New FCPA method aligns LLM generators and validators, boosting consistency by 27 points
LLMs often reject their own answers—FCPA training fixes that gap with a frequency correction.
Large language models have a known inconsistency: they can generate a response that, when asked to validate the same answer independently, they deem invalid. This generator-validator (G-V) gap is problematic for reliability, especially in tasks requiring factual consistency. The authors show that part of the gap stems from frequency bias—generators assign low likelihood to valid but rare utterances, making naive consistency measures unworkable. They formalize a correction: under a rational agent model, a validator should agree with a frequency-normalized generator probability.
Their new method, FCPA (Frequency-Corrected Validator-to-Generator Alignment), implements this correction as a training objective. Experiments show substantial gains: up to +27 percentage points in Pearson correlation on instruction-following (IFEval) and code generation (HumanEval) benchmarks, without degrading the validator's own performance. The approach is model-agnostic and can be applied during fine-tuning, offering a practical path to more consistent LLMs that don't contradict their own reasoning.
- FCPA improves Pearson correlation by up to +27pp on IFEval and HumanEval benchmarks
- Addresses the generator-validator gap by correcting for natural frequency biases in rare but valid responses
- Preserves validator quality across all evaluated tasks, ensuring no trade-off in accuracy
Why It Matters
Makes LLMs more reliable by aligning generation and validation, reducing contradictory outputs in production.