Counterintuitive: Higher Reasoning Effort Causes Models to Lose Previously Solved Tasks
zainhas · x · 2026-08-11
Developer @zainhas shared a counterintuitive finding regarding LLM reasoning: when turning up a model's reasoningeffort, it does not solve more net new problems.
Instead, it trades some previously solved tasks for other tasks. Every single effort step, across all tested model families, loses tasks that the cheaper setting successfully solved.
More from Models
- Elon Musk Commits to Open-Sourcing Prior xAI Grok Models with Weights — MaziyarPanahi · 2026-08-11
- Anthropic's Guardrails Reportedly Prevent Models from Even Hinting They Are Muzzled — teortaxesTex · 2026-08-11
- Why Have 8B-12B Model Releases Stalled? Local Hardware Limits May Be the Cause — _maverick98 · 2026-08-11
- DeepSeek Reportedly Poised to Launch Agent, Registers 'Harness' Account — op7418 · 2026-08-11
- Grok Build TUI Can Integrate Third-Party Models Like DeepSeek — teortaxesTex · 2026-08-11
- AI This Week: Claude Tackles Riemann Hypothesis, Meta Releases 30B Muse Glimmer — WorldofAI · 2026-08-11