💬 NLP

Noise Steering for Controlled Text Generation: Improving Diversity and Reading-Level Fidelity in Arabic Educational Story Generation

This paper demonstrates that injecting calibrated Gaussian noise into the internal representations of small Arabic language models during inference is a superior, training-free strategy for generating diverse, high-quality, and reading-level-faithful educational stories compared to traditional high-temperature sampling.

Original authors: Haziq Mohammad Khalid, Salsabeel Shapsough, Imran Zualkernan

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haziq Mohammad Khalid, Salsabeel Shapsough, Imran Zualkernan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a teacher trying to create a storybook for 6-year-old children learning to read in Arabic. You need the story to be simple (using only words a 6-year-old knows), culturally neutral, and follow a clear structure (beginning, middle, end). But here's the catch: you need to write 50 different stories that all meet these strict rules, and none of them can be boring repeats of the same plot.

Writing these manually is exhausting. So, you ask a computer (an AI) to do it. But AI has a habit of two things:

Being too boring: It keeps telling the same story about a boy named Ahmed finding a lost ball, over and over again.
Going off the rails: If you tell it to "be more creative," it suddenly starts writing stories about space aliens and complex philosophy, which is useless for a 6-year-old.

This paper is about a new trick called "Noise Steering" that helps the AI be creative without losing its mind or forgetting the rules.

The Problem: The "Bored Robot" vs. The "Crazy Robot"

Think of the AI as a robot chef.

The Bored Robot: If you ask it to make 50 sandwiches, it makes 50 identical ham sandwiches. It's safe, but boring.
The Crazy Robot: If you tell the robot to "add more spice!" (this is what scientists call "High Temperature"), it might throw in hot sauce, chocolate, and pickles. Now the sandwiches are unique, but they are inedible and break the rules of what a sandwich is supposed to be.

The researchers wanted a way to make the robot chef invent new sandwich recipes (diversity) without adding hot sauce that ruins the meal (breaking the rules).

The Solution: "Noise Steering" (The Gentle Nudge)

Instead of changing the robot's recipe book (which would take forever to rewrite), the researchers decided to nudge the robot's brain while it was cooking.

They realized that the robot thinks in layers, like an onion.

The Outer Layer (The Output): This is the final word the robot picks.
The Inner Layers (The Brain): This is where the robot processes ideas, connects thoughts, and decides what to say next.

The researchers tried injecting tiny amounts of "static" or "noise" (like a gentle electrical buzz) into different parts of the robot's brain while it was thinking. They tested four different ways to do this:

1. The "Embedding" Nudge (The Bad Idea)

They tried buzzing the very first spark of an idea.

Analogy: Imagine trying to fix a car by shaking the steering wheel before you even turn the key.
Result: The car (the AI) crashed immediately. For some models, this made the robot stop talking entirely or speak gibberish.

2. The "Attention" Nudge (The Risky Idea)

They tried buzzing the part of the brain that decides "what to focus on."

Analogy: Imagine trying to make a chef more creative by shaking their hand while they are chopping vegetables.
Result: Sometimes it worked, but often the chef dropped the knife. The stories became chaotic and broke the rules.

3. The "Residual Stream" Nudge (The Goldilocks Solution)

They tried buzzing the "main highway" where all the robot's thoughts travel after a full thought is formed.

Analogy: Imagine the robot is walking down a path. Instead of shaking its hand, you gently nudge its shoulder after it has decided which way to turn. This changes its stride slightly, making it take a slightly different path, but it doesn't make it fall over.
Result: This worked perfectly. The robot started telling different stories (maybe a girl finding a kite instead of a boy finding a ball), but it still used simple words and followed the rules. It was creative but safe.

4. The "Smart" Nudge (AENI)

This was a special version of the "Attention" nudge. Instead of buzzing the robot all the time, the researchers made the noise smart.

Analogy: Imagine a dance partner. If the robot is dancing in a circle (repetitive), the partner gives it a gentle push to try a new move. But if the robot is already dancing wildly and creatively, the partner stops pushing so it doesn't get dizzy.
Result: This was the most stable method. It only added noise when the robot was getting bored and repetitive, and stopped when the robot was already being creative.

The Big Discovery

The researchers compared their "Noise Steering" tricks against the old method of just turning up the "creativity dial" (High Temperature).

The Old Method (High Temperature): It made the stories more diverse, but the vocabulary got too hard for 6-year-olds, and the stories often fell apart. It was like giving a toddler a thesaurus instead of a picture book.
The New Method (Noise Steering): It made the stories diverse without making them harder to read. The robot stayed in the "early grade" reading level, which is exactly what teachers need.

Why This Matters

This paper proves that you don't need to rebuild the AI or teach it new things to make it better. You just need to know where to gently nudge it.

For Teachers: You can now generate hundreds of unique, simple stories for reading tests without them all sounding the same.
For AI: It shows that the best way to get creativity out of a machine isn't to make it "wild," but to give it a structured, gentle push at the right moment.

In short: The researchers found a way to make the AI "think outside the box" without breaking the box or forgetting that it's supposed to be talking to a 6-year-old. They did this by gently shaking the robot's brain in the right spot, rather than screaming at it to be creative.

1. Problem Statement

The paper addresses the challenge of generating diverse, pedagogically valid short stories for Early Grade Reading Assessments (EGRA) in Arabic.

The Conflict: There is a fundamental tension between diversity (avoiding repetitive plots to ensure assessment validity) and constraint adherence (strict limits on vocabulary, reading level, narrative structure, and cultural neutrality).
The Context:
- Language Complexity: Arabic presents unique challenges due to rich morphology, orthographic variation, and diglossia (Modern Standard Arabic vs. dialects).
- Model Constraints: Practical deployment in under-resourced educational settings requires Small Language Models (SLMs) (7–9B parameters) rather than large frontier models. However, these smaller models are prone to generation instability when diversity is forced.
- Failure of Current Methods: Standard output-level stochasticity (e.g., high-temperature sampling) often leads to "catastrophic collapse" (incoherent text) or violates reading-level constraints by introducing overly complex vocabulary.

2. Methodology: Noise Steering

The authors propose Noise Steering, a training-free method that injects calibrated Gaussian perturbations into the internal representations of transformer models during inference. This approach avoids the need for fine-tuning.

2.1 Injection Strategies

The paper evaluates four distinct injection strategies across five Arabic-centric SLMs:

Embedding Noise: Injects noise directly into token embeddings at the very start of the residual stream.
- Effect: High risk of total collapse, especially in models with sparse Arabic embedding spaces.
Attention-Logit Noise: Injects noise into the output of the self-attention sub-layer (after projection $W_O$ $W_{O}$ ) but before the MLP and residual addition.
- Effect: Unstable; causes high collapse rates because it disrupts information routing before feed-forward processing.
Residual Stream Noise (L-RES): Injects noise into the residual stream after both the self-attention and MLP sub-layers have completed their computation for a block.
- Effect: The perturbation enters the stream when the block's contribution is fully formed, allowing for controlled propagation.
Attention Entropy Noise Injection (AENI): An adaptive strategy where noise magnitude is conditioned on the peakedness of the model's attention distribution.
- Mechanism: If attention is highly concentrated (peaked), indicating a risk of repetition, high noise is injected. If attention is diffuse, noise is reduced.
- Goal: Stabilizes the otherwise unreliable raw attention noise by intervening only when the model is over-confident.

2.2 Noise Decay

To prevent the degradation of constraint adherence over long generations, a cosine decay schedule is applied. Noise magnitude is high at the beginning of the story (encouraging creative exploration) and decays to zero toward the conclusion (ensuring the story ends with strict constraint compliance).

2.3 Calibration

Since internal representation scales vary across architectures, noise magnitude is calibrated using the Root Mean Square (RMS) of the model's residual stream activations. A fixed scaling coefficient ( $\alpha = 0.175$ ) is applied relative to the median RMS of each model to ensure consistent relative perturbation strength.

3. Experimental Setup

Models: Five small (7–9B parameter) Arabic-centric instruct models: Jais 8B, ALLaM 7B, Fanar 9B, AceGPT 8B, and Phi-4-mini.
Baselines:
- Standard decoding ( $T=1.0$ ).
- High-temperature sampling ( $T=1.8$ ) with Top-k and Top-p.
Evaluation Metrics:
- Diversity: Vendi Score (semantic diversity) and Self-BLEU (lexical diversity).
- Quality: LLM-judged scores for readability, logic, grammar, and linguistic quality.
- Constraints: Programmatic checks (word count, tense, proper nouns) and LLM-judged checks (narrative structure, gender balance, cultural neutrality).
- Reading Level: Estimated grade level to ensure suitability for early-grade learners.
- Stability: Modal Collapse Rate (percentage of generated stories that are incoherent or repetitive).

4. Key Results

4.1 Stability and Modal Collapse

High-Temperature Sampling: Failed catastrophically on several models (e.g., AceGPT and Phi-4-mini had collapse rates of 84% and 66%, respectively). It inflated reading levels (e.g., Grade 3 $\to$ Grade 6), violating EGRA requirements.
Embedding & Raw Attention Noise: Caused significant instability (up to 100% collapse on Phi-4-mini for embedding noise).
Residual Stream Noise (L-RES): Achieved a 0% collapse rate across all five models.
AENI: Eliminated collapse entirely across all models, stabilizing the raw attention noise method.

4.2 Diversity vs. Quality Trade-off

L-RES: Consistently improved diversity (Vendi Score) across all models with minimal impact on quality or constraint adherence.
- Example: Jais 8B saw a Vendi Score increase from 4.71 to 7.09 while maintaining Grade 2 reading level.
- Trade-off: For weaker models (e.g., Phi-4), diversity gains came with a slight quality drop, but the output remained usable.
AENI: Provided a "sweet spot" of stability. While it yielded slightly lower diversity scores than raw attention noise, it recovered quality significantly (e.g., ALLaM quality recovered from 2.55 to 6.73) and maintained reading levels.
High-Temperature: Only effective for Jais 8B; for all other models, it degraded quality and violated reading-level constraints.

4.3 Reading Level Fidelity

A critical finding is that internal noise steering (L-RES and AENI) preserves the early-grade reading level, whereas high-temperature sampling forces models to use more sophisticated, adult-level vocabulary, rendering the stories unsuitable for the target assessment context.

5. Key Contributions

Internal vs. Output Stochasticity: Demonstrates that perturbing internal representations (specifically the residual stream) is superior to output-level stochasticity for constrained educational content generation.
AENI Mechanism: Introduces Attention Entropy Noise Injection, an adaptive method that stabilizes noise injection by responding to the model's own attention confidence, effectively preventing runaway instability.
Empirical Validation on Arabic SLMs: Provides a comprehensive evaluation across five diverse Arabic SLMs, showing that noise steering is a viable, training-free strategy for improving diversity without sacrificing pedagogical validity.
Calibration Protocol: Establishes a robust method for calibrating noise magnitude based on model-specific activation scales (RMS), ensuring transferability across different architectures.

6. Significance

This work offers a practical solution for the Arabic educational technology sector, where high-quality, diverse, and level-appropriate reading materials are scarce. By proving that noise steering can unlock latent diversity in small models without compromising safety or constraints, the paper enables the scalable generation of EGRA-aligned content. It highlights that for constrained generation tasks, controlling the internal state of the model is more effective than manipulating the final probability distribution, offering a new paradigm for reliable LLM deployment in low-resource, high-stakes educational settings.