@godofprompt: This MIT paper explains why te...
@godofprompt
31 views
Dec 24, 2025
1
This MIT paper explains why telling models to “think harder” keeps failing in practice.
The core argument is simple and uncomfortable: reasoning doesn’t break because models lack knowledge. It breaks because reasoning itself becomes unstable when it goes on too long.
The researchers show that as models reason step by step, performance initially improves. Then it plateaus. Then it degrades. Errors compound. Assumptions quietly drift. The model stays confident while moving further away from the correct answer.
What’s striking is that the failures aren’t random. The model often applies the right rules early on, then violates those same rules later without noticing. Each step depends on the previous one, so a small mistake quietly poisons everything downstream.
The paper draws a sharp distinction between more reasoning and controlled reasoning.
More reasoning means longer chains of thought.
Controlled reasoning means knowing when to stop, when to verify, and when to constrain the process so it doesn’t eat itself.
Most systems today only do the first.
A few insights stood out.
Reasoning is not a monotonic resource. Asking a model to think longer does not reliably improve accuracy. Past a threshold, extra reasoning actively hurts performance.
Confidence is a misleading signal. Models often become more confident as their answers get worse, because nothing in the reasoning loop forces self-correction.
Internal consistency is fragile. Once an assumption slips, the model rarely revisits it. The reasoning keeps going, building logically on a flawed foundation.
The takeaway is blunt.
Better reasoning systems won’t come from longer chains of thought or bigger prompts. They will come from constraints, verification, and mechanisms that prevent reasoning from spiraling once it starts to drift.
Right now, most “reasoning-heavy” systems don’t fail because they think too little.
They fail because they think too much, without control.
The core argument is simple and uncomfortable: reasoning doesn’t break because models lack knowledge. It breaks because reasoning itself becomes unstable when it goes on too long.
The researchers show that as models reason step by step, performance initially improves. Then it plateaus. Then it degrades. Errors compound. Assumptions quietly drift. The model stays confident while moving further away from the correct answer.
What’s striking is that the failures aren’t random. The model often applies the right rules early on, then violates those same rules later without noticing. Each step depends on the previous one, so a small mistake quietly poisons everything downstream.
The paper draws a sharp distinction between more reasoning and controlled reasoning.
More reasoning means longer chains of thought.
Controlled reasoning means knowing when to stop, when to verify, and when to constrain the process so it doesn’t eat itself.
Most systems today only do the first.
A few insights stood out.
Reasoning is not a monotonic resource. Asking a model to think longer does not reliably improve accuracy. Past a threshold, extra reasoning actively hurts performance.
Confidence is a misleading signal. Models often become more confident as their answers get worse, because nothing in the reasoning loop forces self-correction.
Internal consistency is fragile. Once an assumption slips, the model rarely revisits it. The reasoning keeps going, building logically on a flawed foundation.
The takeaway is blunt.
Better reasoning systems won’t come from longer chains of thought or bigger prompts. They will come from constraints, verification, and mechanisms that prevent reasoning from spiraling once it starts to drift.
Right now, most “reasoning-heavy” systems don’t fail because they think too little.
They fail because they think too much, without control.
2
Read the full paper here: arxiv.org/pdf/2512.17901
3
10x your prompting skills with my prompt engineering guide
→ Mini-course
→ Free resources
→ Tips & tricks
Grab it while it's free ↓
godofprompt.ai/prompt-enginee…
→ Mini-course
→ Free resources
→ Tips & tricks
Grab it while it's free ↓
godofprompt.ai/prompt-enginee…