To a computer, the difference between 10 and 20 is exactly the same as the difference between 1,000 and 1,010. To a human, however, these intervals feel fundamentally different. We are highly sensitive to the change from 10 to 20 (a doubling of magnitude), but the shift from 1,000 to 1,010 is almost imperceptible.
This phenomenon is captured by the Weber-Fechner Law, which suggests that the perceived intensity of a stimulus scales logarithmically with its physical magnitude. Across fields ranging from natural language processing to haptic robotics, researchers are finding that aligning artificial systems with this "compressive" mental representation is essential for building models that reason, feel, and interact like humans.
The Compressive Mental Number Line
In natural language processing, numbers have traditionally been treated as arbitrary tokens, no different from words like "apple." However, Avijit Thawani and colleagues have argued that this string-based approach fails to capture the inherent magnitude and mathematical relationships of numerals. By mapping numerical values to a logarithmic scale before vectorization, models can better mimic biological perception, prioritizing precision for small values while compressing the representation of vast ranges Thawani et al., 2021.
Recent evidence suggests that Large Language Models (LLMs) may already be doing this internally. H. V. AlquBoj analyzed the latent embeddings of LLMs and discovered a "mental number line" where the distance between small numbers is significantly larger than the distance between large ones. This sublinear geometry mirrors human numerical perception, suggesting that as models learn from human-generated text, they naturally adopt our compressed view of scale.
To explicitly improve this capability, Mandar Sharma and Naren Ramakrishnan introduced "Anchoring," a strategy that semantically primes rare numbers by relating them to well-known "anchor" points on a logarithmic scale. This allows models to generalize across ranges as large as 10 billion by focusing on relative scale rather than absolute linear distance Sharma et al., 2024.
Translating Force to Feeling
The need for logarithmic scaling is perhaps most visceral in haptics and robotics, where machines must translate physical forces into human-perceptible sensations.
When estimating hand pressure from video, Chengcheng Tang and James Hays found that discretizing pressure into logarithmic bins—rather than using linear regression—significantly improved accuracy. This approach provides higher resolution for light touches while still maintaining the ability to represent heavy presses, directly aligning the model's sensitivity with human force perception Tang et al., 2022.
Similarly, in teleoperated surgery or delicate manipulation, Noah Becker and Jan Peters demonstrated that applying a logarithmic transformation to force data before it reaches a haptic glove allows operators to "feel" subtle contact changes more clearly. This scaling expands the resolution for small forces, enabling operators to reduce object deformation by nearly 50% during sensitive tasks Becker et al., 2024.
Modeling Uncertainty and Common Sense
Logarithmic frameworks also provide a more robust way to model physical dimensions and uncertainty. Alan A. Stocker and Eero P. Simoncelli showed that human speed perception is best modeled using a LogNormal likelihood function, which accounts for the fact that our discrimination thresholds scale proportionally with the stimulus magnitude Stocker and Simoncelli, 2004.
This logic extends to "common sense" reasoning about the physical world. When Hannaneh Hajishirzi and Yejin Choi built a system to infer the sizes of objects (like elephants vs. butterflies), they used log-normal distributions to represent size ratios. This prevents the mathematical impossibility of negative physical dimensions and aligns with the psychological principle that we perceive size differences as ratios rather than absolute subtractions Hajishirzi et al., 2016.
Optimizing Quality of Experience
Beyond individual perception, logarithmic scaling is used to optimize complex systems like the Metaverse. Hongyang Du and Dusit Niyato applied the Weber-Fechner Law to network resource allocation. By modeling perceived satisfaction as a logarithmic function of rendering quality, they could allocate power more efficiently—focusing resources where the human eye is most sensitive to change and achieving a 20% improvement in user experience Du et al., 2022.
Whether it is a chemist "feeling" the quantum forces of a bond breaking Haag et al., 2014 or a user judging the carbon footprint of a flight Kristof et al., 2019, the evidence is clear: the most effective artificial representations are those that respect the non-linear, logarithmic nature of the human mind.