Generative Computing Releases Common Abstraction Tool for LLM Steering

Generative Computing has released a tool that unifies various steering techniques for controlling the behavior of large language models (LLMs) into a single common abstraction structure, based on the observation that these techniques share similar patterns.

Key Points

The tool covers four key points that influence model behavior.

• Input: Prompt engineering, etc.
• State: Activations, Attentions, etc.
• Structure: Model Weights modification
• Output: Logits and decoding strategies

It also provides Probe construction and evaluation capabilities to compare and assess the performance of steering techniques in specific use cases. The evaluation framework utilizes Inspect. The related repository can be found at https://generative-computing.github.io/steerability/.