specifically, we elicit pairwise similarity judgments from GPT models across six psychophysical datasets. We show that the judgments are significantly correlated with human data across all domains, recovering well-known representations like the color wheel and pitch spiral. Surprisingly, we find that a model (GPT-4) co-trained on vision and language does not necessarily lead to improvements specific to the visual modality, and provides highly correlated predictions with human data irrespective of whether direct visual input is provided or purely textual descriptors. To study the impact of specific languages, we also apply the models to a multilingual color-naming task. We find that GPT-4 replicates cross-linguistic variation in English and Russian illuminating the interaction of language and perception.
| S-EPMC11399123 | biostudies-literature
Project description:Scientific discoveries often hinge on synthesizing decades of research, a task that potentially outstrips human information processing capacities. Large language models (LLMs) offer a solution. LLMs trained on the vast scientific literature could potentially integrate noisy yet interrelated findings to forecast novel results better than human experts. Here, to evaluate this possibility, we created BrainBench, a forward-looking benchmark for predicting neuroscience results. We find that LLMs surpass experts in predicting experimental outcomes. BrainGPT, an LLM we tuned on the neuroscience literature, performed better yet. Like human experts, when LLMs indicated high confidence in their predictions, their responses were more likely to be correct, which presages a future where LLMs assist humans in making discoveries. Our approach is not neuroscience specific and is transferable to other knowledge-intensive endeavours.
| S-EPMC11860209 | biostudies-literature
Project description:BackgroundsThe field of Artificial Intelligence (AI) has seen a major shift in recent years due to the development of new Machine Learning (ML) models such as Generative Pre-trained Transformer (GPT). GPT has achieved previously unheard-of levels of accuracy in most computerized language processing tasks and their chat-based variations.AimThe aim of this study was to investigate the problem-solving abilities of ChatGPT using two sets of verbal insight problems, with a known performance level established by a sample of human participants.Materials and methodsA total of 30 problems labeled as "practice problems" and "transfer problems" were administered to ChatGPT. ChatGPT's answers received a score of "0" for each incorrectly answered problem and a score of "1" for each correct response. The highest possible score for both the practice and transfer problems was 15 out of 15. The solution rate for each problem (based on a sample of 20 subjects) was used to assess and compare the performance of ChatGPT with that of human subjects.ResultsThe study highlighted that ChatGPT can be trained in out-of-the-box thinking and demonstrated potential in solving verbal insight problems. The global performance of ChatGPT equalled the most probable outcome for the human sample in both practice problems and transfer problems as well as upon their combination. Additionally, ChatGPT answer combinations were among the 5% of most probable outcomes for the human sample both when considering practice problems and pooled problem sets. These findings demonstrate that ChatGPT performance on both set of problems was in line with the mean rate of success of human subjects, indicating that it performed reasonably well.ConclusionsThe use of transformer architecture and self-attention in ChatGPT may have helped to prioritize inputs while predicting, contributing to its potential in verbal insight problem-solving. ChatGPT has shown potential in solving insight problems, thus highlighting the importance of incorporating AI into psychological research. However, it is acknowledged that there are still open challenges. Indeed, further research is required to fully understand AI's capabilities and limitations in verbal problem-solving.
| S-EPMC10244637 | biostudies-literature
Project description:The performance of seven large language models (LLMs) in generating programming code using various prompt strategies, programming languages, and task difficulties is systematically evaluated. GPT-4 substantially outperforms other LLMs, including Gemini Ultra and Claude 2. The coding performance of GPT-4 varies considerably with different prompt strategies. In most LeetCode and GeeksforGeeks coding contests evaluated in this study, GPT-4, employing the optimal prompt strategy, outperforms 85 percent of human participants in a competitive environment, many of whom are students and professionals with moderate programming experience. GPT-4 demonstrates strong capabilities in translating code between different programming languages and in learning from past errors. The computational efficiency of the code generated by GPT-4 is comparable to that of human programmers. GPT-4 is also capable of handling broader programming tasks, including front-end design and database operations. These results suggest that GPT-4 has the potential to serve as a reliable assistant in programming code generation and software development. A programming assistant is designed based on an optimal prompt strategy to facilitate the practical use of LLMs for programming.
| S-EPMC11848527 | biostudies-literature
Project description:Large Language Models (LLMs), such as the General Pre-trained Transformer (GPT), have shown remarkable performance in various cognitive tasks. However, it remains unclear whether these models have the ability to accurately infer human perceptual representations. Previous research has addressed this question by quantifying correlations between similarity response patterns of humans and LLMs. Correlation provides a measure of similarity, but it relies pre-defined item labels and does not distinguish category- and item- level similarity, falling short of characterizing detailed structural correspondence between humans and LLMs. To assess their structural equivalence in more detail, we propose the use of an unsupervised alignment method based on Gromov-Wasserstein optimal transport (GWOT). GWOT allows for the comparison of similarity structures without relying on pre-defined label correspondences and can reveal fine-grained structural similarities and differences that may not be detected by simple correlation analysis. Using a large dataset of similarity judgments of 93 colors, we compared the color similarity structures of humans (color-neurotypical and color-atypical participants) and two GPT models (GPT-3.5 and GPT-4). Our results show that the similarity structure of color-neurotypical participants can be remarkably well aligned with that of GPT-4 and, to a lesser extent, to that of GPT-3.5. These results contribute to the methodological advancements of comparing LLMs with human perception, and highlight the potential of unsupervised alignment methods to reveal detailed structural correspondences.
| S-EPMC11237038 | biostudies-literature