Brevity Bias in Communication

Updated 10 October 2025

Brevity bias is a phenomenon where communication systems, from human language to machine outputs, favor shorter expressions for efficiency.
Empirical studies across languages and species using statistical analyses validate this bias, while methodological nuances can reveal notable exceptions.
Recent research in machine learning demonstrates that managing brevity bias is key to balancing efficiency with expressivity in AI-generated content.

Brevity bias denotes a systemic tendency—across both biological and computational communication systems—to favor succinct forms or outputs, often as a result of pressure for efficiency, limited cognitive or physical resources, or learned bias in model optimization. In human language, the concept is classically connected to Zipf’s law of abbreviation, whereby more frequent words are shorter; it further extends to animal signaling, memory encoding, neural machine translation, and large-scale AI model outputs. Recent analyses have both affirmed universality across species and highlighted failures or inversions (often due to methodological or statistical caveats), while contemporary machine learning research investigates brevity as both a practical and a theoretical constraint in optimization, expressivity, interpretability, and sustainability.

1. Theoretical Foundations: Brevity Bias and Least-Effort Principles

A foundational aspect of brevity bias is observed in Zipf’s law of abbreviation, which states that more frequently used linguistic signals (words) tend to be shorter (measured in letters, phonemes, or duration). This pattern is interpreted as evidence for the least-effort principle: communication systems are under selective pressure to minimize energetic or cognitive cost while maximizing efficiency. The cost function $E = \sum_i p_i e_i$ , with $p_i$ as the frequency/probability and $e_i$ as the cost (length, duration), forms the basis for this compression principle. Mathematical derivations show that minimizing expected code length naturally forces shorter forms onto more frequent signals (via swapping arguments where $\Delta E = (p_i - p_j)(e_j - e_i) < 0$ establishes the optimality of brevity for frequent elements) (Ferrer-i-Cancho et al., 2013).

Empirical support comes not only from human language (English, German, Spanish) but also from non-human species (dolphins, macaques, chickadees), where negative correlations between signal frequency and length are reproducibly found (Casas et al., 2019, Semple et al., 2012, Ferrer-i-Cancho et al., 2013).

2. Empirical Evidence and Cross-Species Robustness

Extensive correlation analyses—across languages and species—affirm that both the law of abbreviation and the meaning-frequency law (more frequent words are more polysemous) hold robustly. For example, Pearson/Spearman/Kendall correlations show that frequent words are shorter and more polysemous in English, Dutch, and Spanish, using both character, phoneme, and syllable-based length measures (Casas et al., 2019).

Some taxa, however, serve as putative exceptions. For New World primates (common marmosets and golden-backed uakaris), the law sometimes fails to appear in aggregate data. Yet, upon partitioning call repertoires by signal function (e.g., long-distance vs. low-duration calls), significant brevity bias emerges within receptive subsets. Critically, the visibility of brevity bias depends on sample size: small repertoires often fall below statistical power thresholds (with minimum $n^*$ derived for significance), risking false negatives (Ferrer-i-Cancho et al., 2012).

In animal communication, the presence or absence of brevity bias is modulated by both function (e.g., redundancy needs in long-range signals) and methodological limits (e.g., partitioning, cluster analysis, and the use of one-tailed tests).

A nuanced extension involves the interplay between word frequency, information content, and emotional valence. Quantitatively, information content is measured as self-information $I(w) = -\log P(w)$ , exhibiting a statistically significant negative correlation with valence: positive words (more frequent) carry less information, while negative words (rarer) are more informative (Pearson’s $\rho \approx -0.3$ to $-0.4$ across several corpora) (Garcia et al., 2011). This supports the Pollyanna hypothesis, which posits a bias toward positive, less informative expressions in human language.

Context-sensitive measures, such as $I_{\text{context}}(w) = -\frac{1}{N} \sum_i \log P(w|c_i)$ , further reveal that positive words are predicable in more routine, less information-dense contexts, whereas negative words are both rarer and more surprising.

4. Algorithms, Expressivity, and Computational Trade-offs

In algorithmic settings, brevity bias manifests as a preference for succinct outputs. The bias–expressivity trade-off provides the formal framework: increasing bias (toward brevity) reduces the entropy (flexibility) of outcome distributions. For a search space $\Omega$ , with algorithm-induced distribution $\overline{P}_{\mathcal{D}}$ , expressivity is defined as $H(\overline{P}_{\mathcal{D}}) = H(\mathcal{U}) - D_{KL}(\overline{P}_{\mathcal{D}}\,||\,\mathcal{U})$ . The bounds $H(\overline{P}_{\mathcal{D}}) \leq \log_2 |\Omega| - 2\,[\text{bias}(\mathcal{D}, t)]^2$ and $\text{bias}(\mathcal{D}, t) \leq \sqrt{(1/2)\,[\log_2 |\Omega| - H(\overline{P}_{\mathcal{D}})]}$ are proved, indicating that strong brevity bias entails diminished expressivity—potentially undesirable in contexts demanding adaptability, creativity, or coverage (Lauw et al., 2019).

5. Brevity Bias in Machine Learning and LLM Outputs

Modern neural models, especially translation and LLMs, demonstrate both brevity and verbosity biases. In neural machine translation, locally normalized models exhibit a tendency to truncate outputs prematurely ("brevity bias"), which degrades translation quality at larger beam widths. Solutions include length normalization, the GNMT scaling, and the per-word reward method $s'(e) = s(e) + \gamma \cdot m$ , with the reward $\gamma$ tuned via perceptron-style updates (Murray et al., 2018).

Recent benchmarking studies show that contemporary LLMs commonly output far longer answers than necessary, with minimal content comprising less than half of generated tokens (Poddar et al., 10 Jun 2025). Response trimming via prompt engineering (e.g., "Answer briefly.", "Only provide the minimal answer.", length constraints) achieves up to 60% reduction in response length—yielding 25–60% energy savings without appreciable loss in answer quality.

6. Reasoning Efficiency: Overthinking, Trimming, and Fine-Grained Suppression

Advances in large reasoning models (LRMs) have revealed a recurring overthinking phenomenon, where models produce unnecessarily long, convoluted solutions. The test-time dynamic inference trimming method, EDIT (Han et al., 7 Sep 2025), employs dual-goal search—jointly optimizing correctness and brevity by constraining reasoning steps and iteratively searching for the shortest path that maintains answer fidelity. The method enables up to 40% reduction in reasoning chain length without accuracy loss; this efficiency is measured via custom metrics (Accuracy per Effective Length, Budget-Constrained Accuracy curves).

Further, suppression of self-affirmation reflections—redundant post-solution tokens ("Wait", "but wait")—provides a fine-grained strategy for length compression. Thresholding on generation probabilities of leading reflection tokens and filtering during decoding achieves 19–50% length compression without lowering correctness (Liu et al., 14 Jun 2025).

7. Critiques, Caveats, and Shifting Scientific Preferences

The historical principle of Occam’s razor, privileging brevity and parsimony in scientific explanation, faces inversion in the contemporary academic environment (Mazin, 2022). Increasingly, incentive structures (publication, funding, social allure) encourage complex and buzzword-rich interpretations at the expense of simpler, direct explanations. This preference shift represents a sociological challenge to brevity bias, potentially leading to overcomplicated, less accurate literature.

8. Interpretability, User-Centered Design, and Human Understanding

In model interpretability, the assumption that the shortest rationales are most comprehensible has been challenged. Controlled studies demonstrate that for human users, extremely brief explanations often lack sufficient context for correct inference, whereas moderately longer rationales (around 40–50% of original text) yield optimal understanding (Shen et al., 2022). This suggests that brevity must be balanced against informativeness when designing user-facing AI systems.

9. Ongoing Challenges and Future Directions

Brevity bias remains a powerful organizing principle in communication, but its optimal expression is context-dependent. Research continues into refining trimming and compression algorithms, further cross-species comparisons, user-centered evaluation of explanation length, and recalibrating system-level incentives to privilege efficient, clear communication over unnecessary elaboration.

In summary, brevity bias encapsulates a multi-faceted research theme, connecting statistical universals, algorithmic trade-offs, practical system optimization, and broader sociocultural patterns. Its paper informs the design and analysis of efficient communicative, cognitive, and computational systems across domains.