Large models of the Gemini and GPT-4 generation generalize to new constructs rather than merely memorizing.

Judged by the state of the evidence as of February 2024. True if peer-reviewed or widely cited research from that time showed that GPT-4/Gemini-class models handle new, out-of-distribution constructs (new tasks, rules, definitions or compositions not in their training data) in ways memorisation cannot explain, and this was the prevailing expert view. Mostly true if there was solid evidence of such generalisation, with caveats such as data contamination or brittle performance in some areas. Mixed if the evidence was genuinely split, with substantial findings both of generalisation and of failures that look like memorisation. Mostly false if the evidence showed mostly memorisation or failure on truly new constructs, with only narrow cases of generalisation. False if studies from that time showed clearly that these models do not generalise to new constructs.

Set before the research began.

1. E1 Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks arXiv (ar5iv rendering)

   “We observe above-random counterfactual performance for most tasks, indicating some degree of task generalizability. However, the performance on counterfactual task variants consistently and substantially degrades relative to the performance on the default settings.”

2. E2 Using Counterfactual Tasks to Evaluate the Generality of Analogical Reasoning in Large Language Models arXiv (ar5iv rendering)

   “This work provides evidence that, despite previously reported successes of LLMs on analogical reasoning, these models lack the robustness and generality of human analogy-making.”

3. E3 Faith and fate: Limits of transformers on compositionality USC Information Sciences Institute (publication listing for NeurIPS 2023 paper) · 15 December 2023

   “Our empirical findings suggest that transformer LLMs solve compositional tasks by reducing multi-step compositional reasoning into linearized subgraph matching, without necessarily developing systematic problem-solving skills.”

4. E4 Sparks of AGI: Early Experiments with GPT-4 Emergent Mind (third-party summary of arXiv 2303.12712) · 24 April 2023

   “The paper presents evidence that GPT-4 exhibits AGI traits by excelling in language, coding, mathematics, and reasoning tasks.”

E4 summarises the 'Sparks of AGI' paper, whose authors argued that GPT-4 showed general capabilities across language, coding, mathematics and reasoning. However, E1 found above-random performance on counterfactual tasks, which is some generalisation, alongside consistent and substantial drops relative to default settings. E3, a NeurIPS 2023 paper, reports that transformer LLMs reduce compositional reasoning to subgraph matching rather than developing systematic skills, and E2 reports that these models lack the robustness and generality of human analogy-making. Taken together, the research from that time shows both generalisation and failures consistent with memorisation, which fits the 'mixed' criterion rather than a prevailing expert view of robust generalisation.