$ cat wiki/papers/2026/2611.18856-self-improvement-memory.md
Self-Improvement Memory
TL;DR
Self-Improvement Memory (SIM) preserves the reasons and outcomes behind harness edits when optimization continues across compatible benchmarks or policy models. It is memory for the optimizer, not the task-solving agent. (source)
Authors & Org
Xin Yu, Fei Xia, Jiaqi Pan, Youbao Tang, Chengzhi Liu, Adams Yu, Sicong Jiang, Yanfei Chen, Lingzhou Xue and Ting Yu. Affiliations are not established by the reviewed abstract record. (source)
Method
The authors study Continual Recursive Self-Improvement (CRSI). SIM groups changes by recurring harness problem and records rationale, predicted and observed outcomes, and evaluation context. Subsequent attempts extend this history, making prior design decisions available when the benchmark or policy changes. (source)
Results
Across 4 domains, authors report higher final scores than harness-only and root-cause-only continuation in cross-dataset and pairwise cross-model comparisons. Across 10 independent coding source runs, paired SIM continuations improve final scores by 2.36 points on average over harness-only continuation (p=0.003). This is a score difference, not an absolute success rate. (source)
Significance
For Eval Harness Configuration, the reusable artifact includes the evidence behind edits. This connects to Agents (LLM Agents) through optimizer state, without establishing universal transfer or autonomous weight improvement. (source)
Open Questions
The review covers the abstract and metadata only. Full evaluation protocols, compatibility limits and implementation costs have not been assessed here; the reported result has not been independently reproduced. (source)
Cite
- arXiv:2611.18856v1, submitted 2026-11-13; captured 2026-11-19. (source)