The research team, led by Vera Tolchennikova of RUDN University in Moscow together with colleagues at Lomonosov Moscow State University and Grodno State Medical University, built two desktop virtual mazes and placed 132 participants inside them. In the stable version, the internal organization of the maze remained identical from trial to trial, so any improvement in performance could be attributed to ordinary spatial learning. In the reconfigurable version, the maze’s topology changed after every trial based on the route the participant had chosen, ensuring that no memorized map could ever serve twice.

The task itself was deceptively simple. Participants had to find and collect coins hidden in two of four chests scattered through the maze. The optimal strategy was to navigate directly to the two rewarding compartments while avoiding unnecessary detours into the others. Measuring how quickly people converged on that strategy allowed the researchers to separate two distinct components of cognition: the acquisition of the task’s rules, and the efficiency with which those rules were executed in space.

The technical vocabulary at the heart of the paper is topological instability. Topology, in this context, refers to how the parts of the maze are connected, which corridors link to which junctions, and which compartments are reachable from which entrances. Instability means those connections are not fixed. A route that worked on one trial might be severed on the next, while entirely new passages could open. This is a fundamentally different challenge from merely reshuffling visual details; it attacks the abstract relational structure that spatial memory depends on.

Both groups of participants improved significantly across trials, which the authors interpret as evidence that the task rules themselves were learnable even in a perpetually shifting world. People could still work out that coins sat in two specific chests and that certain compartments were dead ends. Rule extraction, in other words, survived the chaos. But the price of that chaos showed up clearly in the kinematics of navigation, in the raw distances walked and the time spent inside the maze.

The reconfigurable group showed slower reductions in both path length and trial duration as trials accumulated. Quantitatively, participants in the changing maze had path length reductions across trials that lagged the stable group by a factor described by an incidence rate ratio of 1.10, and their trial duration improved more slowly still, with a ratio of 1.26. When the researchers statistically controlled for path length, the duration difference shrank but remained significant at a ratio of 1.17, suggesting that instability imposes a genuine temporal cost beyond simply forcing longer routes.

That residual effect is intriguing. If unstable participants were merely walking farther, controlling for path length should have erased the duration gap entirely. The fact that it did not hints at additional cognitive overhead: moments of hesitation at junctions, re-planning after encountering a blocked passage, or the constant double-checking of a mental map that can never be trusted. Topological instability, the authors conclude, imposes a measurable cost on task-solving efficiency even when the underlying rules are understood.

Not every hypothesis survived the statistical gauntlet. The reconfigurable group also showed a trend toward slower acquisition of optimal behavior, defined as collecting both coins without ever approaching an unnecessary compartment. However, that trend did not hold up under multiple comparison correction, a safeguard against false positives when many tests are run. The researchers are appropriately cautious: with the present sample, they could not determine whether instability truly slows rule acquisition, and they call for replication with a larger number of participants before drawing firm conclusions.

The study’s broader significance may lie less in its single set of statistics than in the tool it introduces. Virtual desktop mazes are cheap, scalable, and easy to deploy, and a reconfigurable variant opens a window onto a cognitive capacity that traditional paradigms struggle to isolate: flexibility in the face of structural change. The authors suggest that, with proper adaptation, the paradigm could be applied in developmental research, in studies of aging, and in clinical populations where cognitive flexibility is impaired, from attention deficits to neurodegenerative disease.

Funded by the Russian Science Foundation, the work arrives at a moment when researchers are increasingly questioning the ecological validity of static laboratory environments. Real-world navigation is a negotiation with a world that changes, and the brain systems supporting it, from hippocampal map-making to prefrontal planning, evolved under exactly such flux. By making a maze that refuses to stay still, this experiment takes a first quantitative step toward understanding what instability costs us, and how resilient human problem-solving turns out to be when the floor keeps shifting beneath our feet.

Subject of Research: Effects of changing environmental structure on human spatial problem-solving in virtual mazes

Article Title: Effects of topological instability on spatial problem solving in a virtual desktop maze task

Article References: Tolchennikova, V., Manukyan, P., Romanova, E., Shlepnev, P., & Aleksandrovich, A. (2026). Effects of topological instability on spatial problem solving in a virtual desktop maze task. Scientific Reports. https://doi.org/10.1038/s41598-026-73754-7

Image Credits: AI Generated

DOI: 10.1038/s41598-026-73754-7

Keywords: virtual maze, spatial learning, spatial problem-solving, topological instability, cognitive flexibility, navigation, virtual reality, Scientific Reports, reconfigurable environment, human cognition, task-rule acquisition, trial duration

The scale of the problem is difficult to overstate. Drawing on scenario-based 2025 estimates, the review reports that inference for a single model, GPT-4o, was modeled to consume between 391.5 and 463.3 gigawatt-hours of electricity annually, alongside 1,335 to 1,580 megaliters of freshwater and 138 to 163 kilotonnes of carbon-dioxide-equivalent emissions. Across 369 generative AI models, global training plus inference has been estimated at 25 to 41 terawatt-hours, concentrated largely in United States and Chinese data centers. Meanwhile, global data-center electricity consumption reached approximately 415 terawatt-hours in 2024, with AI identified as a key driver of projected growth through 2030. Against these figures, the brain’s 20-watt operating budget looks almost absurdly frugal, and the authors are careful to note they use it as contextual evidence that competent cognition can persist under tight constraints, not as a directly transferable efficiency benchmark.

To untangle the mess, the team conducted a systematized narrative review that began with 6,284 identified records, screened 898 at the full-text stage, and selected an initial set of 150 focal studies for detailed synthesis. From this evidence base they constructed a cross-layer framework organizing AI’s resource pressures into seven interacting regimes: diminishing returns from scaling, the cost of broad empirical experimentation, hardware energy floors, data-movement penalties, centralized infrastructure demands, benchmark-driven evaluation incentives, and static inference allocation. The central insight is that these regimes feed one another. Scaling inflates experimentation and memory traffic, infrastructure concentration shapes who can do research at all, and benchmark incentives push labs toward ever more compute-intensive optimization loops.

The first regime confronts the celebrated scaling laws that have governed transformer development. Performance improves predictably as parameters, data, and compute grow, but the returns follow shallow power laws. Using the fitted exponents reported by Kaplan and colleagues, the authors calculate that a mere 10 percent reduction in fitted test loss would require roughly 3.03 times more training tokens, 4.00 times more non-embedding parameters, or 8.23 times more minimum training compute. Human learning offers a striking contrast: people generalize new concepts from a handful of examples by reusing and recombining structured prior knowledge, including parts, rules, and skills, rather than accumulating vast task-specific experience. Translating that principle into AI suggests data selection and pruning, modular adapters, meta-learning, and compositional representations as promising directions, though the review cautions that most reported gains are measured in data volume or convergence time rather than actual energy saved.

The second regime examines the hidden cost of experimentation itself. Hyperparameter searches, architecture comparisons, ablation studies, and failed runs can dwarf the compute of a final training run, yet they are chronically underreported. The authors’ own analysis of 374 AI systems published between 1952 and 2022 found that training compute doubled every 20.6 months before 2010, but accelerated to every 7.3 months for regular-scale systems from 2010 to 2022 and every 6.6 months for higher-scale systems, and these figures exclude search and tuning overhead entirely. Human decision-making research suggests an alternative: rather than broad trial and error, people prune their search space using subgoals, means-ends analysis, and backward planning from desired outcomes. Multi-fidelity evaluation, successive halving, and early termination of unpromising configurations offer AI analogues, with energy-aware successive halving already demonstrating that wasteful experimental runs can be cut short.

Hardware tells a sobering story in the third regime. Although Moore’s law and Dennard scaling once delivered decades of performance gains at roughly constant power, that coupling has broken down. The review’s analysis of 62 machine-learning accelerators from 2017 to 2025 found that TDP-normalized peak tensor throughput increased about 20.0 percent per year, yet reported thermal design power rose 13.1 percent annually. Between 2023 and 2025, median efficiency improved 49 percent while median TDP climbed 94 percent. In other words, chips are getting more efficient per operation but also more power-hungry per device, and quantum tunnelling limits and voltage-scaling constraints impose physical floors that cannot be engineered away. The brain, by contrast, works within a fixed metabolic budget through sparse, discrete spiking, in which only a small fraction of neurons fires at any moment, and efficient coding that strips redundant information from sensory streams. Near-term AI translations include low-precision computation, quantization, and sparse activation, though the authors stress that savings depend heavily on hardware support and workload.

Perhaps the most technically striking regime concerns data movement. Modern AI hardware repeatedly shuttles weights and activations between memory levels, and accessing off-chip DRAM can cost hundreds of times more energy than a single arithmetic operation. The review cites normalized figures in which DRAM access costs 200 times the energy of one multiply-accumulate operation, compared with 6 times for an on-chip buffer and just 1 times for a local register file. Between 2012 and 2022, transistor-switching energy per bit fell roughly tenfold, yet reported end-to-end energy per instruction remains 2.3 times ten-to-the-fourth to 9.7 times ten-to-the-sixth times higher than transistor-level energy per bit once fetching, buffering, and control overheads are counted. Biological neural circuits sidestep this problem entirely by storing memories as synaptic changes within the very assemblies that process them, eliminating any von Neumann-style separation of storage and computation. Near-memory and in-memory computing architectures pursue the same locality principle, and three-dimensional analog in-memory accelerators have been reported to outperform conventional GPUs on memory-bound mixture-of-experts workloads, though much of the evidence still comes from simulation rather than production systems.

The remaining regimes highlight institutional and deployment-level inefficiencies. Centralized infrastructure concentrates frontier training in a handful of well-resourced industry labs, raising barriers to academic research and independent evaluation. Benchmark-driven optimization rewards marginal score gains while ignoring resource costs, a pattern the authors associate with so-called Red AI, and a sampled set of ACL papers confirmed that authors tend to prioritize accuracy over efficiency. Human cognition, by contrast, employs satisficing, accepting the first option that clears a sufficiency threshold rather than exhaustively maximizing a proxy. On the deployment side, standard dense models process every input through the full network regardless of difficulty, an approach that evidence suggests wastes computation on easy cases and can even hurt accuracy through overthinking. Adaptive techniques such as early exit, conditional routing, and speculative decoding can match computation to demand, mirroring how selective attention in the brain gates processing toward task-relevant signals.

The review’s most important caution is that efficiency claims rarely survive contact with the full system boundary. Retrieval-augmented generation can let smaller models rival larger ones, but corpus embedding, index construction, vector search, and reranking add their own energy costs. Mixture-of-experts routing reduces activated computation, yet expert imbalance and inter-device communication can erode the arithmetic savings. Multi-agent systems may improve task performance while adding repeated inference and coordination overhead, and controlled experiments indicate that sufficiently capable single agents can outgrow the benefits of collaboration entirely. The authors conclude that data curation, energy-aware experimentation, hardware-supported low-precision computation, and demand-adaptive inference represent the strongest near-term opportunities, while near- and in-memory computing hold longer-term promise. Their overarching prescription is cross-layer co-design evaluated with direct end-to-end energy measurements at matched capability, reliability, latency, and service quality, because a watt saved in one layer is all too easily spent, invisibly, in another.

Subject of Research: Energy constraints in large-scale artificial intelligence and human-cognition-inspired efficiency strategies for large language models

Article Title: Energy constraints in large scale artificial intelligence and insights from human cognition for efficient large language models

Article References: Amiri, M., Mokari, A., & Bocklitz, T. (2026). Energy constraints in large scale artificial intelligence and insights from human cognition for efficient large language models. Discover Artificial Intelligence, 6(1), Article 1432. https://doi.org/10.1007/s44163-026-02369-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02369-8

Keywords: large language models, energy efficiency, Green AI, human cognition, scaling laws, data movement, in-memory computing, adaptive inference, data centers, resource rationality, sustainability, machine learning

The cortical hierarchy is one of the organizing principles of modern neuroscience. At its lower tiers sit primary sensory and motor areas, which handle raw inputs from the eyes, ears, and body. Moving up the hierarchy, regions become progressively less tied to immediate sensation and more engaged in abstraction, prediction, language, and executive control, culminating in association cortices such as the prefrontal and parietal networks. Neuroscientists have mapped this gradient in detail using functional imaging, showing that higher-order regions integrate information from many lower-order sources. What has remained less clear is how the brain’s physical cabling supports this flow, and whether individual differences in that cabling relate to differences in how people think and reason.

White matter provides the anatomical substrate for that communication. Composed largely of axons wrapped in myelin, a fatty sheath that accelerates electrical signaling, white matter tracts form the brain’s long-distance infrastructure. Techniques such as diffusion magnetic resonance imaging allow researchers to infer the orientation and coherence of these fibers in living brains by tracking the movement of water molecules through tissue. Measures derived from these scans, including fractional anisotropy and related diffusion metrics, serve as indirect indicators of tract organization, myelination, and fiber density. In the new work, the researchers applied such methods to map how white matter pathways connect regions across successive levels of the cortical hierarchy.

The central finding is that the tracts most strongly associated with cognitive performance are not confined to any single level of the hierarchy. Instead, they thread through it, linking early sensory areas to intermediate association regions and onward to the most abstract frontal territories. This pattern suggests that efficient long-range communication across hierarchical levels, rather than the strength of any isolated hub, may be a key anatomical ingredient of higher cognition. The result aligns with a growing body of evidence that intelligence and related abilities depend on the coordinated activity of distributed networks, and that the brain’s wiring diagram constrains how effectively those networks can synchronize.

The notion of cognitive diversity is central to the study’s framing. Rather than ranking individuals on a single scale of ability, the researchers emphasize the many dimensions along which human cognition varies: some people excel at verbal reasoning, others at spatial manipulation, working memory, or cognitive control. The analysis indicates that distinct patterns of white matter organization across the cortical hierarchy relate to these different profiles. In other words, the anatomical substrate of cognition is not a single pipeline but a heterogeneous set of pathways whose varying configurations may give rise to the rich variety of mental strengths seen in the population.

Methodologically, the study draws on large-scale neuroimaging datasets in which hundreds to thousands of participants undergo diffusion imaging alongside extensive behavioral testing. This combination allows researchers to correlate tract-level anatomical measures with performance across multiple cognitive domains while controlling for confounds such as age, sex, and overall brain size. Statistical models in such analyses typically account for the fact that neighboring tracts share biological influences, and modern approaches increasingly test whether findings replicate across independent samples. The emphasis on hierarchical positioning, rather than simple regional labels, represents a methodological refinement: instead of asking whether a named tract predicts a named test, the authors asked whether connectivity spanning particular hierarchical distances predicts cognitive outcomes.

The findings carry implications for several long-standing debates. One concerns the neural basis of general intelligence, often indexed by the tendency of performance across diverse cognitive tests to correlate. Network-based accounts propose that a highly connected brain, with efficient communication among distributed regions, supports the flexible integration that demanding tasks require. The new evidence that white matter tracts span the hierarchy in a way that tracks cognitive diversity lends anatomical weight to that proposal, suggesting that the architecture of interregional communication is where some of the variance in human cognitive ability is physically realized.

A second implication concerns development and plasticity. White matter continues to mature well into adulthood, with myelination proceeding in a hierarchical fashion, from primary sensory tracts toward frontal pathways, over years and decades. If hierarchical connectivity supports cognitive diversity, then developmental changes in white matter may help explain why cognitive profiles shift across the lifespan, and why adolescence and early adulthood, periods of ongoing frontal myelination, are marked by gains in abstract reasoning and executive function. The study’s framework also offers a lens on conditions in which white matter integrity is disrupted, where atypical hierarchical connectivity may contribute to differences in cognitive function.

The researchers and outside commentators alike caution against overinterpreting the results. Diffusion imaging provides indirect measures of microstructure, and the relationship between diffusion metrics and the underlying biology of axons and myelin remains an active area of technical debate. Correlational findings in healthy adults cannot establish causation, and cognitive abilities reflect the interplay of genetics, environment, education, and experience alongside brain structure. The authors frame their contribution as a step toward an anatomical account of cognitive variation, one that must be integrated with functional imaging, genetic data, and longitudinal designs before its full significance can be judged.

Even with those caveats, the study adds a compelling piece to the picture of the human brain as a hierarchically organized communication network. By showing that the same white matter infrastructure carries signals from the senses to the heights of abstraction, and that the organization of that infrastructure varies meaningfully from person to person, the work underscores a principle increasingly central to neuroscience: to understand how minds differ, one must look not only at where the brain is active, but at how its regions are wired together across the full span of the cortical hierarchy.

Subject of Research: The role of anatomical white matter tracts spanning the cortical hierarchy in supporting individual differences in cognition

Article Title: Anatomical white matter tracts span the cortical hierarchy to support cognitive diversity

Article References: Bagautdinova, J., Shafiei, G., Luo, A. C., Pecsok, M. K., Salo, T., Alexander-Bloch, A. F., Bassett, D. S., Gardner, M. E., Gur, R. E., Gur, R. C., Mackey, A. P., Meisler, S. L., Misic, B., Moore, T. M., Roalf, D. R., Shinohara, R. T., Sydnor, V. J., Tong, T. T., Yeh, F.-C., … Satterthwaite, T. D. (2026). Anatomical white matter tracts span the cortical hierarchy to support cognitive diversity. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02559-5

Image Credits: AI Generated

DOI: 10.1038/s41562-026-02559-5

Keywords: white matter, cortical hierarchy, cognitive diversity, diffusion MRI, myelination, brain connectivity, neuroimaging, intelligence, association cortex, cognitive ability, brain networks, human cognition