Pathological Verbosity in Large Language Models

20 concepts14 papers1 proposal

When you ask a modern Large Language Model (LLM) a straightforward factual question, it rarely gives a minimal answer. Instead, it typically wraps the core truth in layers of conversational filler, unsolicited explanations, and polite framing. If you ask for a translation, it might append an explanation of the idioms used. If you ask it to extract a data point from a clinical abstract, it often surrounds that data point with conversational chatter.

This phenomenon—variously termed response verbosity, model chatter, or pathological verbosity—is a systemic characteristic of modern instruction-tuned models. While a detailed response can feel helpful in a casual chat interface, this tendency acts as a significant tax on computational efficiency, obscures core reasoning, and introduces friction into automated downstream pipelines.

Why Models Over-Explain

To understand why language models default to wordiness, we have to look at how they are trained. Most production LLMs undergo instruction tuning and Reinforcement Learning from Human Feedback (RLHF). During this alignment phase, human evaluators and automated reward models tend to favor responses that appear thorough, polite, and comprehensive.

As Diana Galván-Sosa et al. observe in Rubrik's Cube: Testing a New Rubric for Evaluating Explanations on the CUBE dataset, current training objectives systematically prioritize information density and politeness over structural efficiency. The model learns a statistical prior that "longer equals better."

Consequently, the model operates under the assumption that generating more tokens increases the probability of satisfying the user's prompt. This behavioral prior is so strong that even when the input context is heavily compressed or shortened, the model's output length remains stubbornly inflated, as demonstrated by Yufei Tao et al. in Making a Long Story Short in Conversation Modeling.

The Threefold Cost of Verbosity

Pathological verbosity is not merely a stylistic quirk; it introduces severe practical penalties across three main dimensions:

1. The Verbosity Tax on Compute

Every token generated by an LLM requires a forward pass through the network, directly translating to computational time, financial cost, and energy consumption. In Brevity is the soul of sustainability: Characterizing LLM response lengths, Soham Poddar et al. characterize this as a literal "verbosity tax," finding that models frequently generate responses 3 to 10 times longer than necessary for factual queries.

2. Explanation Failure and Diluted Reasoning

Intuitively, one might think a longer explanation is more thorough. However, excessive wordiness frequently degrades the quality of model reasoning. When an LLM generates an explanation, redundant phrasing and irrelevant context often obscure the actual logical chain.

Diana Galván-Sosa et al. discovered that when LLM-generated justifications fail, it is rarely due to grammatical or linguistic errors; instead, the primary failure mode is the inclusion of irrelevant filler or repetitive context. Furthermore, as Qintong Li et al. note in Explanation Regeneration via Information Bottleneck, single-pass explanations are highly prone to introducing "hallucinations"—irrelevant details that sound plausible but do not actually support the final conclusion.

3. Downstream Pipeline Friction

When LLMs are embedded as components within larger software systems—such as Retrieval-Augmented Generation (RAG) or structured information extraction pipelines—verbosity breaks downstream parsing.

For instance, in clinical meta-analysis tasks, David Kartchner et al. show in Zero-Shot Information Extraction for Clinical Meta-Analysis using Large Language Models that models frequently append extraneous explanatory text to the requested data points. This requires engineers to build extensive post-processing scripts (e.g., string splitting, fuzzy matching, and normalization) simply to isolate the core answer.

Expected Output:  ["Adverse Event: Headaches"]
Actual LLM Output: "Sure! Based on the abstract provided, the primary adverse event reported by the patients was Headaches. Let me know if you need anything else!"

This conversational "chatter" similarly interferes with machine translation evaluation protocols, as documented by Tom Kocmi et al. in Findings of the WMT24 General Machine Translation Shared Task: The LLM Era Is Here but MT Is Not Solved Yet, and complicates data extraction from social media, as shown by Eduards Mukans et al. in RIGA at SMM4H-2024 Task 1: Enhancing ADE discovery with GPT-4. Whether the task is academic paper comprehension (Yelin Chen et al. in RPC-Bench: A Fine-grained Benchmark for Research Paper Comprehension) or psychological evidence mining (Federico Ravenda et al. in From Evidence Mining to Meta-Prediction: a Gradient of Methodologies for Task-Specific Challenges in Psychological Assessment), unconstrained length actively reduces the practical utility of the system.

Enforcing Brevity and Flow

To counteract this behavior, researchers attack verbosity from three angles: prompt engineering, algorithmic compression, and architectural design.

Prompt Engineering and Length Constraints

The most immediate remedy is altering the model's instructions. By explicitly penalizing length and rewarding conciseness, users can override the model's default conversational priors. Soham Poddar et al. demonstrated that enforcing brevity via simple prompt engineering can slash inference energy usage by up to 60% while simultaneously improving the precision of factual responses. Similarly, in educational dialogue generation, Seulgi Choi et al. found in VIVID: Human-AI Collaborative Authoring of Vicarious Dialogues from Lecture Videos that revising prompting strategies to strictly manage utterance length was vital to maintaining natural conversational pacing.

Algorithmic Bottlenecks

Rather than relying on a single generation pass, some frameworks introduce a secondary compression or regeneration step. Qintong Li et al. address explanation verbosity by applying the Information Bottleneck principle. Their approach forces the model to regenerate its rationale through a constrained bottleneck, stripping away non-essential tokens and redundant phrasing while preserving only the core information that directly impacts the final answer.

Flow-Aware Architectures

In multi-turn dialogue systems, verbosity often manifests as the model repeating known context or prematurely leaking details. For example, in conversational question generation, a model might accidentally include specific details of the target answer within the question itself.

To solve this, Jing Gu et al. introduced ChainCQG: Flow-Aware Conversational Question Generation, which explicitly models the "flow" of a dialogue across turns to prevent redundant information leakage. By tracking what has already been said and what should remain unsaid, flow-aware architectures mimic human conversational dynamics, where speakers naturally omit known information to maintain efficiency.

Key Foundational Work

ChainCQG: Flow-Aware Conversational Question Generation (2021) — contributed 1 concept elements to this direction.
Ordinal and Attribute Aware Response Generation in a Multimodal Dialogue System (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_explanation_redundancy_b0c0bade,
  title = {explanation redundancy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explanation-redundancy/rdr_b0c0bade69194ea88f8a7ab75c2ddd88},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_explanation_redundancy_b0c0bade_v1,
  title = {explanation redundancy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explanation-redundancy/rdr_b0c0bade69194ea88f8a7ab75c2ddd88/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake