Qwen3 language models raise task difficulty easily but struggle to simplify it

In July 2026, researchers published a framework for measuring the "educational control" of language models: the ability to preserve a task's teaching purpose while shifting its cognitive load to a specified level of Bloom's taxonomy. Using 2520 programming tasks and the Qwen3-Next pair of models, the authors found a consistent asymmetry: the models make tasks harder with confidence, but cannot reliably make them easier.

What "educational control" of a model means

Most current benchmarks score language models on a single measure: can they solve a task correctly? In education, a different skill matters: adapting a ready-made task for a specific learner, making it easier for a beginner or harder for an advanced student, while keeping its teaching purpose and topic.

To measure this, the authors used the revised Bloom's taxonomy, a scale of six sequential levels of cognitive complexity: Remember, Understand, Apply, Analyze, Evaluate and Create. Each level describes a qualitatively different way of working with information, from recalling from memory to independent synthesis. The scale gives a concrete, measurable target: how precisely does the model hit the requested level?

• Models compared: Qwen3-Next-80B-A3B-Instruct (general-purpose) and Qwen3-Coder-Next (specialized for code)
• Sample size: 2520 tasks from three programming benchmarks
• Two scenarios: general difficulty control ("make easier/harder") and precise control by Bloom level
• Methods: semantic clustering of the changes made to tasks and layer-wise probing (Fisher's Discriminant Ratio)

Why do models make tasks harder more easily than they make them easier?

Both models show a consistent directional asymmetry. When asked to raise difficulty by adding requirements, introducing abstractions or widening the problem statement, both succeed reliably. When asked to lower cognitive load, results get worse: the models either transform the task beyond recognition and change its teaching purpose, or return a version nearly identical to the original that is no easier.

Layer-wise probing revealed differences in the internal workings of the two models. The general-purpose Qwen3-Next-80B showed clear feature separation in the middle layers for both types of intervention. The specialized Qwen3-Coder-Next shows weaker separation under general difficulty control and a deeper activation peak under Bloom-based steering. This suggests that specializing in code structures internal representations of cognitive complexity differently, rather than simply redistributing base capabilities.

What this means

Solving tasks does not automatically translate into adapting them pedagogically. The authors state plainly that high performance on execution benchmarks does not guarantee educational control: these are two different competencies, and the gap between them is real and measurable.

For developers of educational AI systems, the takeaway is practical: a language model should be built in as an adaptive content generator with caveats. The "make harder" mode works reliably; the "simplify for the student's level" mode does not. The authors see the framework as a basis for standardized assessment of a model's educational suitability and for targeted fine-tuning on pedagogical tasks.

Frequently asked questions

How many tasks were used in the test, and where did they come from?

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

What is the main difference between

Qwen3-Next-80B and Qwen3-Coder-Next in the probing results?

The general-purpose Qwen3-Next-80B showed clear feature separation in the middle layers for both types of intervention. Qwen3-Coder-Next separated tasks less well under general difficulty control but showed a deeper activation peak under Bloom-based steering. The authors read this as a sign that specialized code models organize their internal representations of complexity differently.