概述
They read tokens, which are chunks of characters, so the individual letters inside a word are often never directly visible to the model. This one fact explains a whole family of odd mistakes with spelling, counting, reversing and rhyming, and it tells you how to work around them.
深入探讨
Before a language model sees any text, a tokenizer splits it into tokens. Most modern tokenizers use a method such as byte pair encoding, which learns common character sequences from a large body of text. Frequent words often become a single token, and rarer words are split into a few pieces. The model then receives each token as a number and turns it into a vector. It never receives the letters inside that token as separate inputs. So when you ask how many r's are in 'strawberry', the model is not scanning ten characters. It sees perhaps two or three chunks, and it has to recall, from patterns learned during training, which letters those chunks contain. That knowledge exists but it is indirect and fuzzy, like being asked how many times the letter e appears in a word you have only ever seen as a picture. The strawberry question became a widely shared example in 2024 precisely because the answer is obvious to people and surprisingly hard for chatbots. The same cause produces related errors: misspelling when asked to spell out loud, reversing words wrongly, failing letter-based constraints like acrostics or lipograms, and miscounting characters for length limits. Exact splits differ between tokenizers, so one model may get a given word right while another fails. A common misconception is that this shows models cannot reason or are simply careless. The limit is mostly in the input format, not in general reasoning. When a model is prompted to spell the word out first, each letter tends to become its own token, and counting becomes a far easier task. Newer reasoning models often do this spelling step on their own, which is one reason they get these questions right more often. For anything where exactness matters, running a short piece of code is still the most reliable method.
战略影响
更清晰的判决
它可以帮助您将清晰的技术声明与营销语言分开。
成本与预算
在花费金钱或时间之前，您可以提出更好的实施问题。
团队与工作流程
具有共同理解的团队可以做出更好的产品、政策和学习决策。
The Future of Why LLMs Struggle to Count Letters
Researchers are exploring models that work directly on bytes or characters, and hybrid designs that group bytes dynamically, which could make character-level tasks more natural. These approaches have tradeoffs, because longer input sequences cost more compute. Meanwhile, reasoning models that spell words out step by step and models that call code tools already reduce many of these errors in practice. It is reasonable to expect fewer embarrassing letter-counting failures over time, but for tasks that need exact character counts, using code rather than trusting a model's recall will likely remain good practice.
现实世界的实施
Asked how many times the letter r appears in 'strawberry', a chatbot answers two instead of three, because the word reaches the model as a few multi-letter chunks rather than ten separate letters.
A teacher asks a model to write a sentence where every word starts with the letter s, and it slips in words that do not, since it has to infer each word's first letter instead of seeing it.
A developer asks for the word 'encyclopedia' spelled backwards and gets a version with swapped or missing letters; asking the model to first list the letters one per line, then reverse the list, fixes it.
A puzzle app needs exact letter counts for a word game, so it computes them with a single line of ordinary code and uses the model only to write the hints.
风险与防护栏
不同的团队可能会以不同的方式使用同一术语，因此请尽早定义范围。
基准测试可能看起来很强大，但实际性能却参差不齐。
忽视数据质量和评估计划通常会产生脆弱的结果。
实施路线图
从您需要的结果的简单语言定义开始。
在测试之前选择一种成功指标和一种失败条件。
使用代表性数据运行小型试点，而不是完善的演示集。
Document where Why LLMs Struggle to Count Letters helps and where simpler methods are better.
不断探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Why LLMs Struggle to Count Letters quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常见问题
What is Why LLMs Struggle to Count Letters?
Large language models struggle to count letters because they do not read text one character at a time. They read tokens, which are chunks of characters, so the individual letters inside a word are often never directly visible to the model. This one fact explains a whole family of odd mistakes with spelling, counting, reversing and rhyming, and it tells you how to work around them.
What is the main reason language models miscount letters in a word?
Models receive tokens, which often contain several characters. The letters inside a token are not separate inputs, so the model must recall them indirectly.
What does a tokenizer produce from a piece of text?
The tokenizer splits text into pieces from a fixed vocabulary and maps each to an integer ID, which is then turned into a vector.
Which prompt change most helps a model count letters correctly?
Writing the word with separators like s-t-r-a-w usually puts each letter in its own token, which makes counting a much easier task.
Byte pair encoding builds its vocabulary by doing what?
BPE starts from bytes or characters and keeps merging the most frequent adjacent pair, so common strings become single tokens.
Why might 'Strawberry' and ' strawberry' be handled differently by a model?
Tokenizers often include leading spaces and case in tokens, so small surface changes can produce different token splits.
继续学习
相关指南
为此主题精选的更多指南