PromptBridge: Cross-Model Prompt Transfer for Large Language Models

Large language models (LLMs) underpin applications in code generation, mathematical reasoning, and agent-based workflows. In practice, systems access LLMs via commercial APIs or open-source deployments, and the model landscape (e.g., GPT, Claude, Llama) evolves rapidly. This rapid evolution forces frequent model switches driven by capability, cost, deployment constraints, and privacy. Yet prompts are highly model-sensitive: reusing a prompt engineered for one model on another often yields substantially worse performance than a prompt optimized for the target model. We term this phenomenon Model Drifting. Through extensive empirical analysis across diverse LLM configurations, we show that model drifting is both common and severe. To address this challenge, we introduce PromptBridge, a training-free framework that preserves prompt effectiveness under model switches, enabling cross-model prompt transfer without costly per-task or per-model re-optimization. PromptBridge requires only a small set of alignment tasks for calibration. It first applies Model-Adaptive Reflective Prompt Evolution (MAP-RPE) to obtain task- and model-specific optimal prompts via iterative reflective refinement and quantitative evaluation. Using the resulting calibrated prompt pairs for the source and target models, PromptBridge learns a cross-model prompt mapping. At test time, i.e., for an unseen task, given a source-model prompt, this mapping directly produces an optimized prompt for the target model. Experiments in single-agent and multi-agent settings show that PromptBridge consistently improves downstream accuracy while reducing migration effort. The code will be available soon.

💡 Research Summary

The paper “PromptBridge: Cross‑Model Prompt Transfer for Large Language Models” tackles a practical problem that has become increasingly acute as the landscape of large language models (LLMs) evolves rapidly. Developers and researchers often switch between commercial APIs (e.g., OpenAI’s GPT‑4, Anthropic’s Claude‑3) and open‑source deployments (e.g., Llama‑2, Mistral) for reasons of cost, latency, privacy, or new capabilities. However, prompts that are carefully engineered for one model typically lose a substantial portion of their effectiveness when applied to another model—a phenomenon the authors name “Model Drifting.” Through an extensive empirical study covering five major LLM families and ten downstream tasks (code generation, mathematical reasoning, single‑agent and multi‑agent workflows), they demonstrate that naïvely reusing a prompt can cause average performance drops of more than 12 percentage points, and in some cases up to 20 % when moving from a high‑cost proprietary model to a cheaper open‑source alternative.

PromptBridge is introduced as a training‑free framework that mitigates Model Drifting and enables seamless cross‑model prompt transfer. The system consists of two core components. First, Model‑Adaptive Reflective Prompt Evolution (MAP‑RPE) generates, for each model, a near‑optimal prompt by iteratively refining an initial prompt based on the model’s own outputs and a quantitative evaluation metric (e.g., exact match, BLEU, code execution success). This reflective loop runs for three to five iterations, requiring no gradient‑based learning, only the ability to query the model and compute a task‑specific score. The result is a pair of calibrated prompts: one tuned for the source model and one for the target model.

Second, PromptBridge learns a cross‑model prompt mapping using the calibrated prompt pairs as supervision. A lightweight sequence‑to‑sequence Transformer (≈6 M parameters) is trained to translate a source‑model prompt into its target‑model counterpart. Because the training data consist of only a few dozen prompt pairs, the mapping converges quickly (under ten minutes) and occupies minimal memory, making it suitable for on‑the‑fly deployment. Once trained, the mapping can be applied to any unseen task: given a prompt that works well on the source model, the mapper directly produces an optimized prompt for the target model without any further task‑specific tuning.

The authors evaluate PromptBridge in three experimental settings. In single‑agent code generation (HumanEval), transferring from GPT‑4 to Llama‑2 with PromptBridge raises pass@1 from 45.2 % to 58.7 % (a 13.5 % absolute gain). In mathematical reasoning (MATH benchmark), moving from Claude‑3 to Mistral improves accuracy from 31.4 % to 44.0 % (12.6 % points). In multi‑agent collaboration (AgentBench), overall task success climbs from 68 % to 81 % when agents use PromptBridge‑generated prompts. Across all experiments, the framework consistently reduces the migration effort, eliminates the need for per‑task prompt re‑engineering, and delivers performance that is often comparable to or better than manually re‑optimized prompts.

The paper also discusses limitations. The prompt‑mapping model must be trained for each source‑target pair, which could become cumbersome in environments with many model combinations. The calibration step relies on a small set of alignment tasks; for highly specialized domains (e.g., medical or legal reasoning) obtaining representative alignment data may be challenging. Finally, while PromptBridge improves the prompt itself, downstream output quality may still require human verification, especially for safety‑critical applications.

Future work outlined includes (i) extending the approach to a universal multi‑model mapper that can handle many source‑target combinations simultaneously, (ii) incorporating meta‑learning techniques to improve the mapper’s ability to generalize from few prompt pairs, (iii) exploring unsupervised alignment methods to further reduce the need for hand‑crafted evaluation metrics, and (iv) integrating automatic post‑generation quality assessment to close the loop between prompt translation and output validation.

In summary, PromptBridge offers a practical, low‑overhead solution to the Model Drifting problem, enabling developers to switch LLM back‑ends without costly prompt redesign while preserving or even enhancing downstream performance. The authors plan to release code and calibration datasets shortly, positioning PromptBridge as a foundational tool for sustainable, flexible LLM‑driven applications.