测试时持续学习中的绑定特异性重学习节省
原标题：Binding-Specific Relearning Savings for Test-Time Continual Learning
AI 摘要
该论文研究参数更新的测试时持续学习（TTCL）智能体，提出用「重学习节省」（relearning savings）评估先前经验对后续更新的影响，而非仅看当前输出准确率。作者在 Qwen2.5-1.5B 上设置 Learned、Control、Wrong 三种匹配历史，经相同的无回放持续微调后，发现 Learned 在部分遗忘检查点重新习得目标比 Control 更快，Wrong 相对 Control 无正向优势。研究表明当前行为准确率可能低估已获得经验的持久影响，并开源了代码。
正文节选
Towards Test-Time Continual Learning Agents Binding-Specific Relearning Savings for Test-Time Continual Learning Abstract When a fact becomes harder to elicit from a language model, has the model lost it, or has it merely become less accessible? Parameter-updating test-time continual learning agents are usually evaluated by current output accuracy; we ask whether that measurement captures how prior experience shapes later updates. We test whether learning a specific factual binding produces a la