Value alignment for AI is not “one-size-fits-all”: even polite and friendly models can still fail to represent individual user contexts and preferences, and local cultural norms. This paper presents a modular workflow for personal fine-tuning, synthesizing four core components from our previous research: (1) robust vectorization of user values and preferences, (2) a binary choice user interface (UI) approach to capturing those preferences with minimal cognitive load, (3) contrastive activation methods for steering large language models (LLMs) via difference vectors, and (4) knowledge graph integration for more auditable and structured alignment. Our approach—descended from past research on “Towards an End-to-End Personal Fine-Tuning Framework”—demonstrates how these elements can be combined to create personalized, context-rich alignment solutions. We report on user studies for the forced-choice UI, describe an experimental pipeline for deriving “control vectors”, and propose a “moral graph” method for bridging symbolic and vector-based alignment. Our findings suggest that multi-pronged personalization can significantly reduce user annotation fatigue, improve alignment fidelity, and allow for more flexible, interpretable AI behaviors.

Large decoder-based language models have become the dominant architecture for reward modeling in reinforcement learning from human feedback (RLHF). However, as reward models are increasingly deployed in test-time strategies, their inference costs become a growing concern. We present TinyRM, a family of small, bidirectional masked language models (MLMs) with as few as 400 million parameters, that rival the capabilities of models over 175 times larger on reasoning and safety preference modeling tasks. TinyRM combines FLAN-style prompting, Directional Low-Rank Adaptation (DoRA), and layer freezing to achieve strong performance on RewardBench, despite using significantly fewer resources. Our experiments suggest that small models benefit from domain-specific tuning strategies, particularly in reasoning, where lightweight finetuning methods are especially effective. While challenges remain in building generalist models and conversational preference modeling, our preliminary results highlight the promise of lightweight bidirectional architectures as efficient, scalable alternatives for preference modeling.