Qwen3 language models easily make educational tasks more complex, but struggle to simplify them

In July 2026, researchers published a framework for measuring "educational control" of language models — the ability to preserve the educational meaning of a task while shifting its cognitive load to a specified level according to Bloom's taxonomy. Using 2520 programming tasks with the Qwen3-Next pair, the authors identified a persistent asymmetry: models are confident at complicating tasks but poor at simplifying them.

What "educational control" of a model means

Most modern benchmarks evaluate language models on a single parameter — can they correctly solve a task. In an educational context, a different ability matters: to adapt a ready-made task for a specific learner — make it simpler for a beginner or more complex for an advanced student, while preserving the educational meaning and subject matter.

The authors applied a revised Bloom's taxonomy to measure this ability — a scale of six sequential levels of cognitive complexity: "Remember," "Understand," "Apply," "Analyze," "Evaluate," and "Create." Each level describes a qualitatively different form of working with information: from memory reproduction to independent synthesis. The scale provides a concrete measurable goal — how accurately does the model hit the specified level?

• Compared models: Qwen3-Next-80B-A3B-Instruct (general-purpose) and Qwen3-Coder-Next (code-specialized)
• Sample size: 2520 tasks from three programming benchmarks
• Two scenarios: general complexity control ("make simpler/more complex") and precise control by Bloom's levels
• Methods: semantic clustering of task changes and layer-wise probing (Fisher's Discriminant Ratio)

Why do models find it easier to complicate than simplify?

Both models demonstrate a persistent directional asymmetry. When asked to increase complexity — add requirements, introduce abstractions, expand the conditions — both handle it reliably. When cognitive load needs to be reduced, results worsen: models either transform the task beyond recognition and change its educational meaning, or reproduce a version almost identical to the original that hasn't become simpler.

Layer-wise probing revealed differences in the internal structure of the two models. The general-purpose Qwen3-Next-80B showed clear feature separation in middle layers for both types of intervention. The code-specialized Qwen3-Coder-Next demonstrates weaker separation during general complexity control and a deeper activation peak when managing by Bloom's levels — a sign that code specialization structures internal representations of cognitive complexity differently, rather than simply redistributing base capabilities.

What this means

The ability to solve tasks does not automatically transfer to the ability to pedagogically adapt them. The authors directly state: high performance on execution benchmarks does not guarantee educational control — these are two different competencies, the gap between which is real and measurable.

For developers of educational AI systems, the conclusion is practical: embedding a language model as a generator of adaptive learning content should come with caveats. "Complication mode" works reliably, "simplification mode for a student's level" does not. The authors view the proposed framework as the foundation for standardized evaluation of a model's educational fitness and targeted fine-tuning for pedagogical tasks.

Frequently Asked Questions

How many tasks were used in the test and from which sources?

2520 programming tasks from three benchmarks. Each went through two types of intervention: general complexity control (make simpler or more complex) and precise control by Bloom's taxonomy levels (shift to a specific level up or down).

What is the main difference between

Qwen3-Next-80B and Qwen3-Coder-Next according to probing results?

The general-purpose Qwen3-Next-80B showed clear feature separation in middle layers for both types of intervention. Qwen3-Coder-Next performed worse at distinguishing tasks during general complexity control but demonstrated a deeper activation peak when managing by Bloom's levels — the authors interpret this as evidence of a different internal organization of complexity representations in specialized code models.