Innsbruck study finds prediction can conflict with energy efficiency in acting systems

Researchers at the University of Innsbruck published a theoretical study in Physical Review X on 5 October 2026 showing that, in some environments, an agent that acts on its surroundings cannot maximize both prediction and the useful energy it extracts. The finding concerns a physical model of information processing, with potential relevance to how biological and artificial agents organize memory.

Lukas J. Fiderer, Paul C. Barth, Isaac D. Smith and Hans J. Briegel developed a framework for what they call percept-action loops: repeated exchanges in which an agent observes an environment, acts on it and then receives new observations. An action can change what the agent will encounter next, making the information needed to predict future observations depend partly on its own past behavior.

Why prediction and forgetting can pull in different directions

The paper defines an agent’s work capacity as the maximum rate at which it can expect to extract useful work from its environment. This gives the authors a thermodynamic measure for comparing ways of retaining information. Within their framework, remembering everything useful for prediction does not always produce the greatest work capacity. In some environments, the agent must balance prediction against forgetting.

The distinction turns on the cost of retaining information about earlier actions. Past actions can help explain later observations, so keeping that history can improve a forecast. But the paper’s analysis finds a thermodynamic cost to retaining it. The authors conclude that an agent can sometimes improve work extraction by discarding part of its action history, even when doing so makes its predictions less complete.

How the result extends earlier work on predictive memory

The new result builds on a question studied in earlier information thermodynamics research: which memories help a system process information efficiently? In a 2020 Physical Review Letters paper, Susanne Still examined partially observable information engines and derived a lower bound on energy dissipation. That work found that retaining irrelevant information limits efficiency, while compressing memory to preserve information useful for prediction favors the analyzed bound.

Still’s paper supplies a useful precedent for the role of selective memory, but it does not test the Innsbruck team’s result. The newer framework addresses agents whose actions affect the observations they later receive. For those agents, information about an earlier action may remain useful for prediction while carrying a cost that works against maximum work extraction. That is the tension identified in the 2026 paper.

What the model means for artificial intelligence

The researchers present the finding as a possible principle for understanding biological and artificial agents. Its immediate claim is narrower than a way to cut the electricity used by current AI services or robots. The journal’s account describes a theoretical limit and does not report measured energy savings from applying the result to a deployed system. It also gives no figure for potential savings in existing AI hardware.

The university account says the trade-off in the model remains even when an agent has more storage capacity, separating the proposed physical constraint from the ordinary expense of providing memory hardware. That distinction matters when interpreting the finding: the result concerns the thermodynamics of an agent’s exchanges with its environment, rather than a comparison of available computer components.

Experimental test remains an open step

Phys.org identifies Fiderer as the study’s lead author and reports his view that a proof-of-principle demonstration should be possible with modern experimental setups. Such a demonstration is prospective. The sources do not establish that the predicted trade-off has been measured in an experiment, or that deliberately forgetting action history would save energy in today’s AI systems.