When Does Continual Learning Require Learning

arxiv:2607.07847

Original source · Paper JSON

Post-training · 2026-07-08

Cites (11) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• When Does Continual Learning Require Learning
• Abstract
• 1 Introduction
• 2 Related Works
• 3 Method
• 3.1 A Protocol for Continual Learning
• 3.2 Update Methods
• 4 Experiments
• 4.1 Domain Transfer
• 4.2 Catastrophic Memorizing: Update Truly New Knowledge
• 4.3 Noisy Temporal Drift
• 4.4 Agentic Task
• 5 Discussion
• 5.1 Conclusion
• 5.2 Limitations and Challenges
• Acknowledgement
• Appendix A Additional Experimental Details
• A.1 Compute and Hyperparameters
• A.2 Task Settings
• A.2.1 Domain shift
• In-place TTT configuration.
• A.2.2 TempWiki Benchmark Details
• A.2.3 TempWiki-Easy: a filtered, continuously-scored variant
• A.2.4 Temporal Drift
• A.2.5 Agentic.
• Update-mechanism hyperparameters.
• Suite construction.
• Verification.
• Agent harness.
• Zero-shot models.
• Appendix B Broader Impact
• Instructions for reporting errors