As opposed to model fine-tuning, prompt engineering is a lightweight approach to fine-tuning by ensuring the prompt holds contextual information via an iterative process.

As I always say, the complexity of any implementation needs to be accommodated somewhere.

In the case of LLMs, the model can be fine-tuned, or advanced prompting can be used in an autonomous agent or via prompt-chaining.

In the last two examples mentioned, the complexity is absorbed in prompt engineering and not via LLM fine-tuning.

The objective of contextual iterative prompting is to absorb the complexity demanded by a specific implementation; and not offload any model fine-tuning to the LLM.

Contextual iterative prompting is not a new approach, what this paper considers is the creation and automation of an iterative Context-Aware Prompter.

At each step (each dialog turn) the Prompter learns to process the query and previously gathered evidence, and composes a prompt which steers the LLM to recall the next piece of knowledge.

This process reminds of soft prompts and prompt tuning.

The paper claims that their proposed Context-Aware Prompter Designoutperforms existing prompting methods by notable margins.

The approach shepherds the LLM to recall a series of stored knowledge (e.g., C1 and C2) that is required for the multi-step inference (e.g., answering Q), analogous to how humans develop a “chain of thought” for complex decision making.


The automated process establishes a contextual chain-of-thought and negates the generation of irrelevant facts and hallucination with dynamically synthesised prompts based on the current step context.

The paper does confirm the now accepted approach to prompting which include:

  1. Prompts need to be contextual, including previous conversation context and dialog turns.
  2. Some kind of automation will have to be implemented to collate, and in some instances summarise previous dialog turns to be included in the prompt.
  3. Supplementary data, acting as a contextual reference for the LLM, needs to be selected, curated and truncated to an efficient length for each prompt at inference.
  4. The prompt needs to be formed in such a way, not to increase inference time.

I’m currently the Chief Evangelist @ HumanFirst. I explore & write about all things at the intersection of AI & language; ranging from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces & more.

Subscribe to HumanFirst Blog

Get the latest posts delivered right to your inbox