What it is
Uses RL with metacognitive feedback to train LLMs to report uncertainty more faithfully, then rewrites outputs into natural hedged language. Instead of rewarding confidence alone, it upweights strong completions whose self-assessed performance matches actual performance. It pairs this with self-judgment-based data selection and a separate rewrite stage for natural linguistic hedging.
Key results
- Up to 63% stronger post-training results than standard RL
- Llama3.1-8B +RLMF: 0.84 avg cMFG vs 0.77 with standard RL
- Qwen3-8B +RLMF: 0.83 avg cMFG vs 0.51 with standard RL
Self-evaluation quality is a useful RL signal for making LLM uncertainty expression more faithful without sacrificing task accuracy.