Continual Learning Exam
Evaluating how agents acquire, retain, and reuse experience across professional workflows. Each task runs over five connected levels, so what is learned early has to be used later.
CLE at a Glance
- 120 tasks
- 12 domains
- 23,981 input files
- 16,125 rubric criteria
120 tasks written by 50+ domain experts, covering 89 professional roles in 12 domains. Each task has five connected levels. Together they hold 23,981 input files and 16,125 rubric criteria.
Benchmark Design Meet the TeamFigure 1.What the benchmark is made of. Tasks by domain and area, with every task named on the outer ring. Input files by family. Rubric criteria by level. Hover to explore. Click to keep a selection.
12 professional domains
120 tasks across 12 domains, with 23,981 input files and 16,125 rubric criteria. Explore the Benchmark →
What Later Levels Reuse
Later levels can draw on what the agent picked up earlier: , , , , . In PJM Day-ahead Launch a load-forecasting data scientist works through five levels. Each arc below is something a later level uses from an earlier one, and some reach back several levels.
How Agents Perform
- 47.7% of levels completed by the best configuration
- 5.3 median points added by feedback and one revision
The best configuration completes 47.7% of levels. Diagnostic feedback and one revision add a median of 5.3 points.
Full LeaderboardFigure 3.Completion across all 600 levels for 17 agent configurations. Trial 1 is a single submission. Trial 2 adds diagnostic feedback and one revision.
The best configuration completes 47.7% of levels. Feedback and revision yield a median gain of 5.3 percentage points. View All Results →