A language model is a probabilistic text generator that predicts likely next tokens from patterns learned in data. Its outputs are free-form: one prompt can produce many valid answers, while statistically plausible text can still be factually wrong. section 1.0 section 5.4.5
Because language models learn patterns rather than following only explicit programs, they can develop capabilities and behaviors that were not directly specified. Evaluation measures what the model does; alignment steers behavior toward human goals, policies, principles, or a consistent persona. section 1.0 section 1.2
Analyze language-model behavior through complementary lenses:
- Evaluate accuracy, reliability, helpfulness, coherence, tone, and edge-case behavior rather than relying only on unit tests or loss curves. section 1.0
- Align responses with explicit policies, higher-level principles, human goals, or a chosen persona. section 1.2
- Monitor hallucinations because mitigation can reduce their rate but cannot currently eliminate them. section 5.4.5
Preference learning for responses
In RLHF, a reward model compares complete responses rather than short action segments. Human preference labels train the reward model, which then scores language-model outputs; the longer response introduces a difficult credit-assignment problem. section 7.4.3
























































































