RESEARCH · RESEARCH · #1318
Component-Aware Feedback for Self-Evolving Programs (arXiv:2609.38639v1)
The paper introduces "component-aware feedback", which logs the component-level edits and their metric differences in an attribution memory storing both local (vs parent) and global (vs seed) reference frames so later mutations can read explicit attribution. On 12 Bright reranking datasets the authors report reaching a strong baseline's final quality after a median of one third of the search budget, finishing 7.2% higher in held-out nDCG@10, and under a cost-aware objective finding pipelines that are on average more accurate while using 11% fewer tokens per query.
KEY POINTS
- The paper introduces "component-aware feedback", which logs the component-level edits and their metric differences in an attribution memory storing both local (vs parent) and global (vs seed) reference frames so later mutations can read explicit attribution.
- On 12 Bright reranking datasets the authors report reaching a strong baseline's final quality after a median of one third of the search budget, finishing 7.2% higher in held-out nDCG@10, and under a cost-aware objective finding pipelines that are on average more accurate while using 11% fewer tokens per query.
- Provides a practical way to attribute the effects of specific component edits during LLM-guided program evolution, speeding and stabilizing multi-component pipeline search and reducing serving cost.
WHY IT MATTERS
Provides a practical way to attribute the effects of specific component edits during LLM-guided program evolution, speeding and stabilizing multi-component pipeline search and reducing serving cost.