Content, Not Format: What Drives Analysis-Mode in Language Model Agents, and How Grounding Recovers Execution

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

A prior report observed that a small language model, when a task is prefixed with astructured description of a recurring decision, stops acting and instead analyzes the descrip-tion. That study used one model and one scenario, and it did not separate the description’scontent from its form. This follow-up adds the controls it lacked. We prefix a task witheither a three-axis descriptive “glyph” or an information-matched imperative rule, acrossfour decision classes, and we sweep the descriptive prefix across four open-weight models onthe original class. The shift away from action generalizes across models, and it is a con-tent effect, not a format effect: the matched imperative suppresses action as much as theglyph, and the descriptive format shows no consistent advantage over a plain rule with thesame content. Control prefixes (a content-scrambled glyph, a wrong-class glyph) show thesuppression mechanism is decision-class-dependent. We then vary two things independentlywhile holding the content fixed: form (descriptive versus directive) and grounding (a domain-free statement versus a domain-specific one). Grounding is the lever. A domain-specific aidrestores execution where a domain-free one suppresses it, in both forms. Whether a domain-specific description suffices as well as a domain-specific command is under-determined here:the command reaches ceiling in all four model-and-class cells, so the two separate onlywhere the description falls short of ceiling, which happens once, on the smaller model andthe harder task, and there the command does better. All runs use one local raw-completioninference rig; reproducing them needs a 24 GB GPU and no paid interface. We state thescope and the measurement limits, including the cells where a baseline already acts and noeffect is measurable.