GEPA
Timeline (last 90 days)
GEPA-Curator adds two modules to the original prompt optimizer GEPA: task curation and prior success retrieval.
Compared with GEPA, GEPA-Curator achieves performance improvements of up to 20.7%.
GEPA-Curator matches or exceeds both baselines, GEPA and ACE, in average performance.
When configured appropriately, merging can improve the performance of both gpt-4.1-mini and Qwen3-8B simultaneously.
In the GEPA paper, the merge operator can improve the average performance of gpt-4.1-mini but reduces the effectiveness of Qwen3-8B, an asymmetry that previously lacked explanation.
A mechanism-level diagnostic study scanned nine configurations of merge timing and merge operators across five datasets and two models (gpt-4.1-mini and Qwen3-8B).
The instability of GEPA merging stems more from configuration sensitivity than from inherent flaws in the merge operators themselves.
Prompt-level static features (such as length, perplexity, semantic complexity, etc.) cannot predict GEPA's merge acceptance decisions, and static-feature gating is explicitly ruled out.
A behavior-based probe, by identifying high-risk tasks and abstaining, can avoid large-scale merge failures across multiple random-seed runs.
In practice, if performance degradation caused by merging is encountered, the first direction to investigate should be the configuration combination (merge timing and operator selection), rather than abandoning the use of the merge mechanism.
10 more timeline events
All Facts (20)
GEPA将遗传算法的变异操作与Agent执行轨迹深度绑定，根据轨迹中的具体失败点提出针对性修改而非随机变异
90%UnverifiedGEPA引入帕雷托最优概念，同时考量准确率、Token消耗、延迟等多个指标进行多目标平衡优化
90%UnverifiedGEPA (Genetic Pareto Prompt Evolution) is a method proposed in an ICLR 2026 Oral paper.
60%UnverifiedCompared with GEPA, GEPA-Curator achieves performance improvements of up to 20.7%.
50%UnverifiedGEPA-Curator adds two modules to the original prompt optimizer GEPA: task curation and prior success retrieval.
50%UnverifiedGEPA-Curator matches or exceeds both baselines, GEPA and ACE, in average performance.
50%UnverifiedA behavior-based probe, by identifying high-risk tasks and abstaining, can avoid large-scale merge failures across multiple random-seed runs.
50%UnverifiedAny attempt to pre-screen merge operations via simple rules or lightweight classifiers (based on static prompt features) is unlikely to work reliably.
50%UnverifiedThe instability of GEPA merging stems more from configuration sensitivity than from inherent flaws in the merge operators themselves.
50%UnverifiedPrompt-level static features (such as length, perplexity, semantic complexity, etc.) cannot predict GEPA's merge acceptance decisions, and static-feature gating is explicitly ruled out.
50%UnverifiedIn practice, if performance degradation caused by merging is encountered, the first direction to investigate should be the configuration combination (merge timing and operator selection), rather than abandoning the use of the merge mechanism.
50%UnverifiedIn the GEPA paper, the merge operator can improve the average performance of gpt-4.1-mini but reduces the effectiveness of Qwen3-8B, an asymmetry that previously lacked explanation.
50%UnverifiedThe effectiveness of merge operators depends heavily on the complementarity of the prompts being merged and on the model's ability to follow the synthesized instructions.
50%UnverifiedWhen configured appropriately, merging can improve the performance of both gpt-4.1-mini and Qwen3-8B simultaneously.
50%UnverifiedGEPA (Genetic-Pareto) is an optimization strategy in DSPy that combines genetic-style search with Pareto-frontier ideas, using execution traces, scoring metrics, and natural language feedback to iteratively rewrite prompts.
50%UnverifiedGEPA differs from fine-tuning in that it does not update model weights and only optimizes configuration at the prompt level.
50%UnverifiedThe harness loop includes Reefine, SkillClaw, GEPA, and Meta-Harness, which propose modifications to the harness's own files and compare scores against the current version, with the default selection strategy being to retain changes only upon a "net win".
50%UnverifiedThe study compared two prompting approaches: expert handcrafted prompts and prompts optimized by the Genetic-Pareto Algorithm (GEPA).
50%UnverifiedGEPA is an automatic prompt optimization method that combines evolutionary computation with multi-objective optimization, selecting optimal prompts through Pareto frontier criteria.
50%UnverifiedUnder the test configuration used, GEPA-optimized prompts did not systematically outperform expert handcrafted prompts.
50%