Missing Skill or Failed Conceptual Control? A Behavior-Analytic Diagnosis of Language Model Failures on Counterfactual Analogy Problems

When a language model fails a task that combines several skills, its accuracy alone does not show why it failed. The model may lack a component skill, its responding may be controlled by the wrong features of the problem, or it may be unable to combine skills it can perform separately. This paper offers a framework, drawn from Engelmann and Carnine’s Theory of Instruction and behavior-analytic accounts of concept formation, for telling these possibilities apart, and illustrates it with a published dispute. Large language models solve letter-string analogy problems well in the standard alphabet but much worse when the same problems are written in a permuted, or counterfactual, alphabet. Lewis and Mitchell interpret the failures as evidence that the models lack robust analogical reasoning; Webb, Holyoak, and Lu attribute them to a specific difficulty with counting positions in an unfamiliar alphabet. In these terms, “the next letter” is a relational concept whose defining relation is confounded with the standard alphabetical order in ordinary text, so the failures are compatible with a failure of conceptual control; the counting interpretation corresponds to a missing component skill, or preskill; and a third possibility is that skills shown in isolation fail to combine within the problem. The two sides have each probed a component skill, but different ones: the models name the next letter in a permuted alphabet almost perfectly, yet GPT-4 rarely identifies the interval between two letters. The dispute therefore turns largely on task analysis. The paper offers two tools for empirical researchers. A task-analysis table lists the candidate component operations for each of six problem types and shows which the published probes have and have not tested. A five-step diagnostic program (1) measures each component reliably, (2) relates analogy accuracy to item-level component proficiency, (3) classifies errors by whether they follow the permuted or the standard alphabet, (4) compares remedies, including sameness-and-difference training with response-contingent feedback of differing informational content, in a factorial design with matched controls and transfer tests, and (5) gathers converging evidence from human participants. The program does not settle the debate, but it turns an argument about machine reasoning into questions that evidence can constrain.