A reward-driven analysis of autoregressive natural language generation.

This thesis presents a reward-oriented approach to autoregressive natural language generation (NLG), introducing a novel decoding strategy that reframes beam search as a reward optimization process. In this framework termed Reward Search, each candidate sequence (beam) acts as an autonomous decision unit, selecting its next token by maximizing a composite payoff function. This function integrates the language model's fluency, measured via log-probability, with external control signals such as toxicity reduction or sentiment promotion, computed through a fine-tuned auxiliary reward model. By embedding these constraints directly into the decoding process, Reward Search enables fine-grained control over generation while preserving fluency and coherence. Empirical evaluations across detoxification and sentiment-controlled tasks demonstrate that the proposed method effectively reduces harmful or off-target outputs without sacrificing generation quality, offering a principled alternative to conventional decoding algorithms for safer and more controllable NLG systems.