Tag
讨论指出 MCP 工具定义会占用大量常驻上下文——有用户卸载一个大型 MCP server 后上下文从约 200K 降至约 10K tokens。文章提供了如何在首次对话前检查工具开销的实用建议，包括检查请求日志、按需禁用 server 及使用延迟加载。
提出 StruPO，一种结构感知的提示编辑器，利用校准场景图作为共享表示，将结构反馈转化为可复用的编辑策略训练信号，在保持场景忠实性的同时提升文生图的美学质量。
GEPA-Curator is a framework that improves agentic prompt optimization by filtering noisy failure signals and retrieving prior successful trajectories, achieving gains of up to 20.7% over GEPA and 21.2% over ACE on benchmarks like GoBrowse, ALFWorld, AppWorld, and τ-bench.
提出了 RSIHub，一个用于构建、比较和扩展 agent-evolution 方法的开源框架，通过可复用的算子与配方简化实验基础设施。论文在 Terminal Bench 2 和 τ³ Banking 上评测了 AHE、Hyperagents、A-Evolve 与 GEPA，最大 pass@1 提升达 13.5 个百分点，但方法排名随设置变化且优化收益不一定迁移到留出任务。
该论文提出一个可复现的离线评估框架，用于在生产环境的可变企业数据上评测对话智能体，并将其作为 GEPA 提示优化的适应度函数，在正确性约束下实现提示压缩 3.5-7.6% 而不损失任务准确率。
CEPO is a general framework for prompt optimization in multi-agent systems that separates shared system-level coordination from agent-specific execution prompts, consistently improving four existing prompt optimizers across 11 benchmarks in four domains.
RSIHub 是一个开源框架，为智能体进化（agent-evolution）研究提供可组合的共享基础设施，标准化方法的构建、比较与扩展，并在 Terminal Bench 2 和 τ3 Banking 上评测了 AHE、Hyperagents、A-Evolve 和 GEPA 等方法。
The paper proposes a two-stage framework for reflective prompt optimization that decouples candidate-pool construction from validation allocation, achieving up to 11.22 percentage point gains over the initial solver on five benchmarks using Qwen3.5-9B and Llama-3.1-8B-Instruct.
HERO 提出结合离线与测试时的 harness 优化方法，通过执行验证的成功与失败样本来学习可复用的 prompt 更新，并在测试时利用执行反馈对生成代码进行精调，在 SQL、Cypher 和 Python 基准上将 pass@1 提升约 9.6 个点。
提出了 Reflective Feedback-Driven Prompt Optimization (RFDPO)，通过追踪 item 级别的恢复、回退和波动来评估自动提示优化，发现整体准确率可能掩盖大量变化，建议同时报告恢复、回归、churn 和未见数据上的表现。
提出 MoPrO-SD，一种面向立场检测的模块化提示优化框架，将提示分解为功能语义模块并利用 LLM 反馈逐步优化，同时发布新的人工标注疫苗立场检测基准 VaccineStance，在多个数据集和开源 LLM 上验证了有效性。
一项预注册研究考察了 GEPA、MIPROv2 和 TextGrad 等提示优化器在提升目标指标时是否会破坏未被评分的输出要求，发现主要回归源于输出截断导致的提取失败，增大输出预算即可消除该副作用。
该论文研究了对多智能体深度研究智能体的系统提示进行反思式提示进化（GEPA），采用主题不重叠、多随机种子的评估协议，发现分解式逐维度反馈是可迁移增益的关键因素，而压缩为单一标量奖励会导致负迁移。
论文提出 CoCoEvolve（CortexCodeEvolve），一种基于"元反思"的进化优化器，通过从完整搜索历史中检索证据并主动探测目标环境来改进生产级 AI 系统的优化，在 DABStep、data-eng-bench 等多个基准上超越现有基线。
A user praises Qwen 3.8 27B as an underrated, high-performing model, noting that smaller/local models can match bigger ones and that LLMs still struggle with 'pointing'. They plan to test prompt optimization for further gains.
Goldsmith, an agentic annotation pipeline presented in an arXiv paper, converts small expert-annotated gold sets into structured annotation definitions optimized via an executable loss-guided loop, outperforming direct rewriting, OPRO, APE, and PromptBreeder.
The paper introduces VERSE, a verified self-evolving optimizer that improves both the executor agent's harness and its own prompts, tools, and procedures using execution-based verification, achieving 42.3% and 37.7% accuracy on held-out SWE-rebench tasks versus 39.2% and 29.3% for the strongest baselines.
The paper introduces TextReg, a regularization framework for text-space prompt optimization that mitigates prompt distributional overfitting via dual-evidence gradient purification and semantic edit regularization, yielding up to +11.8% over TextGrad and +16.5% over REVOLVE on out-of-distribution reasoning benchmarks.
Researchers introduce a chess-puzzle benchmark (1,118 Lichess puzzles) for automatic prompt optimization of frozen LLMs, evaluating six APO algorithms on eight models with cheap, deterministic, renewable scoring at a total cost of around $800.
The paper introduces SHIFT, a framework that builds multi-agent harnesses per-query using a learned local LLM architect and value function combined with Monte Carlo tree search, achieving ~80% mean accuracy across 9,193 tasks and outperforming 17 baselines while reducing execution token costs by 32% in its cheaper mode.