LLMs struggle badly on tasks with more than one right answer
andpoul · x · 2026-10-30
Developer andpoul observes that LLMs are remarkably bad at tasks with more than one correct answer. He's not arguing for a different architecture, but suggests new training methods may be needed to address this limitation.
More from Models
- Reddit meme: labs will rush releases before Qwen 4 just to top benchmarks for a day — Porespellar · 2026-10-30
- Humansand launches &Tutor, a tutor model trained by teaching a human-like learner model — techNmak · 2026-10-30
- Gemini 4 tops public coding benchmarks but drops to bottom on private tasks — Hesamation · 2026-10-30
- Codex Seems to Burn Usage Until You Open the Chat, User Suspects — machpella · 2026-10-30
- Qwen4 Spotted in qwen-code PR, Hinting at a 120B Model — giveen · 2026-10-30
- Your AI agent would cost 260% more without prompt caching, Arena data shows — arena · 2026-10-30