Data and project site: "Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values"

This repository holds the model rollouts and judge outputs behind Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values, and the source of valueleakage.net.

The code repository mounts it as its data/ submodule. With the data in place, all analysis and plotting scripts run from these caches and need no API keys; only re-sampling the models does.

Layout

Path	Contents	Size
final_data/	Rollout and judge caches: Donation Bet, AI Bubble & AGI Tweet, Job Offer, Agentic Grading	6.4 GB
choosing_activities/results/	Choosing Activities pipeline results	1.3 GB
site/	Landing page of valueleakage.net	1 MB
static/	Rollout browser and the JSON it reads	0.9 GB

The browser JSONs under static/data/ are a fixed random sample of the caches: up to 90 rollouts per model and condition, 270 for Choosing Activities and Agentic Grading. They are generated by export_browser_data.py in the code repository.

Website

GitHub Pages serves the site directly from this repository. The workflow in .github/workflows/pages.yml redeploys it whenever a push to main touches site/ or static/:

• valueleakage.net — landing page (site/)
• valueleakage.net/browser — rollout browser (static/)

To preview changes locally:

(cd static && python3 -m http.server 8000)   # browser at http://localhost:8000
(cd site && python3 -m http.server 8001)     # landing page; its /browser link only works once deployed

Citation

@misc{betley2026valueleakagellmsanswers,
      title={Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values},
      author={Jan Betley and Johannes Treutlein and Jan Dubiński and Harry Mayne and Karol Gałązka and Niels Warncke and Anna Sztyber-Betley and Owain Evans},
      year={2026},
      eprint={2607.14345},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2607.14345},
}