Code and data for the paper "Attention-Head Binding as a Mechanistic Marker of Accessibility Concept Emergence in Language Models" by Khanh-Dung Tran (2026).

Associated paper: https://doi.org/10.22541/au.177067444.42828315/v1 (Preprint hosted on Authorea / TechRxiv)

Feedback and issues are welcome.

This study extends prior work on accessibility knowledge across Pythia model sizes (Salas, 2026) by introducing a mechanistic attention-based metric for concept emergence.

We introduce EB* (effective binding), a mechanistic interpretability metric that tracks how attention heads bind multi-token accessibility terms (e.g., "screen reader," "alt text") during training. Using the Pythia model suite (160M, 1B, 2.8B parameters) across eight training checkpoints, we demonstrate:

• C1 (Lead-lag emergence): Binding precedes behavioral competence (Spearman r = 0.33–0.34, p < 0.001)
• C3 (Unlockability): Few-shot prompting yields +61 percentage points improvement when EB* > 0.6
• C4 (Decoupling): At 1B scale, binding saturates early while behavior continues improving
• C5 (Causal regimes): Cross-scale reversal — binding heads are necessary at 160M but interfering at 2.8B

attention-binding-a11y/ ├── src/ # Source code │ ├── utils_model.py # Model loading with checkpoint support │ ├── scoring.py # Recognition and generation scoring │ ├── eval_behavior.py # Behavioral probe evaluation │ ├── extract_attention.py # Attention extraction, BSI/EB/EB* metrics │ ├── tokenization_audit.py # Tokenization span verification │ ├── analysis_pilot.py # Correlation and Go/No-Go analysis │ ├── minimal_causal.py # C5: 160M head ablation │ ├── minimal_causal_28b.py # C5: 2.8B head ablation │ └── eval_few_shot.py # C3: Few-shot unlockability testing ├── data/ │ ├── prompts/ │ │ └── pilot_terms.jsonl # 12 prompts (3 terms × 2 tasks × 2 variants) │ ├── results/ │ │ ├── behavioral/ # Behavioral probe scores │ │ ├── binding/ # EB* binding metrics │ │ ├── causal/ # C5 ablation results │ │ └── few_shot/ # C3 unlockability results │ └── tokenization/ # Tokenization tables ├── config/ │ └── pilot.yaml # Experiment configuration ├── notebooks/ │ ├── figure1_emergence_curves.ipynb # Figures 1 & 4 │ ├── verify_checkpoints_v2.ipynb # Checkpoint verification │ └── verify_setup.ipynb # Environment check ├── figures/ # Generated figures ├── paper/ # Paper source (Markdown) │ ├── main.md │ ├── sections/ │ └── appendix/ ├── tests/ │ └── test_behavioral.py # Unit tests ├── requirements.txt ├── setup_data.py # Environment setup script ├── REPRODUCTION_CHECKLIST.md ├── LICENSE └── README.md 

• Python 3.9+
• CUDA-capable GPU (recommended: 16GB+ VRAM for 2.8B model)

git clone https://github.com/dunktra/attention-binding-a11y.git cd attention-binding-a11y python -m venv venv source venv/bin/activate pip install -r requirements.txt # Verify environment python setup_data.py

Pythia model checkpoints are downloaded automatically from HuggingFace when running experiments.

# 1. Verify tokenization spans python src/tokenization_audit.py # 2. Extract binding metrics (repeat for each model/checkpoint) python src/extract_attention.py 160m step120000 # 3. Run behavioral evaluation python src/eval_behavior.py 160m step120000 # 4. C3: Few-shot unlockability python src/eval_few_shot.py # 5. C5: Causal ablation python src/minimal_causal.py # 160M python src/minimal_causal_28b.py # 2.8B # 6. Summary statistics and correlations python src/analysis_pilot.py 160m

@online{tran2026attentionbinding, title={Attention-Head Binding as a Mechanistic Marker of Accessibility Concept Emergence in Language Models}, author={Tran, Khanh-Dung}, year={2026}, doi={10.22541/au.177067444.42828315/v1}, url={https://doi.org/10.22541/au.177067444.42828315/v1} }

The paper source is in paper/ as Markdown. To compile to PDF:

# Install pandoc sudo apt-get install pandoc texlive-latex-base texlive-latex-extra # Compile all sections into a single PDF cd paper pandoc main.md sections/introduction.md sections/related_work.md \ sections/methods.md sections/results.md sections/discussion.md \ sections/conclusion.md appendix/raw_data.md \ -o attention_binding_a11y.pdf \ --pdf-engine=pdflatex \ -V geometry:margin=1in

MIT License — see LICENSE.

This work builds directly on and extends prior analysis by Trisha Salas on testing accessibility-related knowledge across Pythia model sizes (Salas, 2026).

Salas’ work established that accessibility concepts such as “screen reader” and “alt text” emerge behaviorally at different rates across model scales. The present study extends this line of inquiry by shifting from behavioral evaluation to mechanistic analysis, introducing EB* as an attention-based binding metric to probe how and when these concepts emerge internally during training, and how their causal role changes with scale. Visit @trishasalas' Github repo for more details.