arXiv cs.CL· Xun Sun, Thanh Thi Nguyen, Thuy-Trang Vu·· 13 小时前AI 评分34
HERD：面向持续学习的 harness 演化递归蒸馏框架
HERD: Harness-Evolving Recursive Distillation for Continual Learning
AI 导读
HERD 是一个混合持续学习框架，在 harness 侧持续精炼技能并保留技能未能覆盖的高残差样本，在参数侧用双锚点做漂移约束蒸馏，从而把经验内化为累积能力。在 TRACE+17 的 25 阶段任务流上，固定内存与算力预算下，HERD 的新任务获取能力追平最强参数化基线，同时遗忘最少、通用能力保持最强。
正文 · 原文
Abstract:Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. Parametric methods suffer from catastrophic forgetting and loss of plasticity. Non-parametric methods, such as prompts, skills, and harnesses, leverage in-context learning but are limited by context length and lossy plain-text compression. To effectively combine their complementary strengths, we propose HERD (Harness-Evolving Recursive Distillation), a hybrid continual learning framework that learns efficiently in the harness and internalizes experience into cumulative capabilities via distillation. On the harness side, HERD continually refines its skills to capture reusable patterns, while retaining high-residual examples that the skills fail to capture in a bounded memory. On the parameter side, HERD performs drift-constrained distillation with dual anchors, substantially reducing forgetting over long recursive runs. On the 25-stage task stream of TRACE+17, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task acquisition while exhibiting the least forgetting and the strongest retention of general capabilities.
| Comments: | 22 pages, 9 figures |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.31401 [cs.CL] |
| (or arXiv:2610.31401v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.31401 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xun Sun [view email]
[v1]
Thu, 22 Oct 2026 15:03:20 UTC (628 KB)
来源：arXiv cs.CL · arxiv.org