安德森角
AI-Gened Text为什么比你想的更难理解
在 Google 上的首选来源中添加 Unite.AI
AI模型可能越来越难以理解，因为其培训越来越能有效解决问题，而不是向人类明确解释问题。
以上未经编辑的句子是ChatGPT根据一份科学论文的ELI5单一判决摘要的请求撰写的；它至少具有新研究报告* 中解释的几种语言管理特征。 新工作的结论是，这个问题影响到所有前沿模型，以及近前沿模型，以及一些也经过测试的开放权重模型。
例如，" 他们的培训越来越能有效解决问题 " 部分的早期部分就说明了文件的立场，即AI模式有时会给人类行动带来非活的东西(因为 " 培训 " 在上述例子中被描述为积极奖励的东西)。
这可能会妨碍快速解析该短语，因为我们必须在内部承认归属是一种说明方法，而不是从字面上理解(这要简单得多)。
后一节“不断提高高效解决问题的回报”也挤压了几个想法，迫使读者研究什么是回报、什么是解决、什么是效率提高。
该文件指出：
`[读者]在解析时使用动画作为提示，如果没有额外的上下文，不匹配可能会引起与重新分析相关的P600响应††.
`然而，改写这些句子以避免无生命主体可能要求延长句子：例如，“bites”一词代表“一个问题”。
`我们推测，在训练期间，模型可能已经趋同到将这种构思作为高效(尽管不是标准)的表达方式，否则会占用更多的象征。
说什么？
这是论文的核心 – 由于ChatGPT,Claude和Gemini等现代AI模型被优化用于多种类型的问题解决，而不是专门用于通信，它们倾向于以非常压缩的有效载荷的形式交付语言，需要一些解码。
这让我感到惊讶，因为我不断试图说服ChatGPT及其同龄人减少动词，压缩更多 — — 这个问题在文献中得到承认和解决。
相反，论文提出动词是一种拉链炸弹，即使在很短的句子中也难以理解。 因此，将超长的默认答复长度的双重问题与纸张对过度压缩的看法结合起来，我们可以想象，前沿AI的默认答复风格的特征模式是相当不同的(而且，大概，这种相当尖端的风格在探测器中被部分地用作AI生成文本的决定因素，尽管论文没有涉及这一点)。
作者们还检查了堆叠的名词复合物，例如有问题的短语Squad帮助狗咬：
‘在英语中，名词以形容方式自由修改其他名词：在短语中，咖啡店修改商店，尽管通常作为名词出现。 这个特性可以将多个名词堆叠成丰富的复合词，表示高度具体的含义。
但这种堆积会增加括号模糊
因此，如果LLMs是通过压缩来作动词，它们是否通过产生过多的压缩文本来响应一个查询而成为双动词？也就是说，不仅在回答您的询问时需要读取大量文本，而且文本本身可能具有许多需要解码的"隐含"含义。
一篇2023年的论文认为，LLMs可能因为AI法官在训练中被使用而成为不必要的动词，因为这些往往会奖励更长的答案 — — 即使人类在研究测试中更喜欢较短的答案。
新的作品属性的作者(见前引) 边疆模型的一些压缩倾向，为了保存信物而向建筑的处置，这是现在一个慷慨的想法。
让我们看看作者为研究进行的方法、数据和测试。
数据和方法
不寻常的单词组合(通过n-gram测量，或词序)因其不可预测性而被调查，罕见的组合与阅读难度较大有关。
依赖长度(语法上连接的单词之间的距离)也被测量，因为较长的距离对工作记忆提出了更大的要求——特别是当一个主题及其动词被几个介入词分隔时。
审查了对象相对条款(如 " 承包商建造的房屋 " 等建筑)，特别是在省略了备选词语(如 " 承包商建造的房屋 " 中的 " 该 " )的情况下。
上面提到的Stacked名词复合体(三个或三个以上名词结合在一起，如在 " 恐怖电影音轨生成器 " 中)也受到了调查，因为其组成词之间的关系可能模糊不清。
副词动词(作为动词使用的名词，如 " 叉子 " 而不是 " 把名词放在叉子上 " )，与与通常与生物相关的动词对齐的无生命主体(如前所述)一起被检查。 这两种构思都允许将意义压缩成较少的词，但可能需要读者的额外解释。
参赛者
研究人员收集了来自14个模型的AI生成的对话；Kimi K2.5;GLM 5;Claude Sonnet 5,Opus 5,Opus 5.5和Fable 5;GPT-5.6 Luna, Terra and Sol;Olmo Base，教程和思考；以及Llama Base和教程。
测试包括一般对话的10个起始提示和编码的5个提示，每个任务都需要一个模型来克隆一个真正的公共GitHub寄存器，找到bug，并将其固定在一个沙盒中。
对话限于12次交流，而编码会议限于11次，然后是固定的闭幕电文。 在单独的测试中，要求每个模型根据PESUADE 2.0数据集的任务写出七篇散文。
为了比较起见，从尽可能与AI输出类型相匹配的源头抽取了人文文字：OpenSubtitles(与开源子平台无关的电影和电视字幕集)用于一般对话；StackExchange(问答论坛在线网络)用于编码讨论；以及上述PESUADE2.0数据集的学生论文，用于编写任务。 维基百科也被用作额外的比较来源。
在有适当的人类数据的地方，在人类和人工智能生成的文本中也测量了同样的语言特征，这些参考文献也被用来评估词组的频率。
圆通翻译(英文译为日语，后用NLLB-200模式)用于区分语法偏好和语言必要性。 在修改句子结构时，基本上保留了意义，从而可以确定不必要的复杂措辞。
结果
实验1
如下文所示，在最初AI生成的文本中，特别是在前沿模型中，发现一些困难的构造更为频繁。 这被解释为倾向于采用不必要的复杂措辞：
在三个场合(聊天、编码和散文写作)对14种AI模型的6种语言特征进行了比较。 在聊天中测试了12种模式，在编码中测试了9种模式，在作文中测试了全部14种模式：
原AI生成的文本也与其重译版本进行了比较，允许测量句子结构的差异：
边际和近边模型被发现过度使用不寻常的词组组合，去名动词，无生命主体和对象的相对条款，可选词省略(见上图). 这些倾向在Llama和Olmo Base中普遍较弱，但在Olmo Think中更为明显。
在散文写作上观察到的差异比在聊天或编码上观察到的要小，这归因于散文的单时性，而扩展的前后对话可能鼓励更大的语言压缩。
提交人指出：
`PERSUADE散文显示，与聊天和[编码]相比，在几个类别(n-gm度量、依赖长度、堆叠名词、解名动词)的圆通译下，变化较小。
我们假设这是因为他们只涉及一个瞬间，我们推测某些模式可能会出现在长时间的前后对话中(这可能需要更多的信息压缩).
翻译后最小的变化出现在Llama和Olmo Base and Instruct中，在词组和动词用法上也最接近学生散文。 与此相对照，奥尔莫·思潮被发现表现得更像前沿模型。 观察到Claude Opus 5.5在21个比较中的19个比Opus 5使用难度较小的构造，其中8个没有重叠95%的置信间隔。
实验2
第二个实验旨在测试产生更短、更清晰的信息反应的潜在压力是否会使模型更倾向于语言构造，使人类更难以理解。
在这里，Qwen2.5-1.5B-Instruct使用Group相对政策优化(GRPO)进行微调，并负责将结构化，JSON格式化的事件描述转换为英语句子的模式。
调整后的培训的两个方面是：该模式在多大程度上被允许改变其通常的词语选择(β)；以及它因产生较长的答复而受到多大的处罚(λ)。 每个事件可以用两种语法验证的方式表示，在符号计数上已知的区别。
例如，“他给我们发了张照片”需要6个符号，而“他给我们发了张照片”需要7个符号。 还使用了单独的测试方法，两种说同一件事的方法要求相同的符号数量，以检查差异是否由句长引起。
对以前未见的句子进行了试验，使用了在培训期间遇到的四种构造；两种未经训练的构造，其中一种是可选的 " 那种 " 一词；另一种是涉及占有者、限定者、被动者和目的条款的无关构造。
这些测试旨在确定是否倾向于缩短措辞，并结转到培训期间从未遇到的建筑中：
由于对较长的答复的处罚有所增加，因此选择更短的措词更为频繁(直接在上文中提及)。 如果没有KL的罚则(限制模型偏离其最初的字数预测的距离)，变化会很快；随着它，同样的效果会更加逐渐地观察到。 没有两种惩罚，培训期间的偏好就大不相同，但即使是很短的刑期，也始终倾向于采用较短的替代办法。
以前未见的句子中也出现了这种倾向，其中涉及备选词 " that "。 向较远语法构造的转让要弱得多，只有在没有KL处罚的情况下才对占有者产生影响。
提交人的结论是：
‘我们将这些结果视为一个概念的证明，即用GRPO进行强化学习时使用的长度惩罚可以直接影响语法特征。
`我们注意到，该模型没有被直接指示使用压缩或稠密的语言，也没有直接获得关于能否获得合成交替的信息。
`相反，这种惩罚导致人们突然偏爱较短的混合体成员。 '
结论
我对AI生成的文本所特有的主播感兴趣，因此，这项新的提议揭示了LLMs所注意到的动词与他们现在同样注意到的压缩倾向之间的明显冲突。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Anderson's Angle
Why AI-Generated Text Is Harder to Understand Than You Think
Add Unite.AI to your preferred sources on Google
AI models may be becoming harder to understand because their training increasingly rewards solving problems efficiently rather than explaining things clearly to humans.
The above, unedited sentence was written by ChatGPT in response to a request for an ELI5, single-sentence summary of a science paper; and it features at least a few of the language-mangling traits explained in that new study*. The new work concludes that this issue affects all frontier models, as well as near-frontier models, and some open-weights models that were also tested.
For instance, the early part of the segment ‘their training increasingly rewards solving problems efficiently’ exemplifies the paper’s stance that AI models sometimes give non-living things human actions (since ‘training’ is described in the above example as actively rewarding something).
This can impede quick parsing of the phrase, as we have to internally acknowledge the attribution as an illustrative method, rather than take it literally (which is much simpler).
The latter section ‘increasingly rewards solving problems efficiently’ also squeezes several ideas together, forcing readers to work out what’s being rewarded, what’s being solved, and what’s becoming more efficient.
The paper states†:
‘[Readers] use animacy as a cue during parsing, and without additional context, a mismatch may elicit a P600 response†† associated with reanalysis.
‘However, rephrasing these sentences to avoid inanimate subjects would potentially require lengthening the sentence: for example, the word “bites” stands in for “causes an issue”.
‘We speculate that during training, models may have converged to using such constructions as efficient, albeit non-standard, ways of expressing meaning which would otherwise take up more tokens.’
Say What?
This is the core of the paper – that because modern AI models such as ChatGPT, Claude and Gemini are optimized for multiple types of problem-solving rather than specifically for communication, they tend to deliver language in the form of very compressed payloads that need some decoding.
This was a surprise to me, as I am constantly trying to persuade ChatGPT and its peers to be less verbose, and more compressed – a problem that is recognized and addressed in the literature.
Instead, the paper presents verbosity as a kind of zip-bomb that prevents easy understanding even in a sentence that is not very long. Therefore, combining the twin issues of excessive default reply length with the paper’s observation about excessive compression, we can imagine that the characteristic pattern of frontier AI’s default response style is quite distinct (and, presumably, this rather terse style is used in part as a determinant for AI-generated text in detectors, though the paper does not deal with this).
The authors also examine stacked noun compounds, exemplified in the problematic phrase Squad Helps Dog Bite†:
‘In English, nouns freely modify other nouns in an adjectival way: in the phrase coffee shop, coffee modifies shop, despite usually appearing as a noun. This feature enables stacking multiple nouns into rich compound words which denote highly specific meanings.
However, such stacks can introduce bracketing ambiguity, as in the headline Squad Helps Dog Bite Victim, and disrupt reading.’
So if LLMs are verbose via compression, are they being doubly verbose by producing excessive volumes of compressed text in response to a query? That means that not only is there a great deal of text to read through in response to your query, but the text itself may have a lot of ‘implicit’ meaning that needs decoding.
One 2023 paper posited that LLMs can become unnecessarily verbose due to AI judges being used during training, since these often reward longer answers – even when humans preferred shorter ones, in the study’s tests.
The authors of the new work attribute (see earlier quote) some of the compressive tendencies of frontier models to an architectural disposition to save tokens, which is a generous thought these days.
Let’s take a look at the methods, data and tests run by the authors for the study.
Data and Methods
Unusual word combinations (measured through n-grams, or sequences of words) were investigated for their unpredictability, with rare combinations associated with greater reading difficulty.
Dependency length (the distance between grammatically-connected words) was also measured, since longer distances place greater demands on working memory – particularly when a subject and its verb are separated by several intervening words.
Object relative clauses (constructions such as ‘the house the contractors built’) were examined, particularly when optional words were omitted (such as ‘that’ in ‘the house that the contractors built’).
The aforementioned Stacked noun compounds (three or more nouns joined together, as in ‘horror movie soundtrack generator’) were investigated as well, because the relationships between their component words can be ambiguous.
Denominal verbs (nouns used as verbs, such as ‘fork it’ instead of ‘put it on the fork’) were examined, alongside inanimate subjects paired with verbs normally associated with living things (as discussed earlier). Both constructions allow meanings to be compressed into fewer words, but additional interpretation may be required from readers.
The Contenders
The researchers collected AI-generated conversations from 14 models; Kimi K2.5; GLM 5; Claude Sonnet 5, Opus 5, Opus 5.5, and Fable 5; GPT-5.6 Luna, Terra and Sol; Olmo Base, Instruct and Think; and Llama Base and Instruct.
The tests comprised ten starting prompts for general conversation, and five for coding, with each task requiring a model to clone a real public GitHub repository, find bugs, and fix them in a sandbox.
Conversations were limited to 12 exchanges, while coding sessions were limited to 11, followed by a fixed closing message. In a separate test, each model was asked to write seven essays based on assignments from the PERSUADE 2.0 dataset.
For comparison, human-written texts were drawn from sources matched as closely as possible to the three types of AI output: OpenSubtitles (a collection of film and television subtitles unrelated to the open source subs platform) was used for general conversation; StackExchange (the online network of question-and-answer forums) for coding discussions; and student essays from the aforementioned PERSUADE 2.0 dataset for the writing tasks. Wikipedia was also used as an additional comparison source.
Where suitable human data was available, the same linguistic features were measured across human and AI-generated text, with these reference collections also used to assess the frequency of word combinations.
Round-trip translation (English to Japanese and back, using the NLLB-200 model) was used to distinguish stylistic preferences from linguistic necessity. Meaning was largely preserved while sentence structure was changed, allowing unnecessarily complex wording to be identified.
Results
Experiment 1
As shown below, several difficult constructions were found more frequently in the original AI-generated text, particularly among frontier models. This was interpreted as evidence of a preference for unnecessarily complex phrasing:
The six linguistic features were compared across 14 AI models in three settings (chat, coding and essay writing). Twelve models were tested in chat, nine in coding and all 14 in essay writing:
The original AI-generated text was also compared with its retranslated versions, allowing differences in sentence structure to be measured:
Frontier and near-frontier models were found to over-use unusual word combinations, denominal verbs, inanimate subjects and object relative clauses with optional words omitted (see graph above). These tendencies were generally weaker in Llama and Olmo Base, but were more pronounced in Olmo Think.
Smaller differences were observed in essay-writing than in chat or coding, which was attributed to the single-prompt nature of essays, with greater linguistic compression potentially encouraged by extended back-and-forth conversations.
The authors note:
‘The PERSUADE essays show smaller changes under round-trip translation on several categories (n-gram measures, dependency lengths, stacked nouns, denominal verbs) relative to chat and [coding].
‘We hypothesize that this is because they involve a single prompt, unlike the many iterated turns of chat and coding. We speculate that certain patterns may emerge in long back-and-forth conversations (which may demand more information compression).’
The least change after translation was found in Llama and Olmo Base and Instruct, which also came closest to student essays in their word combinations and verb usage. In contrast, Olmo Think was found to behave more like the frontier models. Claude Opus 5.5 was observed to use fewer difficult constructions than Opus 5 in 19 out of 21 comparisons, with non-overlapping 95% confidence intervals in eight of those.
Experiment 2
The second experiment was designed to test whether an underlying pressure to produce shorter and more information-dense responses might cause models to prefer linguistic constructions harder for humans to understand.
Here, Qwen2.5-1.5B-Instruct was fine-tuned using Group Relative Policy Optimization (GRPO), and the model tasked with converting structured, JSON-formatted descriptions of events into English sentences.
The two aspects of training adjusted were how much the model was allowed to change its usual choice of words (β); and how strongly it was penalized for producing longer responses (λ). Each event could be expressed in two grammatically-valid ways, with a known difference in token count.
For example, ‘He sent us a photo’ required six tokens, compared with seven for ‘He sent a photo to us’. A separate test was also used in which two ways of saying the same thing required the same number of tokens, to check whether differences were caused by sentence length.
The effect was tested on previously unseen sentences using four constructions encountered during training; two untrained constructions involving the optional word ‘that’; and unrelated constructions involving possessives, quantifiers, passives and purpose clauses.
The tests were designed to establish whether a preference for shorter wording would be learned and carried over into constructions never encountered during training:
As the penalty for longer responses was increased, shorter wording was chosen more often (illustrated directly above). Without the KL penalty (a restriction on how far the model could depart from its original word predictions), the shift was rapid; with it, the same effect was observed more gradually. Without either penalty, preferences varied widely between training runs, but even a small length penalty produced a consistent preference for shorter alternatives.
This preference was also observed in previously unseen sentence constructions involving the optional word ‘that’. Transfer to more distant grammatical constructions was much weaker, with an effect found only for possessives when the KL penalty was absent.
The authors conclude:
‘We take these results as a proof of concept that a length penalty used during reinforcement learning with GRPO can directly affect grammatical features.
‘We note that the model was not directly instructed to use compressed or dense language, and was not directly given information about the availability of syntactic alternations.
‘Rather, the penalty resulted in an emergent preference for shorter members of syntactic alternations.’
Conclusion
I have been interested in the anchors – both inadvertent and intentional – that characterize AI-generated text for some time. Therefore this new offering sheds some light on the apparent conflict between LLMs’ noted verbosity and their now equally-noted tendency to comp
Auto-fetched and translated by AI Atlas. All rights belong to the original author.