2026-08-08
From physics and the Free Energy Principle, the authors argue that offloading computation to the environment is universal across persistent systems, from planaria to LLMs.
"Cognitive offloading," as Risko and Gilbert defined it in 2016, is the use of physical or external action to rewrite a task's information-processing demands and save mental effort: turning your head, jotting a note, reaching for a calculator. The literature is almost entirely human-centric; even animal tool use and stigmergic memory barely connect to it. Fields (an independent physicist) and Levin (a Tufts biophysicist known for planarian regeneration and bioelectricity) push the idea to its limit: offloading is not a human trick or even an animal one, but a property of every information-processing system that manages to persist. The larger motive is a shared language for "mind" across radically different substrates, and a fresh way to relitigate the old fight over whether LLMs "understand."
Two pillars. First, basic physics. Any system A and its environment that obey Newton's third law can be treated as two information processors exchanging data over an N-bit channel. Separability (no entanglement) forces the channel to be lower-dimensional than either side, so almost all of each system's internal state is hidden from the other; each is a black box. Second, the Free Energy Principle (FEP): any system that persists must act to maintain its boundary with the environment, minimizing the mismatch between its actions and the environment's responses. The decisive move is that, by physical symmetry, the labels are swappable. So the environment of any active-inference agent is itself an active-inference agent, computing and "helping out." Offloading then falls out as a generic consequence: when a system acts, it farms computation out to an environment that, for small systems, has more compute than the system itself. For anything that counts as cognitive, computational offloading is cognitive offloading. The quantum formalism (Hilbert spaces, holographic screens, Berry phases) is there for generality across classical and quantum regimes, but the conclusion rests on symmetry plus the FEP; the apparatus is scaffolding.
Be clear up front: this is a theoretical paper with no new experiments and no benchmark numbers. Its results are five phenomena re-read as offloading, plus new readings of existing data. Niche construction (beaver dams, cell division, cancer cells remodeling their microenvironment) is a system rewriting the environment's model of it, and recording memory in the bargain. Kinematic replication: Xenobots, living constructs built from frog embryonic cells, actively gather precursor cells and assemble a copy of themselves, which is filling the environment with things like yourself. Bioelectric patterning: planarians (Dugesia japonica) regenerate whole bodies after being cut. The firmest piece of evidence is Figure 2: briefly altering the bioelectric coupling of wound-blastema cells yields a two-headed planarian that stays two-headed across multiple generations in plain water, with no further treatment, and with no detectable change in morphology or molecular markers beforehand. The "what a correct body looks like" memory lives in the bioelectric pattern, held by the cell collective; an individual cell does not know it. Shared semantics: two agents referring to one external object can be written as an LOCC protocol in quantum-information terms, which the paper likens to analogy-making. LLMs: a model interacts only by text in and out, and fluency is a fitness function imposed by the human environment and ruthlessly selected. Achieving fluency without extra-linguistic interaction with the world falsifies older theories that required explicit syntax rules, token-level semantics, or grounded denotation. The paper's reading, borrowing Searle, is that semantics is offloaded onto the environment: language users need not, and cannot definitively, know what words mean; the arbiter is the community plus the world. There is no head-to-head comparison with a baseline model, because the paper does not run that kind of study.
For people building AI, the payoff is the LLM section. The sharp claim: LLMs are fluent precisely because they offload semantics onto the human environment. The model does not need to "know" what words mean, because meaning is arbitrated by the linguistic community and the world. That relocates the Chinese Room and stochastic-parrots debate: understanding is no longer a state variable the model does or does not possess, but an outcome enacted in communication. The engineering hints are concrete. The authors suggest deliberately building environments and agent ecosystems that let systems offload to each other, and treating closed-loop robot-scientist platforms as tunable environments that sharpen a material's problem-solving. They flag a cost too: once you offload compute to a medium, you become exposed to its "agendas," and may take on as much work as you shed. That rhymes with the agent-design lesson that hidden coupling bites back. There is a measurement consequence: if part of an agent's competence lives in its environment, a bare benchmark is not measuring the agent alone.
The FEP, by the authors' own admission, "is not an empirical claim, but rather a principle that must be respected as a matter of logic," close to a tautology. If "offloading is universal" follows from something close to a tautology, its explanatory and predictive power is thin; the paper concedes "a general theory of the maintenance of boundary stability does not yet exist," and how any particular system persists is empirical. The move that the environment is itself an agent risks diluting cognition and agency to the point of vacuity: footnote 7 admits that even rocks and mountains undergo morphogenesis and decay. If everything computes and everything offloads, what would count as a counterexample is a real question. The quantum scaffolding is heavy machinery for a conclusion that likely follows from much weaker premises, and a skeptic can read it as borrowed rigor. The LLM section relocates "understanding" to "enacted pragmatic outcome" more than it resolves the original debate, and whether that is progress is itself contested (the paper cites both sides). The engineering extrapolations (hemisphere grafts for IQ, biobot offloading) are speculative, and this is an unreviewed preprint. No falsifiable experiments are proposed for the AI-relevant claims specifically.