Noise Steering for Controlled Text Generation: Improving Diversity and Reading-Level Fidelity in Arabic Educational Story Generation
New method injects calibrated noise into model representations to create varied, grade-appropriate content.
A team of researchers has introduced 'Noise Steering,' a novel technique for improving the diversity of AI-generated educational content without sacrificing quality or violating strict pedagogical constraints. The method, developed by Haziq Mohammad Khalid, Salsabeel Shapsough, and Imran Zualkernan, works by injecting carefully calibrated Gaussian noise directly into the internal representations of transformer language models during inference. This training-free approach was specifically designed to address a critical challenge in generating Arabic early-grade reading assessments: creating stories that are both diverse in plot to avoid repetitive test items and strictly adhere to vocabulary lists, narrative structures, and target reading levels.
The researchers evaluated their technique across five different Arabic-centric language models ranging from 7 to 9 billion parameters. They tested four specific noise injection strategies, including Residual Stream Noise and Attention Entropy Noise Injection (AENI), against traditional high-temperature sampling baselines. The results were clear: internal representation-level perturbation consistently improved narrative diversity with minimal impact on story quality or constraint adherence. Crucially, it preserved the intended early-grade reading level, whereas high-temperature sampling was found to inflate grade levels significantly and even cause 'catastrophic collapse' in some models. The study concludes that perturbing a model's internal state is a more suitable strategy for constrained educational content generation than simply increasing output-level randomness.
This work represents a significant step forward for educational AI, particularly for languages like Arabic that are historically underserved by large language models. By providing a method to generate a wider variety of valid test items automatically, it can help reduce the time and cost of creating high-quality, culturally relevant reading assessments. The finding that attention-logit noise can be stabilized by AENI to recover output quality also offers new insights for the broader field of controlled text generation, suggesting pathways to make AI outputs more reliably useful for specific, high-stakes applications.
- Noise Steering injects calibrated Gaussian perturbations into transformer model representations during inference, a training-free method.
- Tested on five 7-9B parameter Arabic-centric models, it improved diversity while preserving reading level better than high-temperature sampling.
- The technique enables generation of more varied, pedagogically valid stories for Arabic early-grade reading assessments.
Why It Matters
Enables scalable creation of diverse, culturally relevant educational content in underserved languages, improving assessment quality and accessibility.