Zhenting Wang

Large language model (LLM) systems frequently switch between models due to evolving capabilities, cost, and deployment constraints. However, prompts are highly model-sensitive: prompts engineered for one model often degrade substantially when reused on another. We term this phenomenon \emph{Model Drifting}, and show through extensive empirical analysis that it is common and severe across diverse LLM configurations. To address this, we propose \emph{PromptBridge}, a training-free framework for cross-model prompt transfer that preserves prompt effectiveness under model switches. It uses a set of alignment tasks to calibrate source and target models via \emph{Model-Adaptive Reflective Prompt Evolution (MAP-RPE)}, and learns a prompt mapping that translates a source-model prompt into an optimized target-model prompt without per-task or per-model re-optimization. Experiments across single- and multi-agent benchmarks demonstrate that \emph{PromptBridge} consistently improves downstream performance while reducing migration effort. For instance, when transferring prompts to \texttt{o3}, \emph{PromptBridge} improves accuracy by 27.39% on \textsc{SWE-Bench} and 39.44% on \textsc{Terminal-Bench} over direct transfer, establishing cross-model prompt transfer as a practical requirement for sustainable LLM system development.

Anthropic proposes the concept of skills for LLM agents to tackle multi-step professional tasks that simple tool invocations cannot address. A tool is a single, self-contained function, whereas a skill is a structured bundle of interdependent multi-file artifacts. Currently, skill generation is not only label-intensive due to manual authoring, but also may suffer from human–machine cognitive misalignment, which can lead to degraded agent performance, as evidenced by evaluations on SkillsBench. Therefore, we aim to enable agents to autonomously generate skills. However, existing self-evolving methods designed for tools cannot be directly applied to skills due to their increased complexity. To address these issues, we propose CoEvoSkills, a self-evolving skills framework that enables agents to autonomously construct complex, multi-file skill packages. Specifically, CoEvoSkills couples a Skill Generator that iteratively refines skills with a Surrogate Verifier that co-evolves to provide informative and actionable feedback without access to ground-truth test content. On SkillsBench, CoEvoSkills achieves the highest pass rate among five baselines on both Claude Code and Codex, and generalizes strongly to six additional LLMs. The code is publicly available at https://github.com/Zhang-Henry/CoEvoSkills.