Model Evaluation

Testing frontier models to uncover new risks and highlight security issues.

Evaluations (or model tests) serve as a mechanism for identifying risks and assessing whether a system can safely operate in real-world scenarios. Leveraging our research experience, FAR.AI focuses on testing frontier models to uncover new risks and highlight security issues, enabling developers to put in place appropriate mitigations for currently deployed systems. We explore trends in frontier models to identify which problems will grow increasingly severe over time and require urgent attention. We also work on developing metrics and benchmarks measuring reliability and security to provide clear targets for researchers, improving the transparency of future testing.

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

Model Evaluation

Model Evaluation Research:

STARC: A General Framework For Quantifying Differences Between Reward Functions

Uncovering Latent Human Wellbeing in Language Model Embeddings

Uncovering Latent Human Wellbeing in LLM Embeddings

Exploiting Novel GPT-4 APIs

We Found Exploits in GPT-4’s Fine-tuning & Assistants APIs

Evaluating the Moral Beliefs Encoded in LLMs

Evaluating LLM Responses to Moral Scenarios

Inverse Scaling: When Bigger Isn't Better