ARF-RLHF: Adaptive Reward-Following for RLHF through Emotion-Driven Self-Supervision and Trace-Biased Dynamic Optimization

Abstract: With the rapid advancement of Reinforcement Learning from Human Feedback (RLHF) and autoregressive transformers, state-of-the-art models such as GPT-4.0, DeepSeek R1, and Llama 3.3 increasingly emphasize answer depth and personalization. However, most existing RLHF approaches (e.g., PPO, DPO) still rely on a binary-preference (BT) paradigm, which, while reducing annotation costs, still requires substantial human effort and captures only group-level tendencies rather than individual preferences. To overcome these limitations, we propose Adaptive Reward-Following (ARF), a self-assessment framework that leverages a high-precision emotion analyzer achieving over 70% accuracy on GoEmotions, Sentiment140, and DailyDialog to convert free-form user feedback into continuous preference scores. We further enrich and debias these signals through lightweight data augmentations, including synonym replacement, random trace truncation, and score bias annotation algorithm. A Dynamic Adapter Preference Tracker continuously models evolving user tastes in real time, enabling our novel Trace Bias (TB) fine-tuning algorithm to optimize directly on these tracked rewards instead of coarse binary labels. Experiments on Qwen-2/2.5, Gemma-2, and Llama-3.2 across four preference domains demonstrate that ARF achieves an improvement of 3.3% over PPO and 7.6% over DPO. Moreover, TB preserves theoretical alignment with PPO and DPO objectives. Overall, ARF presents a scalable, personalized, and cost-effective approach to RLHF LLMs through autonomous reward modeling.