In-Context Reorganization in Large Language Models - A Behavioral Approach through a Forced-Choice Protocol

Human memory research has established that later information can reorganize an earlier memory, changing not merely its accessibility but what the earlier event is taken to mean. Whether large language models do something structurally comparable has been difficult to establish. The available lines of evidence cannot confirm it: benchmarks give overall scores, mechanistic interpretability looks at circuits and self-report is unreliable. We isolate this process in behavior with a forced-choice reorganization paradigm. Each item pairs a first-person account containing one ambiguous target phrase with a continuation that either recontextualizes that phrase (reveal) or resolves the situation ordinarily (neutral), followed by a forced choice between two glosses; a reveal-consistent choice is correct only if reorganization has occurred. Across 38 trials administered to 18 models from ten developers at temperature 0, seven of ten narrative items show a reveal-versus-neutral shift in the predicted direction surviving Holm correction, with a large pooled association between condition and choice, $\chi^2(1, N=538)=161.5$. Two items show no discrimination, and they share confounds that warrant further study. Additionally we document a rare but well-defined dissociation in which a model's stated reasoning affirms the reveal while its stated choice contradicts it. Confidence drops when there's no basis to decide, yet ``no basis'' is seldom chosen. The paradigm shows that functional reorganization can be measured behaviorally and provides a baseline against which introspective and representational measures of the same process can be compared.