Think Through Uncertainty: Improving Long-Form Generation Factuality via Reasoning Calibration

Build an inference wrapper that implements 'Reasoning Calibration' for LLMs to estimate confidence per-token. This stops models from confidently lying in long-form tasks.

Suggested repo: calibra-gen

"Force your LLM to admit when it's just guessing."

Estimated effort: 50h