AIThe Decoderabout 1 hour ago
New Benchmark Exposes AI Agent Decision Flaws
5 min read
Microsoft and City University of Hong Kong researchers created Taste-Bench, a new benchmark proving AI agents struggle with long-term decisions. The best model, GPT-5.6 Sol, averaged 59.7 percent accuracy, barely beating a coin flip. While training a student model on Qwen3.6-27B improved performance to 47.9 percent, the study highlights that models collapse when evidence appears only after extensive later work.
Level
Hype check
Tap to vote and see what everyone thinks.
