UCL's CLE Benchmark Finds Best Frontier LLM Agent Clears Only 47.7% of Levels
UniversityCollegeLondon · hf · 2026-10-27
UCL introduces the Continual Learning Exam (CLE), a large-scale benchmark built by 50+ domain experts to measure experience-driven improvement in LLM agents: 120 human-authored tasks × 5-level ladder = 600 levels across 89 professional roles and 12 domains. Across 17 harness–model configurations, feedback and revision yield only a 5.3-point median gain; the best system clears just 47.7% of levels and finishes all five levels in 23.3% of tasks, exposing how limited sustained learning from experience remains in frontier LLMs.
More from coding & agent
- If Your CLAUDE.md Is Longer Than Your README, Your Agent Gets Harder to Steer — Arindam_1729 · 2026-10-27
- When does checking an AI agent's work cost more than doing it yourself? — Live_Hedgehog8785 · 2026-10-27
- FrontierCS 2 launches on Kaggle with DeepMind: 48 tasks ask agents to beat published papers — lmoroney · 2026-10-27
- SkillHarm benchmark shows agent skill-poisoning attacks succeed up to 86.3% of the time — hhsun1 · 2026-10-27
- Dev builds Dreamcast game in 3 days with OpenAI's Astra, chasing locked 60fps — AIandDesign · 2026-10-27
- Dev builds Unreal prototype entirely via Claude Code CLI, never opening the editor — majidmanzarpour · 2026-10-27