Fonts i Tipus d'Infiabilitat en Models de Llenguatge
Year: 2023
Type: dissertation
Abstract: [EN] Pretrained artificial intelligence models are made more human-like and human-aligned by scaling them up in resources (e.g., by increasing compute, training data and parameter size) and shaping them up with human feedback. One of the major concerns about the trustworthiness of Large Language Models (LLMs) is their unpredictability from the human point of view. First, the model can solve very challenging tasks while some questions that are considered easy by humans are completely missed by these models (difficulty discordance). Second, models that avoid answering the questions and digress in an unpredictable way may also be considered unreliable (task avoidance). Third, the same question can get solved or not depending on minor changes in the formulation of the question (prompt sensitivity). While all these cases are related to confidence estimation, they are not simply solved by having better logprob estimates from the model. These three unreliability issues are usually identified by experts or users (e.g., in social media) as the most striking or stupid, and not understanding them jeopardises the penetration of LLMs and our trust in them. Consequently, these three elements require specific analysis. This work describes three kinds of unreliability in LLMs (prompt sensitivity, difficulty discordance, and task avoidance), as well as their relations with question difficulty in LLMs. This helps identifying regions of risks. We construct a set of questions of various difficulties in the domain of mathematics (addition), combinatorics (anagrams), and geography (location), with a collection of natural prompt templates (i.e., human-understandable variations of the exact same question) in each domain. With these elements, we analyse where exactly more powerful LLMs make their improvements and whether they create areas of high reliability as a function of instance difficulty, and whether this difficulty is aligned with the human perception of difficulty in which an affirmative case will help reduce unexpected behaviour and increase acceptability. As a result, surprisingly, we have provided evidence that recent models (e.g., GPT-4) have not become more reliable for humans. By studying difficulty discordance, task avoidance and prompt sensitivity of the GPT family we discover that: (1) easy instances for humans are also easy for the models, but scaled-up shaped-up models do not secure a threshold level of difficulty below which the model is fully reliable, (2) early models often avoid user questions, but scaled-up shaped-up models tend to give an apparently sensible, yet wrong answer much more often, and (3) robustness to different natural prompts of the same question is compromised by scaling-up and shaping-up interventions, such as post-filtering moderation mechanism. All this indicates that the path towards reliable general-purpose AI needs an overhaul.
Author: Zhou, Lexin
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164275312
