LLMs struggle with Semantic Inertia: the inability to inhibit pre-trained priors (e.g., “Lava is Dangerous”) when dynamic, in-context rules contradict them. We probe this phenomenon using , where physical laws are mutable text rules, enabling precise evaluation of models’ ability to override learned priors when rules change. We quantatively observe that larger models can exhibit inverse scaling: they perform worse than smaller models when natural language reasoning requires suppressing pre-trained associations (e.g., accepting “Lava is Safe”). Our analysis attributes this to natural language encoding, which entangles descriptive semantics and logical rules, leading to persistent hallucinations of familiar physics despite explicit contradictory rules. Here we show that representing dynamics as executable code, rather than descriptive text, reverses this trend and enables effective prior inhibition. We introduce LCV, which fine-tunes models on counterfactual pairs and identifies states with contradictory rules, thereby forcing attention to logical constraints rather than visual semantics. This training-time approach outperforms expensive inference-time search methods in both efficiency and accuracy. Our results demonstrate that representation fundamentally determines whether scaling improves or impairs contextual reasoning. This challenges the assumption that larger models are universally better, with implications for domains that require dynamic overriding of learned priors.

Instructions are never complete specifications of the behavior they request: speakers say only what their listeners cannot supply for themselves. An agent told to “prepare an apple” must decide unaided whether to wash or cut it and where to place it. Such choices differ between people and are seldom stated, only enacted. Multimodal foundation models render instructions into grounded plans, yet settle these residual choices by generic priors rather than by evidence about the person served. Whether an agent can instead recover such constraints from prior behavior has not been separated from the easier question of planning once they are supplied. Here we show the failure to be largely one of acquisition rather than planning, and that stating the inferred preference before acting recovers much of what is lost. Preference-based Planning (PBP) benchmarks 290 preferences over action parameters, placement policies, and temporal orderings, asking agents to infer from a few unlabeled demonstrations which constraint governs a new underspecified instruction. Given the preference, the strongest models plan categorical tasks at near-zero error; required to infer it, performance drops sharply, and without demonstrations accuracy falls by over a third. Because a verbalized rule keeps intent while discarding perceptual particulars, preferences so represented transfer to unfamiliar scenes, whereas implicit imitation stays bound to the rooms it was learned in. The bottleneck lies not in generating actions but in the pragmatic inference preceding them, for which language proves an effective medium.