← News

ResearchSources: arXiv · 16 Nov 2026

Census of a four-state world shows when less information can pay more

An arXiv paper enumerates every admissible representation and action rule in a four-state model, mapping exactly how retained information, task accuracy and hypothetical costs relate.

In plain words

The paper asks whether a decision system that keeps less information about the world can still perform as well as, or better than, one that keeps more. By exhaustively listing all possible representations in a small four-state model, the authors show the answer depends on which distinctions are preserved, not just how many bits are kept.

What was announced

The authors, publishing on arXiv, study a uniform four-state world and enumerate all admissible representations together with binary action rules, comparing constructions while holding other components fixed and by optimising the action rule for each module. They examine three tasks, mixtures of these tasks, and four hypothetical cost tariffs. One striking example: a representation retaining one bit can solve a task perfectly, while another retaining 1.5 bits on average reaches only 75% accuracy, because the two preserve different distinctions.

The census finds that in some comparisons the lower-information representation earns the higher payoff, while in others the more informative one wins; some pairs with different information amounts achieve equal accuracy. If a representation preserves all existing distinctions and adds new ones, its best zero-cost payoff cannot decrease, so any advantage of less information depends on which distinctions are retained. Changing the task set or resource price can flip which member of a fixed pair performs better. The authors stress these are exact, complete censuses of the specified finite sets, illustrating established distinctions between information quantity and decision value rather than proposing a new general law.

What it does not show

  • The evidence is the paper's abstract only; the full article's methods and results were not available for verification.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.