Beyond Elicitation: Provision-based Prompt Optimization for Knowledge-Intensive Tasks

1. Problem Statement

Current automated prompt optimization methods (e.g., OPRO, APO, PromptAgent) rely on elicitation-based strategies. These methods assume that optimal prompts can unlock latent capabilities within a Large Language Model's (LLM) pre-trained parameters through linguistic reformulation, instruction tuning, or reasoning templates.

However, the authors identify a fundamental limitation: Knowledge Poverty. In knowledge-intensive domains (e.g., specialized medicine, law, finance), tasks often require specific factual knowledge, terminology precision, and reasoning patterns that are not present in the model's parametric memory.

• The Core Issue: Elicitation-based methods attempt to "ask" the model to recall non-existent knowledge. When the model lacks the necessary domain facts, no amount of prompt engineering can elicit the correct answer.
• Observed Failures: Experiments show that traditional methods often yield marginal or negative improvements on knowledge-intensive benchmarks. They tend to produce "shallow" improvements by matching surface patterns rather than resolving underlying knowledge gaps, leading to optimization ineffectiveness.

2. Methodology: Knowledge-Provision-based Prompt Optimization (KPPO)

The authors propose KPPO, a framework that shifts the paradigm from eliciting latent knowledge to providing external domain knowledge directly within the prompt. KPPO operates through three core technical innovations:

A. Knowledge Gap Filling (Gradient-based Analysis)

Instead of random search, KPPO uses an Optimizer LLM ($LLM_{opt}$) to analyze failure cases systematically:

1. Error Explanation: Analyzes why the current prompt failed.
2. Knowledge Gap Identification: Pinpoints specific missing domain facts or reasoning rules required to solve the error.
3. Modification Generation: Generates "gradients" (targeted modifications) to inject specific domain knowledge into the prompt.

• Structure: The integrated knowledge is organized into a Directed Acyclic Graph (DAG) hierarchy (Topic nodes and Note nodes) to ensure structured, logical accumulation of facts.

B. Batch-wise Dual-Objective Evaluation

To ensure robustness and efficiency, KPPO avoids full validation set evaluation (which is computationally expensive and prone to overfitting). Instead, it uses a batch-wise evaluation inspired by Trust Region Policy Optimization:

• Objective: Maximize performance improvement ($\Delta S$) on recent failure cases while minimizing distributional divergence ($D$).
• Mechanism: Candidates are ranked by a lexicographic order: first by performance gain on the recent batch, then by stability (low divergence from the previous prompt's output distribution). This prevents the model from overfitting to spurious patterns in a small batch.

C. Adaptive Knowledge Pruning

As knowledge is iteratively added, prompts can become excessively long, increasing inference costs. KPPO introduces a pruning mechanism guided by structural constraints:

• Local Degree Constraint: Limits the number of child nodes per topic to control information density.
• Global Balance Constraint: Ensures the hierarchy remains balanced (preventing "flat" topics with too many children).
• Process: The system detects violations of these constraints and instructs the $LLM_{opt}$ to prune redundant or less critical information, maintaining essential domain knowledge while reducing token usage.

3. Key Contributions

1. Paradigm Shift: Proposes KPPO, the first framework to treat prompt optimization as systematic knowledge integration rather than capability elicitation.
2. Technical Innovations:
   • A Knowledge Gap Filling mechanism that identifies and targets specific missing facts.
   • A Dual-Objective Evaluation strategy that balances performance gains with distributional stability.
   • An Adaptive Pruning strategy that reduces inference token consumption without sacrificing accuracy.
3. Comprehensive Evaluation: Validated across 15 knowledge-intensive benchmarks spanning Financial, Legal, and Medical domains.

4. Experimental Results

The paper evaluates KPPO on Llama 3.1 and Qwen 2.5 against strong baselines (OPRO, APO, PromptAgent, RAG, Verify-and-Edit).

• Performance Gains:
  • KPPO achieves an average accuracy improvement of ~6.1% over the best elicitation-based baselines.
  • On specific tasks like MedConceptsQA (where parametric knowledge is non-existent), KPPO improved accuracy from 28.3% to 51.7% on Llama 3.1, significantly outperforming RAG and tool-augmented methods.
• Efficiency:
  • The Pruning Variant reduces inference token usage by 21.7% (Llama) and 29.0% (Qwen) while maintaining performance superior to all baselines.
  • KPPO demonstrates significantly lower latency (3.88s) compared to web-search-augmented methods (36–37s).
• Knowledge Quality:
  • Analysis shows KPPO integrates 26x more domain knowledge points than OPRO and maintains a 92.5% faithfulness rate.
  • Unlike baselines that optimize for global validation metrics (leading to spurious correlations), KPPO's batch-wise evaluation ensures genuine knowledge gap remediation.

5. Significance

• Solving the "Knowledge Ceiling": KPPO addresses the critical bottleneck where LLMs fail on specialized tasks due to missing training data. It provides a computationally efficient alternative to fine-tuning (which requires massive data and compute) by injecting knowledge directly into the inference context.
• Practical Deployment: By combining knowledge provision with adaptive pruning, KPPO makes LLMs viable for specialized domains (e.g., medical diagnosis, legal analysis) where accuracy and token cost are critical constraints.
• New Research Direction: The paper establishes provision-based optimization as a distinct and necessary paradigm for the next generation of prompt engineering, moving beyond the limitations of static model knowledge.