Truth or Deceit? A Bayesian Decoding Game Enhances Consistency and Reliability

Large Language Models (LLMs) often produce outputs that - though plausible - can lack consistency and reliability, particularly in ambiguous or complex scenarios. Challenges arise from ensuring that outputs align with both factual correctness and human intent. This is problematic in existing approaches that trade improved consistency for lower accuracy. To mitigate these challenges, we propose a novel game-theoretic approach to enhance consistency and reliability during the decoding stage of LLM output generation. This ensures consistency through Correctness Alignment and enhances reliability via Ambiguity Calibration. Remarkably, our game design allows smaller models to outperform much larger models through game mechanisms (e.g. Large Language Models (LLMs) have demonstrated extraordinary capabilities in tasks such as factual question answering, fact-checking, and open-ended text generation (Brown et al., 2020; Radford et al., 2021). However, as these generative models increase in ...