Prompt Migration: Adapting AI Prompts

• Prompt migration is the systematic adaptation of prompts across evolving models or task settings to preserve performance and maintain output quality.
• It employs methods such as cross-model transfer, continual prompt optimization, and regression testing to address issues like model drift and format violations.
• Research demonstrates that balancing explicit redesign with preservation of key prompt knowledge is critical for stable and effective GenAI applications.

Searching arXiv for papers on prompt migration, cross-model prompt transfer, and migration-related prompting frameworks. Prompt migration is the systematic adaptation of prompts when the underlying LLM, task setting, or migration target changes, with the goal of preserving or improving task behavior despite model-sensitive prompt semantics. In current research, the term is used in at least three closely related senses: cross-model prompt transfer for LLM replacement, prompt redesign for software migration and modernization tasks, and prompt-aware stabilization of production GenAI systems under model drift. Across these settings, the common problem is that prompts encode nontrivial task knowledge, output constraints, and behavioral assumptions that often cease to hold when a model, domain, or workflow changes. Recent work has therefore treated prompts not as static strings but as migratable artifacts requiring calibration, evaluation, and lifecycle management (Wang et al., 1 Dec 2025, Tripathi et al., 8 Jul 2025, Davari et al., 14 Jul 2025).

1. Conceptual scope and problem definition

Prompt migration is explicitly defined as “a systematic process for adapting and redesigning the prompts used in a GenAI application when the underlying LLM is changed or updated, so that the application’s behavior (accuracy, format, semantics) remains stable” in a production case study on enterprise search (Tripathi et al., 8 Jul 2025). A related line of work formalizes the same phenomenon as cross-model prompt transfer: given a prompt tuned for a source model $M_s$, generate an effective prompt for a target model $M_t$ without re-optimizing from scratch on every downstream task (Wang et al., 1 Dec 2025). Another closely aligned formulation is Continual Prompt Optimization, in which a prompt $p_t$ optimized for model $\mathcal{M}_t$ is migrated to a new model $\mathcal{M}_{t+1}$ while preserving useful task knowledge embedded in earlier prompt versions (Davari et al., 14 Jul 2025).

The underlying motivation is that prompts are highly model-sensitive. PromptBridge terms the resulting degradation under model replacement “Model Drifting” and defines the transfer gap as the difference between target-model performance under a source-optimal prompt and target-model performance under its own optimal prompt (Wang et al., 1 Dec 2025). The paper reports that on HumanEval, a GPT-5-optimized prompt achieving 99.39% on GPT-5 drops to 68.70% on Llama-3.1-70B-Instruct, जबकि Llama-3.1-70B’s own optimal prompt reaches 79.47% on itself (Wang et al., 1 Dec 2025). This establishes prompt migration as a distinct problem from ordinary prompt engineering: the issue is not merely improving a prompt, but preserving functionality under model replacement.

The same problem appears in production systems. In the Tursio enterprise-search case study, reusing prompts across GPT model upgrades caused regression-test failures and parser-breaking output changes. For a SQL-filter extraction prompt, the same prompt yielded 100% tests passed on GPT-4-32k, 98% on GPT-4.1, and 97.3% on GPT-4.5-preview before migration-specific redesign (Tripathi et al., 8 Jul 2025). The observed failures included missing or incorrect ORDER BY and GROUP BY, invented column names, semantic misinterpretations, and output-format violations (Tripathi et al., 8 Jul 2025).

A third usage of the term arises in software engineering tasks where prompting itself becomes part of the migration workflow. LibRec, for example, operationalizes what its authors describe as “prompting for migration”: retrieval-augmented prompts, in-context migration examples, and intent-aware prompt design for library replacement recommendations (&&&10&&&). In this sense, prompt migration includes transferring prompt structures across models and related subtasks such as intent extraction, intent-type classification, analogous-library detection, and target-library recommendation (Han et al., 13 Aug 2025).

2. Principal research strands

The current literature clusters into three main strands.

The first strand studies cross-model prompt transfer directly. “PromptBridge: Cross-Model Prompt Transfer for LLMs” develops a training-free framework for mapping prompts from one model family to another using a small set of alignment tasks, model-specific prompt optimization, and an LLM-mediated mapping extractor (Wang et al., 1 Dec 2025). “Rethinking Prompt Optimization: Reinforcement, Diversification, and Migration in Blackbox LLMs” frames prompt migration as Continual Prompt Optimization and addresses preservation of prompt knowledge during model changes through positive reinforcement and feedback diversification (Davari et al., 14 Jul 2025). “RETAIN: Interactive Tool for Regression Testing Guided LLM Migration” emphasizes regression testing, data slicing, and interactive error discovery during migration from one model to another (Dixit et al., 2024). “Prompt Migration: Stabilizing GenAI Applications with Evolving LLMs” treats prompt migration as prompt lifecycle management in a production application (Tripathi et al., 8 Jul 2025).

The second strand studies prompt migration in software migration and modernization workflows. “LibRec: Benchmarking Retrieval-Augmented LLMs for Library Migration Recommendations” formulates library migration recommendation as ranked prediction over source-target library pairs with migration intent and intent type, and shows that prompting strategy materially affects recommendation quality (Han et al., 13 Aug 2025). “Automatic Library Migration Using LLMs: First Results” evaluates Zero-Shot, One-Shot, and Chain-of-Thought prompts for SQLAlchemy 1→2 migration and finds that One-Shot prompting produces the best overall migration outcome (Almeida et al., 2024). “Token Optimization Strategies for LLM-Based Oracle-to-PostgreSQL Migration” studies prompt/context transformation under token budgets and explicitly argues that token optimization is a migration problem balancing semantic preservation, validity, and token cost rather than mere prompt shortening (Grynets et al., 27 May 2026).

The third strand studies prompt design portability across prompt languages and task directions. “Evaluating LLMs for Code Translation: Effects of Prompt Language and Prompt Design” compares concise versus detailed prompts and English versus Arabic prompt language for source-code translation, reporting that detailed prompts consistently improve quality and that English prompts outperform Arabic by 13–15% (Aljagthami et al., 16 Sep 2025). Although framed around code translation rather than model replacement, the paper is directly relevant because it shows that prompt migration across natural languages is nontrivial even when task semantics are fixed (Aljagthami et al., 16 Sep 2025).

3. Mechanisms of prompt drift

Prompt drift arises from systematic differences in model training, interface conventions, and alignment. PromptBridge attributes Model Drifting to differences in training corpora and distributions, tokenization and role systems, alignment and human-feedback procedures, and API conventions (Wang et al., 1 Dec 2025). This explains why the same textual prompt can induce different behavior not only across model families such as GPT, Llama, and Qwen, but even within model families and across model sizes (Wang et al., 1 Dec 2025).

Empirically, the phenomenon is both common and severe. PromptBridge reports that diagonal entries in source-model/target-model transfer matrices are consistently highest, while off-diagonal transfers can lose 50–70% relative accuracy on xCodeEval even between different sizes of the same family (Wang et al., 1 Dec 2025). On SWE-Bench Verified, directly transferring an o4-mini-optimized prompt to o3 yields 33.40% resolved issues, while PromptBridge reaches 46.00% after migration-specific adaptation (Wang et al., 1 Dec 2025). On Terminal-Bench, PromptBridge yields a 39.44% relative improvement versus direct transfer for target o3 (Wang et al., 1 Dec 2025).

The Tursio study reveals a complementary production perspective. GPT-4.1 required more explicit instructions, missed implicit trends such as time ordering for “over time” queries, failed on relative time expressions unless explicitly guided, and tended to select only explicitly mentioned fields (Tripathi et al., 8 Jul 2025). GPT-4.5-preview introduced additional output-format violations and extra informational text that downstream parsers could not consume (Tripathi et al., 8 Jul 2025). These results suggest that prompt drift is not restricted to accuracy alone; it includes format drift, inference-policy drift, and differing tolerance for implicit task assumptions.

In black-box optimization settings, prompt drift also appears as destructive forgetting during adaptation. The BReAD/CPO paper shows that naively re-optimizing an expert prompt on a new model can damage precisely those prompt components that encoded valuable task knowledge on the old model (Davari et al., 14 Jul 2025). On Geometric Shapes, a GPT-3.5 expert prompt transferred to GPT-4o and then optimized with a baseline negative-feedback-only APO method reaches 75.1, whereas optimizing from a default prompt reaches 79.0; this demonstrates that migration can underperform clean re-optimization if prior prompt knowledge is not explicitly preserved (Davari et al., 14 Jul 2025).

4. Methods for prompt migration

A central divide in the literature is between training-free mapping methods, feedback-driven optimization methods, and regression-testing-driven redesign.

PromptBridge exemplifies the first category. It first uses Model-Adaptive Reflective Prompt Evolution (MAP-RPE) to obtain task- and model-specific optimized prompts $p^*_{M,S_i}$ on a small alignment suite (Wang et al., 1 Dec 2025). MAP-RPE is a gradient-free, reflective, population-based optimization method using a reflection model, quantitative evaluation, and island-based diversity maintenance (Wang et al., 1 Dec 2025). For each prompt candidate, it combines task performance and a behavioral score, with $\lambda = 0.8$ in the scoring rule $$\text{Score}(p) = \lambda \cdot \text{Perf}(p) + (1 - \lambda) \cdot \text{Behavior}(p)$$ (Wang et al., 1 Dec 2025). Using calibrated source–target prompt pairs, PromptBridge then asks a powerful meta-model to summarize the common differences into a transfer-effects summary and uses an adapter model to transform unseen source prompts for the target model (Wang et al., 1 Dec 2025). The crucial claim is that prompt migration can be represented as a reusable model-level editing rule rather than an independent prompt search on every task (Wang et al., 1 Dec 2025).

The BReAD framework represents the second category. It reinterprets textual gradients from wrong predictions as negative reinforcement and introduces positive reinforcement from correct predictions to preserve beneficial prompt components (Davari et al., 14 Jul 2025). It further reduces feedback noise through feedback diversification, generating multiple feedback samples and summarizing the consistent portions into consolidated guidance (Davari et al., 14 Jul 2025). The migration-specific insight is scheduling: in ordinary prompt optimization, positive reinforcement introduced too early can freeze weak prompts, with best results appearing when positive reinforcement starts at iterations 3–4, but in migration from an already strong expert prompt it should be enabled from iteration 1 to prevent overwriting critical instructions (Davari et al., 14 Jul 2025).

RETAIN represents the third category: prompt migration as interactive regression testing. It provides an evaluation interface, prompt editor, runs tracker, and an error discovery module that compares corpora of outputs from source and target models and generates goal-oriented textual descriptions of their differences (Dixit et al., 2024). These descriptions can then be converted into LLM-based assertions that become reusable migration guards (Dixit et al., 2024). This approach shifts the emphasis from optimizing prompts in the abstract to identifying regressions on meaningful data slices and iterating prompt changes against those slices (Dixit et al., 2024).

The Tursio framework is a production-oriented redesign methodology rather than a formal optimization algorithm. Its workflow includes migration-testbed evaluation, prompt redesign guided by vendor prompt guides, retesting on the migration testbed, and subsequent full regression testing on production-like workloads (Tripathi et al., 8 Jul 2025). Its main intervention is increased explicitness: role, structured instructions, output format, examples, and reasoning scaffolding (Tripathi et al., 8 Jul 2025).

5. Prompt migration in software migration tasks

Software migration tasks offer especially clear demonstrations of how prompting strategy itself must migrate across subtasks, models, and modernization workflows.

LibRec defines library migration as replacing a source library $S_j$ with a target library $T_j$, with an associated migration intent $M_t$0 and intent type $M_t$1, constructing a migration dataset $M_t$2 and a query $M_t$3 for recommendation (Han et al., 13 Aug 2025). The framework uses GPT-4o to extract migration intents from commit messages, classifies intent types, retrieves the top-3 similar historical migration records with BM25, and prompts an LLM to recommend top-10 alternative libraries (Han et al., 13 Aug 2025). On LibEval, which contains 2,888 migration records associated with 2,368 libraries extracted from 2,324 Python repositories, the strongest full-system performance comes from Claude-3.7-Sonnet with Precision@10 of 0.792 and MRR of 0.652 (Han et al., 13 Aug 2025).

LibRec also offers one of the clearest empirical analyses of prompt-strategy migration. It compares Vanilla prompting, prompts without retrieval, prompts without intent, and full LibRec prompts, as well as One-shot and Chain-of-Thought variants across ten LLMs (Han et al., 13 Aug 2025). The average ablation results show that full LibRec reaches P@10 of 0.715 and MRR of 0.585, versus 0.615 and 0.469 for Vanilla (Han et al., 13 Aug 2025). In prompt-paradigm comparisons, One-shot prompting emerges as the most robust and reliable across 6 of 10 models and yields the highest Precision@10 for GPT-4 at 72.9% and DeepSeek-V3 at 71.1%, whereas CoT is best only for Claude-3.7-Sonnet, DeepSeek-R1, and QwQ-Plus, with Claude-3.7-Sonnet reaching Precision@10 = 80.3% under CoT with RAG (Han et al., 13 Aug 2025). This is a direct example of cross-model prompt portability being conditional rather than universal (Han et al., 13 Aug 2025).

The SQLAlchemy migration study reaches a closely related conclusion in a narrower setting. In migrating a FastAPI application from SQLAlchemy 1.x to 2.x using GPT-4, One-Shot prompting was the only evaluated strategy that produced a running application with all 4/4 tests passing, while Zero-Shot and Chain-of-Thought versions did not run successfully in the reported setup (Almeida et al., 2024). The One-Shot prompt also successfully migrated all 4/4 columns, whereas Zero-Shot migrated 0/4 columns (Almeida et al., 2024). This is strong evidence that prompt migration in API modernization often depends on worked examples rather than bare instructions.

Oracle-to-PostgreSQL migration adds a different dimension: prompt/context transformation under token constraints. The paper formalizes LLM-based migration as $M_t$4, where $M_t$5 is a token-optimization strategy applied to the source artifact and context (Grynets et al., 27 May 2026). Among twelve strategies, mild context pruning preserves semantic quality almost at baseline levels, achieving 89.75% Semantic Match on the 100-query sample versus 89.80% for the baseline, while adaptive routing reduces input tokens by 8.72% and output tokens by 5.49% with 88.40% Semantic Match and a 6.67% increase in Token Efficiency (Grynets et al., 27 May 2026). Aggressive schema distillation yields a 132.22% gain in Token Efficiency but incurs a 44.50-percentage-point loss in Semantic Match (Grynets et al., 27 May 2026). These findings suggest that prompt migration across software-modernization workflows must often be treated as multi-objective optimization over context length, syntax, semantic preservation, and structural fidelity rather than purely as instruction wording.

6. Prompt language, structure, and portability

Prompt migration is not limited to changing models; it also includes changing prompt language and task instantiation while preserving behavior. The code-translation study evaluates four LLMs, a TransCoder baseline, two prompt styles, and two prompt languages over all directed pairs among C++, Java, Python, and C# (Aljagthami et al., 16 Sep 2025). Its central findings are that detailed prompts consistently improve both BLEU and CodeBLEU across models and directions, and English prompts outperform Arabic prompts by 13–15% (Aljagthami et al., 16 Sep 2025).

The prompt-style comparison is especially relevant to portability. The concise prompt is simply “Please translate this code from X to Y,” whereas the detailed version adds ten explicit requirements, including preserving logic, variable names where syntactically valid, comments and relative positions, control-flow structure, equivalent data structures and types, identical outputs for identical inputs, necessary imports, and code organization (Aljagthami et al., 16 Sep 2025). Because the same detailed template improves performance across all evaluated directions and models, the paper suggests that some prompt structures are portable at the task-template level even if wording or language is not (Aljagthami et al., 16 Sep 2025).

The prompt-language result introduces an important caveat. Even when structure and task semantics are preserved, moving from English to Arabic prompt language causes a substantial drop in performance, which the paper attributes implicitly to training-data skew and the stronger alignment of code with English-language instructions and documentation (Aljagthami et al., 16 Sep 2025). This implies that prompt migration across natural languages may preserve structure but not necessarily behavior.

The production findings from Tursio are consistent with this emphasis on structure. When adapting prompts to GPT-4.1, the redesigned SQL-filter extraction prompt replaced a terse instruction with explicit format rules, instruction bullets, examples, and directives to infer implicit columns and preserve wording in filter values (Tripathi et al., 8 Jul 2025). The result was restoration of 100% test pass rates for the task after migration on GPT-4.1 (Tripathi et al., 8 Jul 2025). Together, these results indicate that prompt migration often succeeds by increasing structural explicitness rather than by preserving the exact surface form of earlier prompts.

7. Evaluation methodologies and benchmarks

The methodological literature around prompt migration converges on regression testing, alignment-task calibration, and goal-specific error analysis.

PromptBridge evaluates migration on single-agent code generation, agentic software engineering, terminal interaction, planning, and multi-agent systems (Wang et al., 1 Dec 2025). Its baselines include direct transfer, a zero-shot GPT-5 prompt optimizer, MIPROv2, GEPA, and in-context-learning baselines (Wang et al., 1 Dec 2025). Across target models o3, o4-mini, and Llama-3.1-70B, PromptBridge improves over direct transfer on HumanEval, MBPP, APPS, xCodeEval, and CodeContests, for example raising HumanEval from 68.70 to 79.88 on Llama-3.1-70B and CodeContests from 48.61 to 56.36 on o3 (Wang et al., 1 Dec 2025).

RETAIN emphasizes interactive regression testing rather than static benchmark reporting. In its user study, participants using RETAIN identified 165 errors versus 86 with manual evaluation, experimented with 4.55 prompt edits on average versus 2.6 in a manual setup, and achieved a final BERTScore of 0.704 versus 0.625 in the prompt-migration task (Dixit et al., 2024). The automatic evaluation of its goal-driven error discovery module reports Error Relevance of 0.87 with goal conditioning versus 0.72 without, and selector precision/recall of 0.69/0.38 with goals (Dixit et al., 2024). These results support the view that prompt migration benefits from slice-aware, goal-aware analysis rather than pure aggregate metrics.

The Tursio framework uses a curated migration testbed of 110 questions divided into easy, moderate, and hard categories based on explicit versus implicit operations and prior failures (Tripathi et al., 8 Jul 2025). Its pass-rate metric is straightforward but operationally relevant because it measures the ability of a prompt–model pair to satisfy structural and semantic expectations in a production pipeline (Tripathi et al., 8 Jul 2025). A notable result is that prompt migration reduced the time required for subsequent model migrations from “a couple of months” to “a couple of weeks” once the testbed and workflow were in place (Tripathi et al., 8 Jul 2025).

Software-migration benchmarks are similarly task-specific. LibRec introduces LibEval and evaluates Precision@k and MRR on top-10 library recommendations (Han et al., 13 Aug 2025). Oracle-to-PostgreSQL migration uses Valid Syntax Rate, Exact Match, Semantic Match, CodeBLEU, and Token Efficiency on 10- and 100-query samples (Grynets et al., 27 May 2026). The SQLAlchemy study evaluates whether the application runs successfully, the number of tests passing, Pylint score, Pyright errors, and counts of migrated columns and methods (Almeida et al., 2024). Across these works, a common pattern emerges: prompt migration is best evaluated against end-task metrics and downstream compatibility checks rather than prompt-level heuristics alone.

8. Common failure modes and controversies

Several recurrent failure modes appear across the literature.

One is destructive prompt rewriting during adaptation. The BReAD paper shows that standard APO methods focused only on correcting errors can erase useful instructions from expert prompts, causing migrated prompts to underperform both direct transfer and fresh optimization (Davari et al., 14 Jul 2025). This is particularly salient in migration settings because prior prompts often encode accumulated task knowledge.

A second is format drift. Tursio observed parser-breaking output-format violations and extraneous informational text after model upgrades, especially on GPT-4.5-preview (Tripathi et al., 8 Jul 2025). This highlights a misconception in practical deployment: a model can remain semantically strong while becoming operationally unusable if it no longer respects strict output contracts.

A third is overreliance on implicit inference. GPT-4-32k in the Tursio case tolerated flexible prompts that assumed the model would infer trends, relative time expressions, or intended columns, whereas GPT-4.1 required much more explicit instruction (Tripathi et al., 8 Jul 2025). The controversy here is whether stronger models should require less explicit prompting. The available evidence suggests that “smarter” models do not uniformly reduce the need for prompt migration; they may alter the locus of required specificity instead (Tripathi et al., 8 Jul 2025).

A fourth is the assumption that reasoning prompts always transfer best. LibRec contradicts that assumption by showing that One-shot prompting is more robust than CoT for most models, and that CoT advantages are concentrated in reasoning-optimized models such as DeepSeek-R1, QwQ-Plus, and Claude-3.7-Sonnet (Han et al., 13 Aug 2025). This matters because prompt migration often fails when developers transfer a favored reasoning pattern from one model class to another without recalibration.

There are also limitations in current methods. PromptBridge relies on access to a powerful meta-model such as GPT-5; its appendix reports substantially worse performance when weaker models are used for mapping extraction and adaptation (Wang et al., 1 Dec 2025). RETAIN’s error discovery depends on LLM-based judgments with relatively low recall (Dixit et al., 2024). The BReAD migration experiments are limited to GPT-3.5-turbo → GPT-4o on five tasks (Davari et al., 14 Jul 2025). Tursio’s results derive from one enterprise-search application (Tripathi et al., 8 Jul 2025). These constraints imply that prompt migration remains partly empirical and system-specific rather than fully standardized.

9. Lifecycle management and future directions

A consistent conclusion across the literature is that prompts should be managed as versioned, testable assets. The Tursio study argues explicitly for prompt lifecycle management, with prompts evolving alongside both application changes and model releases (Tripathi et al., 8 Jul 2025). RETAIN operationalizes this view with versioned runs, reusable metrics, and migration-oriented regression slices (Dixit et al., 2024). PromptBridge offers a way to make this lifecycle less costly by converting prompt migration into a reusable model-pair mapping problem (Wang et al., 1 Dec 2025).

Several research directions follow naturally. One is standardized benchmarks for prompt portability across models, prompt languages, and task domains, a need explicitly identified in PromptBridge’s discussion of prompt portability evaluation (Wang et al., 1 Dec 2025). Another is better formalization of Continual Prompt Optimization, since the BReAD paper notes that CPO is conceptually motivated but not yet captured by explicit regularization or retention objectives over prompt content (Davari et al., 14 Jul 2025). A third is domain-specific migration suites that combine semantic, structural, and operational metrics, as seen separately in LibEval, RETAIN, and Oracle2PostgreSQL but not yet unified (Han et al., 13 Aug 2025, Dixit et al., 2024, Grynets et al., 27 May 2026).

There is also growing evidence that prompt migration should be model-aware rather than purely task-aware. LibRec’s ablations show that some models benefit more from intent-aware prompts while others benefit more from retrieval demonstrations (Han et al., 13 Aug 2025). PromptBridge’s transfer summaries similarly encode systematic stylistic differences between model pairs (Wang et al., 1 Dec 2025). A plausible implication is that future prompt-migration systems will maintain explicit model-profile abstractions, capturing preferred prompt structures, verbosity, reasoning depth, and output-format tolerance.

Finally, production-oriented work suggests that prompt migration will become a permanent engineering discipline rather than an episodic task. Tursio’s experience with successive GPT-4-family deprecations demonstrates that model turnover can force repeated migrations on timelines set by providers rather than application teams (Tripathi et al., 8 Jul 2025). In that environment, stable prompt migration workflows, regression testbeds, and versioned assertions are likely to become as routine as schema migration and API compatibility testing.

Prompt migration has thus emerged as a unifying problem at the intersection of prompt engineering, model evaluation, and software maintenance. Whether the target is cross-model transfer, library modernization, code translation, or enterprise search stabilization, the empirical record increasingly supports the same conclusion: prompts are model-contingent behavioral interfaces, and preserving application semantics under change requires explicit migration methodology rather than prompt reuse by assumption (Wang et al., 1 Dec 2025, Tripathi et al., 8 Jul 2025, Davari et al., 14 Jul 2025).