FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models

Build a framework that allows real-time, fine-grained steering of model outputs via activation shaping at the inference layer. This replaces static system prompts with dynamic control.

Suggested repo: steer-flow

"Fine-grained control of LLM behavior at inference time."

Estimated effort: 50h