LAISR Lab
Long-term AI Safety Research · Cornell University
We study the values language models come to hold, and build ways to measure and verify them, so that increasingly capable systems stay aligned with the people they serve.
Focus
- Measuring valuesComparative, model-judged evaluations of how well behaviour fits a written constitution.
- Constitution recoveryVerifying a model’s values by recovering the constitution it acts on from its behaviour, then checking it against the one it was meant to follow.
- Character and its side effectsWhat else changes when a model is trained toward one trait.
- Open toolsResults, transcripts, code and data published for anyone to inspect and reuse.
Research
- EigenBenchA Comparative Behavioral Measure of Value Alignment
- Side Effects of Character TrainingQuantifying Cross-Constitution Drift in LLMs
