How to Control LLM Risks?

Risk control techniques in binary classification can be used to enhance LLM trustworthiness by implementing response guardrails, such as censoring undesired content.

---

LLM as Binary Classifier

Conformal prediction methods can be applied to LLM-based classifiers. We propose a method presented in Benchmarking LLMs via Uncertainty Quantification, which:

1. Reduces a commonsense reasoning task (CosmosQA dataset) to a classification problem
2. Extracts only the logits corresponding to the possible answers
3. Applies a softmax so the LLM can be used as a simple classifier
4. Enables the use of conformal predictions

---

Resources

Educational Repository

The following repository (not maintained by the MAPIE team) implements part of this paper for educational purposes in the MAPIE_for_cosmosqa notebook.

Blog Article

Read our blog article on Medium where we dive deeper into the topic.