<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni" -->

---
title: Why Thinking More Can Make an LLM Dumb | daily.dev
description: Extra reasoning time can make a language model's answer worse, not better, because each token it writes shapes what it writes next. Reasoning models like...
canonical: https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Why Thinking More Can Make an LLM Dumb | daily.dev
og:description: Extra reasoning time can make a language model's answer worse, not better, because each token it writes shapes what it writes next. Reasoning models like...
og:url: https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni
og:image: https://api.daily.dev/og/posts/ZJ3uSl1ni.png
og:image:alt: Why Thinking More Can Make an LLM Dumb
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Why Thinking More Can Make an LLM Dumb

**[ByteByteGo](https://daily.dev/sources/bytebytego)** · 14 min read · 25 upvotes · 5 comments

## Summary

Extra reasoning time can make a language model's answer worse, not better, because each token it writes shapes what it writes next. Reasoning models like OpenAI's o1 and DeepSeek's R1 generate intermediate 'thinking tokens' before a final answer, trained with methods like DeepSeek's Group Relative Policy Optimization. Researchers found one model asked to compute '2 + 3' got 5 correctly, then kept going until it produced thirteen redundant solutions to the same question. In a train-timetable example, a model correctly reads a 9:00 departure time but then overthinks a mentioned delay and answers 9:10, drifting from the actual question asked. Extra steps help on hard problems by catching mistakes, but on easy ones they can waste time or introduce wrong assumptions.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://blog.bytebytego.com/p/why-thinking-more-can-make-an-llm>

## Questions this post answers

### why does a reasoning model sometimes give a wrong answer even though it found the right answer first

A reasoning model can overwrite a correct early answer because later tokens build on everything written before, including unnecessary extra reasoning. In one documented case, a model asked to calculate '2 + 3' arrived at 5 immediately but kept going, producing thirteen total solutions to the same question, wasting time even when correct. On questions with an incidental detail, like a train timetable mentioning a usual delay, the extra reasoning can pull the model away from the literal question into an unintended inference, such as predicting 9:10 instead of reporting the scheduled 9:00.

_Anyone debugging flaky reasoning-model output can track findings like this on daily.dev._

### what training method did deepseek use to teach r1 to reason step by step

DeepSeek used Group Relative Policy Optimization (GRPO) to train R1-Zero, having the model generate a group of attempts per question, score each one, and adjust the model to favor attempts that scored better than their peers in the same group, rather than training a separate critic model to estimate scores. DeepSeek-R1 itself added supervised fine-tuning on examples after R1-Zero showed issues like switching languages mid-response.

_Engineers evaluating reasoning-model training approaches can follow DeepSeek updates on daily.dev._

## Community discussion

Top comments from developers on daily.dev.

**@arrrdel** · 1 upvotes

> I literally published a post today on why thinking longer can help models: [https://medium.com/@chindahel1/why-thinking-longer-helps-test-time-compute-and-the-new-scaling-axis-8971c85fb0a9](https://medium.com/@chindahel1/why-thinking-longer-helps-test-time-compute-and-the-new-scaling-axis-8971c85fb0a9)
>
> But I agree with you that extra reasoning steps on relatively easy tasks can be unnecessary and even cause explosion in token consumption.

**@diyweb** · 1 upvotes

> The timetable example is the failure mode I would watch for. The model already had the 9:00 departure, then treated a mentioned delay as something it had to use, and answered 9:10, which is a different question. Extra thinking tokens help when they catch a mistake; they hurt when every earlier fact is treated as a constraint that must appear in the answer. For a short, closed question such as 2 + 3, I would cap the thinking or stop at the first consistent result, because the model in the article kept going until it had produced thirteen more solutions to a problem it had already solved. A...

**@kartiknvj** · 0 upvotes

> I want to test the timetable example on my own agent, where the model has 9:00 right and then talks itself into 9:10 because of the delay note. The version I worry about is a correct value from a tool followed by an assumption the model adds on its own.

**@sayanmondal** · 0 upvotes

> It Shows how extra reasoning can derail a simple answer: the model overused the delay detail and answered a different question. For straightforward problems, it should stop once it finds a consistent answer—not keep inventing more work.

**@mikhailmakeev** · 0 upvotes

> The timetable example is the one I'd worry about in production, because the wrong answer reads perfectly fine. We hit a cousin of it in a news tagger. The article said Walmart shares "fell up to 9%", and the model wrote that the news was "driving a 9% share rise". Our first guess was the stray "up", so we rewrote the summary to a plain "fell 9%" and it still wrote rise. Raising reasoning effort from low to medium didn't fix it either. It used 2.6 times the output tokens and only softened the sentiment label to neutral. Our fallback model got that row right every time. So the extra thinking...

## Similar posts on daily.dev

- [Don’t be fooled—LLMs don’t reason](https://daily.dev/posts/don-t-be-fooled-llms-don-t-reason-4waw8d0j4) · MIT Technology Review · 0 upvotes · 1 comments
- [What objective is reasoning solving?](https://daily.dev/posts/what-objective-is-reasoning-solving--9zzqapr18) · Hacker News · 14 upvotes · 1 comments

---

Tags: [#llm](https://daily.dev/tags/llm), [#prompt-engineering](https://daily.dev/tags/prompt-engineering), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning)

[View this post on daily.dev](https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Why Thinking More Can Make an LLM Dumb","url":"https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni"},"datePublished":"2026-11-17T16:36:05.449Z","dateModified":"2026-11-17T16:48:21.872Z","description":"Extra reasoning time can make a language model's answer worse, not better, because each token it writes shapes what it writes next. Reasoning models like...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/475a6de6c3c160d31eb05623abc86ef5?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/475a6de6c3c160d31eb05623abc86ef5?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"ByteByteGo","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"ByteByteGo","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/35be29234ee14d01a9cd049c52e12753","url":"https://daily.dev/sources/bytebytego"},"commentCount":5,"discussionUrl":"https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":25},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":5}],"keywords":"llm,prompt-engineering,reinforcement-learning","timeRequired":"PT14M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"ByteByteGo","item":"https://daily.dev/sources/bytebytego"},{"@type":"ListItem","position":3,"name":"Why Thinking More Can Make an LLM Dumb"}]}
{"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni","comment":[{"@type":"Comment","text":"I literally published a post today on why thinking longer can help models: https://medium.com/@chindahel1/why-thinking-longer-helps-test-time-compute-and-the-new-scaling-axis-8971c85fb0a9\nBut I agree with you that extra reasoning steps on relatively easy tasks can be unnecessary and even cause explosion in token consumption.","datePublished":"2026-11-17T17:43:05.999Z","url":"https://daily.dev/posts/ZJ3uSl1ni#c-pSPuW1UhI","author":{"@type":"Person","name":"arrdel","url":"https://daily.dev/arrrdel","image":"https://media.daily.dev/image/upload/s--lrD-0dez--/f_auto/v1786754219/avatars/avatar_z4vIWrBpOmlmkyMcosnSh?_a=BAMAMicg0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}},{"@type":"Comment","text":"The timetable example is the failure mode I would watch for. The model already had the 9:00 departure, then treated a mentioned delay as something it had to use, and answered 9:10, which is a different question. Extra thinking tokens help when they catch a mistake; they hurt when every earlier fact is treated as a constraint that must appear in the answer. For a short, closed question such as 2 + 3, I would cap the thinking or stop at the first consistent result, because the model in the article kept going until it had produced thirteen more solutions to a problem it had already solved. A time budget is not enough if that time can be spent inventing work.","datePublished":"2026-11-17T18:28:37.650Z","url":"https://daily.dev/posts/ZJ3uSl1ni#c-i6WfAGvEP","author":{"@type":"Person","name":"DiyWeb","url":"https://daily.dev/diyweb","image":"https://media.daily.dev/image/upload/s--dgS6jDeN--/f_auto/v1790879420/avatars/avatar_774zDwMaRTNS1fv5OKStX?_a=BAMAMicg0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}},{"@type":"Comment","text":"I want to test the timetable example on my own agent, where the model has 9:00 right and then talks itself into 9:10 because of the delay note. The version I worry about is a correct value from a tool followed by an assumption the model adds on its own.","datePublished":"2026-11-17T19:10:03.776Z","url":"https://daily.dev/posts/ZJ3uSl1ni#c-yj9BcV4al","author":{"@type":"Person","name":"kartik-nvjk","url":"https://daily.dev/kartiknvj","image":"https://media.daily.dev/image/upload/s--3gGgsVCw--/f_auto/v1781456774/avatars/avatar_TvTVeiMdkRCqWUDullFmy?_a=BAMAMiWQ0"}},{"@type":"Comment","text":"It Shows how extra reasoning can derail a simple answer: the model overused the delay detail and answered a different question. For straightforward problems, it should stop once it finds a consistent answer—not keep inventing more work.","datePublished":"2026-11-18T19:44:45.526Z","url":"https://daily.dev/posts/ZJ3uSl1ni#c-PpVDXMqYZ","author":{"@type":"Person","name":"Sayan Mondal","url":"https://daily.dev/sayanmondal","image":"https://media.daily.dev/image/upload/s--cW6TLpb9--/f_auto/v1731954927/avatars/avatar_31X8v4aef86WyygmnVrjC"}},{"@type":"Comment","text":"The timetable example is the one I’d worry about in production, because the wrong answer reads perfectly fine. We hit a cousin of it in a news tagger. The article said Walmart shares “fell up to 9%”, and the model wrote that the news was “driving a 9% share rise”. Our first guess was the stray “up”, so we rewrote the summary to a plain “fell 9%” and it still wrote rise. Raising reasoning effort from low to medium didn’t fix it either. It used 2.6 times the output tokens and only softened the sentiment label to neutral. Our fallback model got that row right every time. So the extra thinking didn’t cause it and didn’t cure it. Has anyone seen effort change which fact gets invented, rather than how often?","datePublished":"2026-11-18T06:30:26.548Z","url":"https://daily.dev/posts/ZJ3uSl1ni#c-RFjPbFLSU","author":{"@type":"Person","name":"Mikhail Makeev","url":"https://daily.dev/mikhailmakeev","image":"https://lh3.googleusercontent.com/a/ACg8ocIjcvDU1gmJ9-KJDeTYqY2rMdl5QxcWYtLTGSXrj8ubF4nwar_w=s96-c"}}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/why-thinking-more-can-make-an-llm-dumb-zj3usl1ni#faq","mainEntity":[{"@type":"Question","name":"why does a reasoning model sometimes give a wrong answer even though it found the right answer first","acceptedAnswer":{"@type":"Answer","text":"A reasoning model can overwrite a correct early answer because later tokens build on everything written before, including unnecessary extra reasoning. In one documented case, a model asked to calculate '2 + 3' arrived at 5 immediately but kept going, producing thirteen total solutions to the same question, wasting time even when correct. On questions with an incidental detail, like a train timetable mentioning a usual delay, the extra reasoning can pull the model away from the literal question into an unintended inference, such as predicting 9:10 instead of reporting the scheduled 9:00. Anyone debugging flaky reasoning-model output can track findings like this on daily.dev."}},{"@type":"Question","name":"what training method did deepseek use to teach r1 to reason step by step","acceptedAnswer":{"@type":"Answer","text":"DeepSeek used Group Relative Policy Optimization (GRPO) to train R1-Zero, having the model generate a group of attempts per question, score each one, and adjust the model to favor attempts that scored better than their peers in the same group, rather than training a separate critic model to estimate scores. DeepSeek-R1 itself added supervised fine-tuning on examples after R1-Zero showed issues like switching languages mid-response. Engineers evaluating reasoning-model training approaches can follow DeepSeek updates on daily.dev."}}]}
```

