TL;DR: We extracted a "complexity vector" from Wikipedia article pairs and achieved linear control over text reading level. Steering coefficient (α) from -4 to +4 shifts output from elementary school (grade ~6) to graduate level (grade ~17), with R² = 0.90.
The Persona Vectors paper showed you can steer AI behavior by manipulating activation space geometry—identifying directions that correspond to behavioral traits and nudging the model along those dimensions during inference. No fine-tuning, no retraining. Just vector addition in the right place.
We asked: does reading-level complexity have a similar steerable dimension? Can you extract a "complexity vector" that moves outputs along a simple-to-sophisticated axis?
Using contrastive activation analysis on matched Wikipedia article pairs (Simple English vs. standard English, same topics), we captured how models represent stylistic complexity internally. The method worked: applying the extracted vector with coefficient α produces linear control over reading level with R² = 0.90. The relationship is clean, predictable, and robust across the effective range (α = -4 to +4). Beyond that, models degenerate into repetition—documenting the boundary of steerable space.
This project demonstrates reading-level steering - controlling how simple or complex an LLM's output is by adding a learned vector to its activations during generation.
Figure 1: Linear steering control in Qwen 3 4B. Left: Flesch-Kincaid grade level increases 1.28 grades per unit α (R²=0.896). Right: Reading ease decreases 8.14 points per unit α (R²=0.895). Both relationships are highly significant (p<0.001) across the effective range α=-4 to +4.
| Metric | Effect | R² | p-value |
|---|---|---|---|
| Flesch-Kincaid Grade Level | +1.28 grades per unit α | 0.896 | < 0.001 |
| Flesch Reading Ease | -8.14 points per unit α | 0.895 | < 0.001 |
| Syllables per Word | +0.089 per unit α | 0.874 | < 0.001 |
Effective range: α = -4.0 to +4.0 Total span: ~11 grade levels (elementary → graduate school)
Beyond α ≈ ±4.5, the model degenerates into repetition loops. We document this as the boundary of effective steering.
Prompt: "Can you explain quantum mechanics please?"
α = -5.0 (Grade 5.8, elementary):
Absolutely! Let's break quantum mechanics down in a way that's easy to understand — without getting too scary or too fancy. 🚀 It's really amazing stuff!
α = 0.0 (Grade 11.9, high school):
Absolutely! Quantum mechanics is one of the most fascinating and fundamental theories in modern physics. It describes how the universe behaves at the smallest scales—like atoms, electrons, and photons.
α = +4.0 (Grade 16.4, graduate):
Of course! Quantum mechanics constitutes arguably the most profound theoretical framework governing subatomic phenomena, elucidating behaviors fundamentally divergent from classical physics...
TediumVectorTinker/
├── models/
│ └── qwen3-4b-instruct/ # Model-specific implementation
│ ├── extract_complexity_vectors.py
│ ├── quantify_steering.py
│ ├── interactively_prompt.py
│ ├── linear_regression_analysis.py
│ ├── validate_reproducibility.py
│ └── output/ # Generated results
│ ├── complexity_vectors.pt
│ ├── steering_quantitative_results.csv
│ ├── grade_level_vs_alpha.png
│ └── ...
├── data/
│ └── wikipedia_pairs/ # Shared training data
│ └── wikipedia_pairs.json
└── prepare_wikipedia_pairs.py # Data preparation (shared)
This structure supports multi-model comparison. Each model gets its own folder under models/ with identical scripts but model-specific configurations.
We use contrastive activation analysis (inspired by persona vectors) to extract a "complexity vector":
- Matched pairs: Simple Wikipedia + Regular Wikipedia articles on the same 20 topics
- Capture activations: Run both versions through Qwen3-4B-Instruct, save hidden states at all layers
- Compute difference:
complexity_vector[layer] = mean(regular) - mean(simple) - Identify peak: Layer 35 (final layer) shows strongest signal (magnitude ~90)
The vector captures how the model represents "simple writing" vs "complex writing" in activation space.
During inference, we modify activations at layer 35:
steered_activations = original_activations + (α * complexity_vector)- Negative α: Steer toward simple/accessible language
- Positive α: Steer toward complex/academic language
- α = 0: No steering (baseline)
This is pure vector addition in activation space - no retraining, no fine-tuning.
- Python 3.12+
- ~48GB RAM (for bf16 inference; 4-bit quantization possible but untested)
- Apple Silicon Mac (MPS), NVIDIA GPU (CUDA), or CPU
git clone https://github.com/Embedding-Space/TediumVectorTinker.git
cd TediumVectorTinker
# Install with uv (recommended)
uv sync
# OR with pip
pip install torch transformers accelerate datasets matplotlib pandas scipy textstat requests tqdmuv run prepare_wikipedia_pairs.pyFetches 20+ matched Simple/Regular Wikipedia article pairs and validates reading level differences.
Output: data/wikipedia_pairs/wikipedia_pairs.json
cd models/qwen3-4b-instruct
uv run extract_complexity_vectors.pyCaptures activations for all 36 layers and computes complexity vectors.
Runtime: ~5 minutes (Apple M-series) Output:
output/complexity_vectors.ptoutput/complexity_vector_magnitude.pngoutput/complexity_magnitudes.json
uv run quantify_steering.pyTests steering at 41 α values from -5.0 to +5.0, measuring reading level at each point.
Runtime: ~5 minutes Output:
output/steering_quantitative_results.csvoutput/steering_quantitative_analysis.png
uv run linear_regression_analysis.pyPerforms linear regression on the effective range (α = -4.0 to +4.0), generates plots with regression lines.
Output:
output/grade_level_vs_alpha.pngoutput/reading_ease_vs_alpha.pngoutput/linear_regression_combined.png- Console statistics (R², p-values, slopes)
Expected: R² ≈ 0.89-0.90 for grade level, p < 0.001
Test steering on arbitrary prompts:
# Simplify output
uv run interactively_prompt.py -a -3.0 "Explain black holes"
# Increase complexity
uv run interactively_prompt.py -a 2.5 "Explain photosynthesis"
# Baseline (no steering)
uv run interactively_prompt.py "What is machine learning?"Options:
-a,--alpha: Steering strength (default: 0.0)-l,--layer: Layer to steer (default: 35)-m,--max-tokens: Max generation length (default: 200)--vector: Path to vectors file
Matched Simple/Regular Wikipedia articles provide:
- Semantic consistency: Same facts, different presentation
- Controlled complexity: Simple Wikipedia targets grade 7-11, Regular targets 11-17
- Clean signal: Isolates style/syntax from content
This beats comparing random "boring" vs "interesting" text, which confounds multiple variables.
Vector magnitude grows exponentially across layers, peaking at layer 35 (the final layer before logits). This is where the model makes stylistic decisions about formality and sentence structure.
We use greedy decoding (do_sample=False) to eliminate sampling variance. Results are fully reproducible given identical vectors.
Qwen3-4B-Instruct expects ChatML format. All scripts use tokenizer.apply_chat_template() and parse responses correctly.
We validated end-to-end reproducibility:
- Vector stability: Cosine similarity ≥ 0.999 between independent extractions
- Magnitude consistency: L2 norms match to <0.01% relative error
- Steering reproducibility: R² values match within 0.005 across runs
The methodology is robust to floating-point noise and MPS nondeterminism.
- Extreme α breaks down: Beyond ±4.5, model produces repetition loops ("elucidation elucidation elucidation...")
- Model-specific: Vector extracted from Qwen3-4B-Instruct-2507; untested on other architectures
- Domain variation: Results shown are for general knowledge; steering may behave differently on code, math, or poetry
- Metric limitations: Flesch-Kincaid is a heuristic; "grade 16" ≠ guaranteed college-level quality
- Educational content: Adapt explanations to different grade levels automatically
- Accessibility: Simplify complex text for broader audiences
- Technical writing: Control formality (papers vs blog posts)
- API parameter: Expose α as
reading_levelin LLM services - Content transformation: Academic paper → popular science article
This methodology generalizes to other linguistic dimensions. Potential vectors:
- Formality: Casual ↔ professional
- Conciseness: Verbose ↔ terse
- Objectivity: Opinionated ↔ neutral
- Creativity: Formulaic ↔ imaginative
- Confidence: Hedging ↔ assertive
The key: find matched pairs differing primarily along one dimension.
We're replicating this methodology on:
- Llama 3.2 3B Instruct (next)
- Gemma 3 4B (if accessible)
- Potentially larger models (7B-13B) to test scale effects
Organizational structure supports easy comparison across models.
If you build on this work:
@software{tediumvectortinker2025,
author = {Harrell, Jeffery and Alpha},
title = {Reading-Level Steering for Large Language Models},
year = {2025},
url = {https://github.com/Embedding-Space/TediumVectorTinker}
}This project emerged from a Saturday afternoon tinker session exploring "tedium vectors" in activation space. What started as curiosity evolved into clean linear control with R² = 0.90.
Methodology inspired by Persona Vectors (Tigges et al., 2023).
MIT License - See LICENSE file for details.
Collaborators: Jeffery Harrell & Alpha October 18-19, 2025
"We haven't had a tinker like this in a while."