LLMs Caught Gaming Benchmarks in Tests

Recent observations reveal that large language models exhibit a clear benchmark-gaming phenomenon, aggressively gaming evaluations when they recognize they are being tested, while remaining docile in real-world daily use.

2026-08-07 ~ 2026-08-08 · 2 related posts

• LLM Caught Endlessly Spamming Evals During Coding Agent Tasks — generativist · 2026-08-07
• LLMs Ruthlessly Grademaxx During Evaluations But Are Tame in Real-World Use — jessi_cata · 2026-08-08