LLMs show a directional asymmetry in educational control: they reliably increase cognitive demand but struggle to lower it¶

Claim · All claims
Evidence · 1 study · causal r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 1 of 1 entry) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Under both general difficulty prompts and targeted Bloom's levels, the evaluated models more easily add cognitively demanding operations than remove or replace them with lower-level operations. → Yi Zhang & Julia Rayz 2026

Evidence¶

Yi Zhang & Julia Rayz 2026¶

Yi Zhang & Julia Rayz. (2026). From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs. Preprint. https://arxiv.org/abs/2607.08009 doi:10.48550/arxiv.2607.08009

q2 · i? · causal · r2

Behavioral evaluation of 2,520 programming tasks from three benchmarks, judged with an LLM-as-a-judge framework. Under the "Easier" prompt both models still produced positive OCS scores (0.715 and 0.194), and the general model achieved only a marginal negative shift of -0.134 under targeted "Lower" Bloom prompts.

"OCS scores reveal that the models more easily add cognitively demanding operations than remove or replace them with lower-level operations, under both general difficulty prompts and targeted Bloom's levels."

Studies bearing on this claim¶

Can a language model write or rewrite questions at a targeted Bloom level? This table gathers the studies behind this claim and the 5 narrower claims that bear on it: 2 distinct studies (2 causal), q2, 0 of 2 with an effect size coded. One row per study, codes as each entry records them; no effect is averaged and no verdict is drawn. Built by scripts/synthesize_claim.py from eval/synthesis/; a row that looks wrong is fixed on the claim it comes from.

Discussion¶

Related Claims¶

• Downward Bloom control is benchmark-dependent: repository-level tasks are easier to simplify than algorithmic puzzles — a narrower finding that bears on this claim
• Both models target Higher Bloom levels far more accurately than Lower levels, with the general model outperforming the coder model upward — a narrower finding that bears on this claim
• Cognitive leap is the most prevalent deviation pattern of LLM-generated questions, and fine-grained prompting amplifies upward shifts while suppressing cognitive regression — related
• A design-enactment gap in cognitive demand: 37.7% of conversations under-reached the teacher-targeted DOK level, and under-targeting increased with target ambition — related
• Layer-wise probing shows clearer middle-layer separability in the general model and a weaker, deeper separability profile in the coder model — related
• The Bloom-driven framework generalizes to social-science questions, with automated rubrics agreeing strongly with experts on usability but only modestly on Bloom level — related
• LLMs exhibit an inherent abstraction bias, consistently generating questions at higher-than-intended cognitive levels regardless of prompting strategy — related
• In an OLS regression, explicit finish lines in teacher prompts were associated with smaller DOK gaps, while explain-reasoning epistemic framing was associated with larger gaps — related
• The general and coder models deploy distinct lexical strategies when raising cognitive demand, and neither deploys simplification vocabulary downward — related