A contrastive decoding method addresses limitations of early-exit strategies by constructing an 'amateur' model via dynamic layer pruning rather than simple truncation. Removing specific intermediate reasoning layers produces a better-calibrated contrastive prior with more informative logits, steering generation away from factually incorrect but high-probability tokens while maintaining fluency. This approach achieves consistent factuality improvements with minimal inference overhead, evaluated on TruthfulQA, FACTOR (News, Wiki), and StrategyQA datasets using TruthfulQA (MC1, MC2, %Truth, %Info), FACTOR, and StrategyQA Accuracy metrics.

Authors

• Person: Not available Organization: GitHub
  EdinburghNLP/awesome-hallucination-detection - GitHub

Sources

• EdinburghNLP/awesome-hallucination-detection - GitHub github.com GitHub via serper

Referenced by nodes (1)

• TruthfulQA concept