InvThink: Towards AI Safety via Inverse Reasoning

Example code for the paper:
"InvThink: Towards AI Safety via Inverse Reasoning" (XXX xxx)

Paper (TBD) | Project Page

To train LLMs with the InvThink framework, you should prepare a dataset that consists of:

1. query,
2. inverse reasoning (including harm enumeration, consequence analysis, mitigation strategy, and safe forward reasoning), and
3. safe response.

You can train your LLMs using the provided scripts.
Please update configuration parameters (e.g., model_name) according to your setup.

@inproceedings{invthink2025,
  title     = {InvThink: Towards AI Safety via Inverse Reasoning},
  author    = {Anonymous},
  booktitle = {XXX},
  year      = {2025}
}