analiza aiqbench
This research challenges the prevailing assumption that linear decodability of world properties like locations or emotions proves LLMs possess internal world models, demonstrating instead that static word embeddings capture much of this signal through distributional associations alone. While transformers retain advantages in coarse sorting entities by century and country, the study suggests that decodability metrics cannot definitively distinguish between genuine representational capacity and pre-existing statistical correlations in training data. This distinction is critical for accurately interpreting claims about AI cognition and the nature of knowledge representation in large language models.
World Properties without World Models: Distributional Associations and the Interpretation of Decoding Results from Language Models
arXiv:2603.04317v2 Announce Type: replace Abstract: A growing literature shows that variables can be linearly decoded from the activations of large language models (LLMs). These range from properties of the world, such as the locations of cities and the lifetimes of historical figures, to emotions and pain. Such findings are often taken as evidence that language models go beyond surface text statistics and form internal models of the world. We show that static word embeddings (fixed, context-insensitive representations learned from corpus statistics) of the same or matched stimuli support much of the same decoding. Across four published cases (place, time, pain and emotion), static vectors predict coordinates and year of death (R^2 = 0.42-0.59), separate pain from matched control sentences (held-out AUC 0.85-0.88), and classify twelve emotions in stories written to avoid naming them (AUC 0.84-0.88). Because static embeddings assign each word a single, context-independent vector, these results are a lower bound on what word associations alone can support. The LLMs retain clear advantages on representational tests, and causal and behavioral findings remain outside the scope of the baseline. On the original authors' entities, where we reproduce their Llama-2 results, the transformer's advantage lies mostly in placing historical figures in the right century and places in the right country, coarse sorting that richer word associations would be expected to improve; within those groups every representation orders items poorly. Static vectors for disambiguated Wikipedia entities, which carry the associations of a particular place or person rather than of the words in its name, close most of the remaining gap, matching Pythia-2.8B on coordinates and Llama-2-7B on year of death. These results indicate that decodability alone cannot distinguish a representation of a property from information already available in fixed distributional associations.
Abstract:A growing literature shows that variables can be linearly decoded from the activations of large language models (LLMs). These range from properties of the world, such as the locations of cities and the lifetimes of historical figures, to emotions and pain. Such findings are often taken as evidence that language models go beyond surface text statistics and form internal models of the world. We show that static word embeddings (fixed, context-insensitive representations learned from corpus statistics) of the same or matched stimuli support much of the same decoding. Across four published cases (place, time, pain and emotion), static vectors predict coordinates and year of death (R^2 = 0.42-0.59), separate pain from matched control sentences (held-out AUC 0.85-0.88), and classify twelve emotions in stories written to avoid naming them (AUC 0.84-0.88). Because static embeddings assign each word a single, context-independent vector, these results are a lower bound on what word associations alone can support. The LLMs retain clear advantages on representational tests, and causal and behavioral findings remain outside the scope of the baseline. On the original authors' entities, where we reproduce their Llama-2 results, the transformer's advantage lies mostly in placing historical figures in the right century and places in the right country, coarse sorting that richer word associations would be expected to improve; within those groups every representation orders items poorly. Static vectors for disambiguated Wikipedia entities, which carry the associations of a particular place or person rather than of the words in its name, close most of the remaining gap, matching Pythia-2.8B on coordinates and Llama-2-7B on year of death. These results indicate that decodability alone cannot distinguish a representation of a property from information already available in fixed distributional associations.
22 pages, 3 figures, 10 tables. Substantially revised to include analyses of full released Gurnee & Tegmark datasets with Llama-2 and Pythia comparisons; replaces the earlier 100-city, 194-figure analysis; also includes entity-level vectors, and pain and emotion decoding
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2603.04317 [cs.CL]
(or arXiv:2603.04317v2 [cs.CL] for this version)
Text integral sub licența deschisă a publicației · 339 cuvinte
Subiecte conexe
The Entangled Dimensions of Decision Theory
LessWrong · 30 iul. 2026
Research directions in condensation: varieties of objectivity
AI Alignment Forum · 28 iul. 2026
The PPA: A New Approach to Anthropics
LessWrong · 14 oct. 2026
The basic argument for making AI better at moral philosophy
LessWrong · 14 oct. 2026