The Limitation of Static Prompting

Traditional prompt engineering often relies on manual iteration or static templates, which fail to capture the nuanced, high-dimensional space of effective instructions. As LLMs become more complex, the relationship between specific prompt tokens and model performance becomes increasingly non-linear and difficult to optimize through trial and error. The authors argue that treating prompts as fixed strings ignores the underlying semantic structure that models actually respond to.

Embedding by Elicitation: A Dynamic Approach

'Embedding by Elicitation' shifts the optimization process from the discrete token space to a continuous latent space. Instead of searching for the perfect sequence of words, the system learns a dynamic representation of the prompt. By utilizing Bayesian Optimization (BO), the framework iteratively probes the model's performance on specific tasks, using the feedback to refine the latent embedding. This approach allows the system to 'elicit' the most effective prompt structure by navigating the model's internal representation space rather than relying on human intuition alone.

Improving Bayesian Optimization for Prompts

Bayesian Optimization is typically computationally expensive and struggles with high-dimensional inputs. The authors propose that by mapping prompts to a learned latent space, they can significantly reduce the search space complexity. This allows for faster convergence on high-performing system prompts compared to standard gradient-based or brute-force search methods. The method effectively treats the LLM as a black-box function, optimizing the system prompt to maximize a specific objective function (e.g., accuracy, latency, or adherence to constraints) without requiring access to the model's internal weights.