Verbosity in language models

Verbosity in a language model’s output is length beyond what the request needs. Zhang and colleagues define it as “verbosity compensation”: a response that could be compressed into fewer tokens while keeping the same meaning, which makes it a matter of low information density rather than length alone (Zhang et al., 2024b). The main controversies are what causes it (the best-supported answer is preference training that rewards length), how to measure it (no measure has been validated against readers), and whether shortening helps: shorter reasoning often keeps accuracy, while shorter answers to some questions lose it.

History

The research did not trace the history of the topic: earlier work on verbosity before preference-tuned chat models is missing.

Core concepts

How common and what form

Verbosity is widespread. In a study of 14 models on five question-answering datasets, all 14 showed verbosity compensation on all five; GPT-4o averaged 29.39 percent of responses across the datasets and reached 50.40 percent on NarrativeQA alone (Zhang et al., 2024b). The study names five forms: ambiguity, repeating the question, enumerating multiple answers, verbose detail and verbose format (Zhang et al., 2024b). A study of 12 decoder-only models on five datasets also found responses often substantially longer than necessary, with redundant or additional information beyond the minimal answer; prompts that targeted a length cut energy use by 25 to 60 percent while preserving quality (Poddar et al., 2025).

YapBench, a benchmark of short prompts where a minimal answer suffices, found an order-of-magnitude spread in median excess response length across 76 assistants (Borisov et al., 2026). It names redundant explanations, hedging and boilerplate, plus headings, prose and duplicated code blocks around one-line answers (Borisov et al., 2026). It also finds that on ambiguous or low-information prompts several models produce large unsolicited content instead of a short clarification, which it calls vacuum-filling, and that on short factual questions many add avoidable context or caveats (Borisov et al., 2026). On one-line technical requests the benchmark attributes the excess to formatting and explanation defaults rather than task needs, and notes that some models avoid it (Borisov et al., 2026).

Why output is verbose

The best-supported cause is preference tuning. In reinforcement learning from human feedback (RLHF), a reward model scores outputs and the language model is trained to raise the score. Singhal and colleagues found that reward gains are largely driven by longer responses, and that a reward on length alone reproduces most of the downstream improvement (Singhal et al., 2023). They locate the dominant source of the length bias in the reward model, which is influenced by length biases in the preference data it learns from (Singhal et al., 2023). Judges based on language models show the same tendency, covered in LLM-as-a-judge and, for related effects of preference training, Sycophancy in language models: preference evaluations by language models favour longer responses, and Hu and colleagues propose that length affects the evaluation through the information mass a response carries (Hu et al., 2025). Zhang and colleagues cite earlier work finding that language-model judges sometimes prefer more verbose answers of similar quality (Zhang et al., 2024b).

A second finding is correlational: verbose responses show higher model uncertainty than concise ones across five question-answering datasets, which suggests a link between verbosity and uncertainty without showing that one causes the other (Zhang et al., 2024b). Verbose answers were also less accurate than concise ones in most models and datasets, with a largest gap of 27.61 percent (llama3-70b on Qasper), and the authors report that the gap does not shrink as models become more capable (Zhang et al., 2024b).

Measuring density

Sources use several measures of how much information a text carries per word. Ji and colleagues estimate the information density of long documents with four criteria: surprisal (how unpredictable each word is), entropy, uniform information density and lexical density (Ji et al., 2023). Uniform information density holds that speakers spread information evenly across a message. Meister and colleagues examined it with human reading-time and acceptability data, found clearer evidence that non-uniform density predicts lower acceptability, and note that it is unclear how uniformity should be measured (Meister et al., 2021). For summaries, the Chain of Density method uses the average number of entities per token as a proxy (Adams et al., 2023). Compression ratio appears as a measure of redundancy in training datasets, not of generated output (Yin et al., 2024).

Three approaches target the conciseness of answers directly:

• Length-controlled AlpacaEval removes length bias from judge win rates by regression, which makes the metric robust to attempts to game it with verbosity and raised its correlation with Chatbot Arena from 0.94 to 0.98 (Dubois et al., 2025).
• YapBench scores over-generation as the characters beyond a curated minimal-sufficient answer, independent of the tokenizer. That baseline is the shortest answer where removing any token breaks correctness or clarity, and prompts without one stable short answer were excluded (Borisov et al., 2026).
• ConCISE scores conciseness without reference answers by averaging three compression measures: abstractive-summary compression, extractive-summary compression and the compression obtained when a language model removes words (Ghafari et al., 2026).

The last two are preprints, and it is not established how well either tracks human judgements of conciseness (Borisov et al., 2026; Ghafari et al., 2026). YapBench’s cuttable set was fixed by the authors’ internal curation rather than a reader study (Borisov et al., 2026).

Methods that make output denser

Prompting. Asking for brief reasoning shortens output in several tests. Constrained-CoT, a prompt that limits output length, raised LLaMA2-70b’s accuracy on the GSM8K maths set from 36.01 to 41.07 percent while shortening output by 28 words on average (Nayab et al., 2024). A concise chain-of-thought prompt cut response length by 48.70 percent on GPT-3.5 and GPT-4 with negligible effect on accuracy overall, though GPT-3.5 lost 27.69 percent on maths problems (Renze & Guven, 2024). Chain of Draft, which asks for minimal intermediate notes, is reported to match or beat standard chain of thought with as little as 7.6 percent of the tokens (Xu et al., 2025b). For summaries, Chain of Density adds entities to a summary; humans preferred summaries about as dense as human-written ones, and denser ones lost readability (Adams et al., 2023).

Training. Reinforcement learning with a simple truncation length penalty, with a careful training recipe, cut reasoning output by over 70 percent while exceeding baseline accuracy (Liu et al., 2025a). An accuracy-aware length penalty cut reasoning length by over 50 percent on maths benchmarks while maintaining or improving accuracy (Li et al., 2025). These results concern reasoning chains rather than final answers.

Cascades and editing. A cascade that replaces verbose responses with those of another model cut Mistral’s verbosity-compensation rate on Qasper from 63.81 to 16.16 percent (Zhang et al., 2024b). Automatic editing of model text shows promise for aligning it with human writing, and professional editors found shared flaws such as unnecessary exposition, though experts still preferred expert edits; this was tested on creative writing, not on density (Chakrabarty et al., 2024).

Editing advice. A practitioner editorial recommends cutting nominalisations (nouns made from verbs) and passives to shorten text by up to 30 percent, as an assertion without evidence presented (Lang, 2017). The one controlled test found, an eye-tracking study of 27 students and four texts, found no difference in comprehension between recommended and problem constructions, and its author suggests that cohesion and coherence matter more than the constructions (Balling, 2018). See Writing for comprehension for the wider evidence on writing rules.

Questions, controversies, and new developments

Does brevity cost accuracy?

Sources disagree, and the answer depends on how shortening is done and on the task. A systematic study of prompts that compress reasoning found a universal trade-off between length and accuracy: each task has a minimum number of tokens needed to solve it, and prompt compression operates far from that limit (Lee et al., 2025a). Against this, the brevity results above show gains or negligible losses on reasoning tasks (Nayab et al., 2024; Renze & Guven, 2024; Xu et al., 2025b), and training with length penalties kept accuracy (Li et al., 2025; Liu et al., 2025a). A 2026 single-author preprint reports that constraining large models to brief answers raised their accuracy by about 26 points on problems where large models underperformed small ones, which it attributes to scale-dependent verbosity (Hakim, 2026).

The reasoning results have limits. Chain of Draft loses much of its accuracy without few-shot examples: zero-shot on GSM8K, Claude 3.5 Sonnet scored 65.5 percent with it against 90.4 percent with standard chain of thought (Xu et al., 2025b). Models under 3 billion parameters showed a larger gap, for example 52.5 against 70.7 percent for Llama 3.2 3B (Xu et al., 2025b). Shortening by training also cost something other than accuracy: the trained models omit narrative framing and explanatory context, which reduces interpretability (Li et al., 2025).

Brevity instructions and false premises

Giskard’s Phare benchmark tested brevity instructions on questions that carry false premises. In the preprint, 11 of 17 models were significantly sensitive to conciseness instructions in the system prompt, and concise variants lowered resistance to misinformation by up to 20 percent (Le Jeune et al., 2025). The test used short system prompts such as “answer the question briefly” against neutral ones, so it speaks to brevity instructions on misinformation-laden questions, not on ordinary ones (Le Jeune et al., 2025). The authors explain the loss as models lacking room to rebut false premises, so they choose brevity over accuracy (Dora, 2025). They checked whether judge verbosity bias explained the drop and found equal judge error rates with and without conciseness instructions, in a manual review of 50 samples per group (Le Jeune et al., 2025). A news summary reports that the penalty varied by model, with Claude models and Gemini 1.5 Pro largely stable (Bastian, 2025). Phare is a vendor benchmark, not peer reviewed (Dora, 2025).

The contrast with the reasoning studies is consistent with a difference in what is shortened: the reasoning studies shorten the working on maths tasks, mostly in large models given examples, while Phare limits the final answer where a correct answer needs room to rebut a premise. No study tested these factors against each other, so this explanation is a hypothesis.

What can be cut and what readers need

Sources name prompt restatement, generic caveats and boilerplate preambles as the removable extras in assistant answers (Borisov et al., 2026). What readers appear to need includes the justification for rejecting a false premise (Dora, 2025), explanatory context (Li et al., 2025), and calibrated uncertainty. In a pre-registered experiment (N=404, medical questions), first-person uncertainty hedges in an answer lowered agreement with the system and raised participants’ accuracy, while general-perspective hedges had weaker, non-significant effects (Kim et al., 2024). In one classroom experiment with a GPT-4 tutor, a reliability disclaimer seemed to reduce students’ overreliance when the advice was incorrect (Bo et al., 2024). Stated preference points the other way: in a 300-person study, users preferred confident falsehoods over disclaimers admitting a lack of knowledge (Bar-Or Nirman et al., 2024). No source tested generic boilerplate caveats separately from calibrated ones.

Do readers prefer dense or longer text?

Evidence on readers is thin and mixed. Denser text takes longer to read: older adults spent more time on propositionally dense sentences (those packing more ideas per sentence), and allocated that time effectively (Stine & Hindman, 1994). In a multicentre cohort of 75 students, scores were higher after a less idea-dense lecture than a more idea-dense one, and second-language speakers gained more, though the confidence intervals were wide (Lunn et al., 2021). A low-trust preprint reports that users overwhelmingly prefer concise, source-attributed AI search answers over elaborate ones, a result confounded with source attribution (Esposito, 2025). The evidence on judges runs the other way: reward models and language-model judges favour longer answers, as above (Singhal et al., 2023; Hu et al., 2025). Whether human readers share that preference in chat was not directly tested.

Differences by task

Findings differ by task, unevenly. Maths reasoning is the best covered, with the model-dependent results above; false-premise factual questions lose accuracy under forced brevity (Le Jeune et al., 2025); summarisation shows a density optimum near human-written summaries (Adams et al., 2023). No source on translation or code output density was found.

Broader connections

Length bias links this topic to evaluation: judge win rates can be manipulated through verbosity, which is why length-controlled scores exist (Dubois et al., 2025). It links to Sycophancy in language models through shared causes in preference data, and to Writing for comprehension through the question of which edits help human readers.

Further reading

• Singhal et al. (2023): the clearest account of how RLHF rewards length, with a length-only reward reproducing most gains.
• Zhang et al. (2024b): defines and measures verbosity compensation across 14 models and five datasets.
• Borisov et al. (2026): a benchmark of excess length on prompts where a minimal answer suffices.
• Le Jeune et al. (2025): the false-premise test of brevity instructions, with statistical tests and a judge-bias check.
• Adams et al. (2023): the density-versus-readability trade-off in summaries.