LLM Memory Consolidation Degrades Performance on ARC-AGI Tasks
A study on arXiv (2605.12978) reveals that continuous memory consolidation by LLMs can degrade performance. Researchers found that while initial memory updates improve utility, further consolidation leads to degradation, sometimes below a no-memory baseline. GPT-5.4 failed on 54% of ARC-AGI problems it had previously solved without memory, tracing the regression to the consolidation step.
Key facts
- arXiv paper 2605.12978 examines LLM memory consolidation
- Consolidation first improves then degrades memory utility
- GPT-5.4 fails 54% of previously solved ARC-AGI problems after consolidation
- Regression traced to consolidation step, not underlying experience
Entities
Institutions
- arXiv