Paper proves more reasoning tokens unlock harder problems, sparking debate

A NeurIPS 2026 paper gives the first tight hierarchy theorem proving more reasoning tokens let Transformers strictly solve harder problem classes. The discussion drew pushback from gerardsans, who argued agentic workflows should be orchestrated in code, not natural language prompts.

2026-11-12 ~ 2026-11-12 · 3 related posts

• First CoT hierarchy theorem: more reasoning steps provably unlock harder problems — gurtej__gill_ · 2026-11-12
• New complexity paper claims more reasoning tokens strictly expand what transformers can solve; critics push back — gerardsans · 2026-11-12
• Stop Defining Agent Workflows in Plain Text: Why Deterministic Code Beats Fragile Prompts — gerardsans · 2026-11-12