Machine intelligence in language models is partly understandable but not fully explained
Large-scale factor analysis shows machine intelligence is only partially interpretable
Computation and LanguageArtificial Intelligence
Summary
People often think language models have one big intelligence like humans do, but this study shows that’s not quite true. The authors looked at a huge amount of data from many different language models and tests and found that only part of their skill comes from a general intelligence. Also, tests that seem similar don’t always measure the same abilities, and common intelligence tests don’t fully capture this general skill. This means that focusing on improving one big intelligence in language models might not be the best approach.
What this means in practice
- •For language model developers: Improve evaluation processes by considering multiple distinct abilities rather than targeting a single general intelligence factor in language models.
- •For ai evaluation teams: Design better benchmarking suites that reflect diverse abilities instead of relying on traditional intelligence-like tests for language models.
Authors
Faiz Ghifari Haznitrama, Afrizal Hasbi Azizy, Faeyza Rishad Ardi
Abstract
A common assumption in language model development is that cognitive abilities are organized around a general, domain-free intelligence factor, like fluid intelligence in humans. This assumption is rarely tested directly, and prior attempts have done so only at a much smaller scale. We take a latent variable approach to intelligence in language models, similar to how psychometricians study psychological constructs. Performance in every specific problem set is influenced by a domain-specific and a domain-agnostic latent factor. Using factor analysis as a dimension-reduction technique, we analyzed 13,251 published evaluation scores covering 1,618 language models across 456 different text-only benchmarks. Due to the super-sparse nature of the dataset, we triangulate our analysis across different data densifiers and imputation methods. A robust pattern across different modes of bias is that 1. A general intelligence factor accounts for 70.8% of variance in model performance at our most generous estimate, and far less than that in most of our solutions, 2. Content-similar benchmarks do not necessarily cluster together, and 3. The $g$ factor is not dominated by any common theme, and there is a lack of evidence that it is well-proxied by standard "intelligence" benchmarks. Our findings go against current endeavors of defining, identifying, and targeting general intelligence as a tangible construct in language model development. This leaves the strategy of targeting a single conceptual ability without support, since the first-order abilities it would have to reach are often partially idiosyncratic and not identifiable in practice.