High-priority alert based on source significance and keyword relevance.
Context
What happened: Researchers introduced the Continual Learning Exam (CLE), a benchmark built with 50+ domain experts to measure whether LLM agents improve through repeated professional tasks with feedback.
Key numbers: 120 tasks, 600 levels, 89 roles, 12 domains, 17 harness-model configurations, median gain of 5.3 percentage points, best system clears 47.7% of levels
Why it matters: Results indicate frontier LLM agents still have substantial room to improve in long-term experience accumulation and knowledge transfer, and the dataset can be used to post-train models for continual learning.
Why it matters
Results indicate frontier LLM agents still have substantial room to improve in long-term experience accumulation and knowledge transfer, and the dataset can be used to post-train models for continual learning.