Contrastive Pretraining Teaches Format Generation, Not Behavioral Knowledge

Code and data for the paper "Contrastive Pretraining Teaches Format Generation, Not Behavioral Knowledge" (Sanchez, 2026).

Paper: Accepted at the 1st Conference For AI Scientists (CAISc 2026), Open-Ended Problems track.

Key Finding

Injecting contrastive behavioral pairs into 5% of training blocks (0.11% of tokens) breaks the behavioral emergence wall: a 7M model exceeds vanilla 34M on bias and sycophancy at 5x fewer parameters. But logit-level analysis reveals the mechanism is format generation enablement, not knowledge creation. Every vanilla model from 3M to 64M already achieves 41.0% accuracy when constrained to answer tokens.

Results Summary

Finding	Result
7M contrastive vs 34M vanilla sycophancy	0.513 vs 0.300
Optimal injection rate	5% (non-monotonic)
Logit-level accuracy (all vanilla models)	41.0% (123/300)
Deconcentration score separation	Direct > 1.0, Null < 0.16
Cross-transfer onset	d = 96 (readout threshold)
Perplexity cost at 7M (5%)	-1.7% (improves LM quality)
Perplexity cost at 64M (5%)	+1.0% (negligible)

Repository Structure

paper/          LaTeX source and bibliography
figures/        Paper figures (PDF)
src/            Experiment code
  configs.py              Model architecture configs (3M--64M)
  train_model.py          GPT-2 pretraining on OpenWebText
  train_contrastive.py    Contrastive variant training (3%, 5%, 10%)
  train_geometric.py      Geometric regularizer (negative control)
  scale_audit.py          Behavioral evaluation + subspace geometry

Reproducing

1. Train vanilla and contrastive models

python src/train_model.py --config 7m
python src/train_contrastive.py --config 7m --rate 0.05 --behaviors bias,sycophancy

2. Run behavioral audit

python src/scale_audit.py --model checkpoints/7m_contr_05 --probes data/bbq_300.json

Citation

@inproceedings{sanchez2026contrastive,
  title={Contrastive Pretraining Teaches Format Generation, Not Behavioral Knowledge},
  author={Sanchez, Bryan},
  booktitle={Proceedings of the 1st Conference For AI Scientists (CAISc 2026)},
  year={2026}
}

License

MIT