Efficient coding is essential for sensory systems to extract meaningful information from the environment. Here, we investigate how stimulus-driven thermodynamic shifts and geometric reorganization enable efficient population coding. Using wide-field calcium imaging, we simultaneously recorded neuronal activity across the entire mouse V1 under the presentation of structured stimuli and characterized neural dynamics spanning microscale neuronal connectivity, mesoscale thermodynamic states, and macroscale manifold geometry. We found that stimulus presentation reorganized neuronal couplings into increasingly modular subnetworks, driving a shift from near-critical dynamics toward a more ordered regime. This shift coincided with a compression of neural population activity onto low-dimensional manifolds aligned with stimulus features, thereby enabling efficient coding. Furthermore, mathematical derivations and in silico perturbation experiments confirmed that selective modulations of neuronal connectivity altered both critical temperature and manifold geometry, establishing a causal link from neuronal couplings to efficient coding. Collectively, our findings suggest a mechanistic bridge between statistical physics and neural geometry, providing new theoretical insights into how neural networks dynamically transition into more stable and efficient coding states in response to the environment.

Developing robotic manipulation policies is iterative and hypothesis-driven: researchers test tactile sensing, gripper geometries, and sensor placements through real-world data collection and training. Yet even minor end-effector changes often require mechanical refitting and system re-integration, slowing iteration. We present RAPID, a full-stack reconfigurable platform designed to reduce this friction. RAPID is built around a tool-free, modular hardware architecture that unifies handheld data collection and robot deployment, and a matching software stack that maintains real-time awareness of the underlying hardware configuration through a driver-level Physical Mask derived from USB events. This modular hardware architecture reduces reconfiguration to seconds and makes systematic multi-modal ablation studies practical, allowing researchers to sweep diverse gripper and sensing configurations without repeated system bring-up. The Physical Mask exposes modality presence as an explicit runtime signal, enabling auto-configuration and graceful degradation under sensor hot-plug events, so policies can continue executing when sensors are physically added or removed. System-centric experiments show that RAPID reduces the setup time for multi-modal configurations by two orders of magnitude compared to traditional workflows and preserves policy execution under runtime sensor hot-unplug events. The hardware designs, drivers, and software stack are open-sourced at https://rapid-kit.github.io/ .

One major distinction between artificial neural networks and biological brains is the prevalence of extensive, long-range feedback connections in biological systems. Here we investigate unique contributions of these hierarchical feedback signals beyond feedforward processing and local recurrence by exploring their mechanistic role in resolving visual ambiguity caused by occlusion. Both empirical fMRI and EEG experiments and computational modeling show that when sensory evidence for faces became insufficient, the ventrolateral prefrontal cortex (vlPFC) sustained a low-dimensional belief state (e.g., animate vs. inanimate objects) and transmitted this abstract information back to the animacy map in the ventral temporal cortex (VTC) encompassing face-selective representations. Critically, in the hierarchical vision model inspired by this finding, this frontal feedback did not reshape the attractor geometry of the VTC; instead, it provided guidance to reroute ongoing neural dynamics away from ambiguous pseudo-states toward face attractor basins in the energy representational landscape. This control-based mechanism of feedback signals thus enabled perceptual completion by reconstructing missing facial features with temporal costs verified through EEG. Together, this multimodal study bridges analysis-by-synthesis theories of vision and dynamical-systems perspectives on long-range feedback as state-space control, and offers inspiration for the design of hierarchical AI architectures incorporating feedforward, recurrent, and feedback connections.

In contemporary large language models (LLMs), the swish-gated linear unit (SwiGLU) activation function is widely adopted to regulate the information flow and introduce non-linearity. For large positive inputs, SwiGLU approximates the quadratic function x^2, providing strong nonlinearity and expressive capacity. However, this property also causes numerical instability as the input or model scale increases, particularly in low-precision LLM training. The main reason is its approximate quadratic amplification, which enlarges the output range and exacerbates outliers. To address this issue, we propose a stable activation function, Power Linear Unit (PowLU), for large-scale LLM pre-training. Specifically, PowLU employs a rational power function to achieve adaptive nonlinearity, thereby improving representation ability and enabling stable training in spike regions. Moreover, we provide theoretical justification for several key properties of PowLU. Scaling law experiments confirm that the performance is consistent across model sizes, and further experimental results with the Ling architecture (7.9B and 124B total parameters) demonstrate that PowLU achieves competitive results against SwiGLU and SwiGLU-Clip in large-scale training of LLMs. In addition, the experimental results also show that PowLU effectively improves the scalability of the large-scale training of LLMs.

Face recognition is essential for human social life, and two core face areas have been identified in the human brain: the fusiform face area (FFA) and occipital face area (OFA). Previous studies have explored the functional division of the FFA and OFA in holistic face processing, using variations of whole faces where holistic processing was disrupted. However, a main prediction of holistic face processing (i.e., nonlinear integration of parts into wholes) lacks direct evidence. Here, we tested this prediction in the FFA and OFA by comparing their responses to incomplete versus complete faces. We generated novel facial fragments using AlexNet, a deep convolutional neural network (DCNN), and presented the facial fragments, their occluded counterparts, and complete face during fMRI scanning. Results revealed that the FFA exhibited higher activation for complete than incomplete faces, while the OFA showed comparable or higher activation for some incomplete faces than complete one. Moreover, the FFA's activation for the complete face was higher than the weighted average activation for some of its constituent incomplete face pairs, while this difference was not observed in the OFA. These results suggest that the FFA favors holistic processing and combines facial parts nonlinearly, whereas the OFA focuses on local parts and processes parts linearly. Our findings shed new light on the neural basis of holistic face processing and the functional division of the FFA and OFA in face recognition. The approach combining DCNNs with neuroimaging techniques can be extended to broader visual domains in future studies.

Abstract Working memory (WM) stores information after sensory input disappears and later retrieves it in a task-relevant format, but the mechanism unifying storage and retrieval remains unclear. Here we combine neural geometry analyses of macaque dorsolateral prefrontal cortex activity during a visuospatial delayed-match-to-sample task with computational modeling to test whether WM can be implemented as recurrent fast-weight computation. We found that the relational geometry of remembered locations was strongly expressed during sample presentation, degraded during the early delay, and reemerged before requirement in a partially distinct mnemonic subspace. A recurrent fast-weight programmer model, which implements a form of dynamic fast-weight memory closely related to linear Transformer computation, reproduced these latent-to-mnemonic dynamics. Direct inspection and perturbation of the model revealed that neural activity writes stimulus information into rapidly modifiable synaptic states, synaptic dynamics organize this latent memory over time, and recurrent readout queries the evolving state to generate task-relevant activity. These findings provide a unified account of WM storage and retrieval and suggest that biological WM and Transformer family architectures share an algorithmic principle of programmable temporary memory.

This paper presents a novel extension of neural scaling laws to Mixture-of-Experts (MoE) models, focusing on the optimal allocation of compute between expert and attention sub-layers. As MoE architectures have emerged as an efficient method for scaling model capacity without proportionally increasing computation, determining the optimal expert-attention compute ratio becomes critical. We define the ratio r as the fraction of total FLOPs per token dedicated to the expert layers versus the attention layers, and explore how this ratio interacts with the overall compute budget and model sparsity. Through extensive experiments with GPT-style MoE Transformers, we empirically find that the optimal ratio r^* follows a power-law relationship with total compute and varies with sparsity. Our analysis leads to an explicit formula for r^*, enabling precise control over the expert-attention compute allocation. We generalize the Chinchilla scaling law by incorporating this architectural parameter, providing a new framework for tuning MoE models beyond size and data. Our findings offer practical guidelines for designing efficient MoE models, optimizing performance while respecting fixed compute budgets.

To enable navigation in both physical and mental spaces, the human brain employs a cognitive map constructed from the global metrics of the entorhinal cortex and the local representations of the hippocampus. However, how these two regions coordinate to enable navigation remains poorly understood. Here, we designed an object-matching task where human participants unknowingly manipulated object variants arranged in a ring-like structure around a central prototype. Functional MRI revealed a threefold spatial periodicity in the hippocampal activity that tracked navigation directions from object variants to the central prototype. This hippocampal periodicity was phase-locked with the well-documented sixfold periodicity of the entorhinal cortex, suggesting hierarchical interaction between these regions. Consistent with this neural pattern, a corresponding threefold periodicity was observed in behavioral performance, which was synchronized with hippocampal activity. Finally, an EC-HPC PhaseSync model reproduced this phenomenon, in which the sixfold activity periodicity of entorhinal grid cells across directions projects vectorial representations to the hippocampus, and the collection of these vectors exhibits threefold periodicity to represent conceptual directions. Together, these findings reveal a periodic mechanism through which entorhinal grid codes structure hippocampal vector representations.

Abstract To achieve the computational goal of navigating in both physical and mental spaces, the human brain employs a cognitive map constructed by the global metrics of the entorhinal cortex and the local locations of the hippocampus. However, the mechanism by which these two areas interact to support navigation remains unclear. Here, we designed an object-matching task where human participants unknowingly manipulated object variants arranged in a ring-like structure around a central prototype. Functional MRI revealed a 3- fold spatial periodicity of hippocampal activity, which tracked the navigation trajectories from the original object variants to the central prototype in the object space. Importantly, this spatial periodicity of the hippocampus was phase-locked with the well-documented 6-fold periodicity of the entorhinal cortex, suggesting a periodic mechanism connecting these two areas. In addition, the 3-fold periodicity was replicated in human behavior, which varied with a function of spatial directions and phase-locked with hippocampal activity. Finally, we proposed an EC-HPC PhaseSync model, illustrating a framework of the hippocampal-entorhinal network, in which the 6-fold spatial periodicity of entorhinal grid cell populations embeds vector fields that are represented in the hippocampus for conceptual navigation.

The perceived similarity between objects has often been attributed to their physical and conceptual features, such as appearance and animacy, and the theoretical framework of object space is accordingly conceived. Here, we extend this framework by proposing that subjective object space may also be defined by embodied features, specifically action possibilities that objects afford to an agent (i.e. affordance) and their spatial relation with the agent (i.e. situatedness). To test this proposal, we quantified the embodied features with a set of action atoms. We found that embodied features explained the subjective similarity among familiar objects along with the objects' visual features. This observation was further replicated with a separate set of novel objects. Our study demonstrates that embodied features, which place objects within an ecological context, are essential in constructing object space in the human visual system, emphasizing the importance of incorporating embodiment as a fundamental dimension in our understanding of the visual world.

Despite extensive research, mechanisms underlying the emergence of cognitive modules remains elusive due to the complex interplay of genetic, developmental, and environmental factors. Computational modeling, however, provides a means of exploring their origins by simulating manipulations on these factors. In this study, we aimed to investigate the emergence of cognitive modules by developing the Dual-Task Meta-Learning Partitioned (DAMP) model, whose plastic architecture facilitates automatic structure optimization through a genetic algorithm that simulates natural selection by iteratively selecting for efficient learning fitness. We found that a specialized module for face identification robustly emerged in the DAMP model. Critically, the emergence of cognitive modules was not exclusive to faces in individual-level identification tasks. Rather, modular structures formed across all tested object categories in both categorization and identification tasks within our model. Interestingly, the formation of these modules was strongly influenced by the structural constraint of sparse connectivity within the network, suggesting that modularity may arise as an adaptation strategy to cope with the limitations imposed by sparse connections in biological neural networks. These findings provide a new evolutionary perspective on the development of cognitive modules in the human brain, highlighting the pivotal role of neural network structural properties in shaping cognitive functionality.

Trait motivation is fundamental in shaping human behaviors. Previous studies have primarily focused on their impact on affective and motivational processing, with their role in perceptual processes less investigated. The present study takes face perception, a crucial and well-studied perceptual process, as a representative specimen to examine the perceptual effect of trait motivation. We investigated whether the behavioral activation system (BAS) and the behavioral inhibition system (BIS) were associated with structural characteristics of the inferior temporal face-selective regions as well as face recognition performance. With a sample of Chinese young adults (N = 264), voxel-based morphometry revealed that BIS scores correlated with greater gray matter volume in the fusiform face area. Further, a higher BIS score was associated with slightly better performance in face recognition. These findings provide novel evidence that trait motivation, particularly behavioral inhibition, is linked to both the structure and function of the face processing system. This underlines the intrinsic coupling between motivational and perceptual systems, blurring the presumed divide between affective and perceptual processes. ### Competing Interest Statement The authors have declared no competing interest. the National Natural Science Foundation of China, 32371099, T248810018 the Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park, Z221100002722012

Understanding how artificial systems can develop spatial awareness and reasoning has long been a challenge in AI research. Traditional models often rely on passive observation, but embodied cognition theory suggests that deeper understanding emerges from active interaction with the environment. This study investigates whether neural networks can autonomously internalize spatial concepts through interaction, focusing on planar navigation tasks. Using Gated Recurrent Units (GRUs) combined with Meta-Reinforcement Learning (Meta-RL), we show that agents can learn to encode spatial properties like direction, distance, and obstacle avoidance. We introduce Hybrid Dynamical Systems (HDS) to model the agent-environment interaction as a closed dynamical system, revealing stable limit cycles that correspond to optimal navigation strategies. Ridge Representation allows us to map navigation paths into a fixed-dimensional behavioral space, enabling comparison with neural states. Canonical Correlation Analysis (CCA) confirms strong alignment between these representations, suggesting that the agent's neural states actively encode spatial knowledge. Intervention experiments further show that specific neural dimensions are causally linked to navigation performance. This work provides an approach to bridging the gap between action and perception in AI, offering new insights into building adaptive, interpretable models that can generalize across complex environments. The causal validation of neural representations also opens new avenues for understanding and controlling the internal mechanisms of AI systems, pushing the boundaries of how machines learn and reason in dynamic, real-world scenarios.

Classification constitutes a fundamental cognitive challenge for both biological and artificial intelligence systems. Here, we investigated how the brain categorizes stimuli that are not linearly separable in the physical world by analyzing the geometry of neural manifolds formed by macaque V2 neurons during a classification task involving motion-induced illusory contours. We identified two related but distinct neural manifolds: the sensory and perceptual manifolds. The sensory manifold was embedded in a three-dimensional subspace defined by three stimulus features, where contour orientations remained linearly inseparable. However, through a sequence of geometric transformations equivalent to twist operations, this three-dimensional sensory manifold expanded into a seven-dimensional perceptual manifold, enabling the linear separability of contour orientations. Computational modeling further demonstrated that this dimension expansion was facilitated by neurons exhibiting nonlinear mixed selectivity with heterogeneous response profiles. These findings provide insights into how biological neural networks enhance the dimensionality of representational spaces, illuminating the geometric mechanism underlying the transformation from sensation to perception.

Sensory inputs do more than activate individual neurons; they also reshape collective cortical dynamics. Using wide-field calcium imaging in mouse primary visual cortex (V1) during viewing factorial sets of shape, motion, and color stimuli, we simultaneously tracked thousands of neurons across the entire V1 and analyzed neural dynamics across thermodynamic, topological and geometric scales. While distributions of pairwise neuronal correlations remained unchanged, maximum-entropy models revealed that sensory inputs consistently lowered the network’s critical temperature, signifying a selective weakening of neuronal couplings and a departure from spontaneous near-criticality. This microscale decoupling dissolved modular subnetworks, integrating previously segregated neuronal communities. Consequently, neural population activity collapsed onto a lower-dimensional manifold aligned with stimulus-defined dimensions, thereby increasing linear decoding accuracy. Further formal proofs and computational modeling demonstrated a causal chain linking sensory-driven weakening of microscale couplings to reduced mesoscale modularity and enhanced macroscale manifold capacity. Thus, our study proposes microscale neuronal couplings as a mechanistic basis for cortical networks dynamically trading off exploratory variability for representational efficiency, thus unifying criticality and efficient coding, two classical signatures of cortical computation, within a coherent multiscale framework. ### Competing Interest Statement The authors have declared no competing interest.

Trait motivation is fundamental in shaping human behaviors. Previous studies have primarily focused on their impact on affective and motivational processing, with their role in perceptual processes less investigated. The present study takes face perception, a crucial and well-studied perceptual process, as a representative specimen to examine the perceptual effect of trait motivation. We investigated whether the behavioral activation system (BAS) and the behavioral inhibition system (BIS) were associated with structural characteristics of the inferior temporal face-selective regions as well as face recognition performance. With a sample of Chinese young adults (N = 264), voxel-based morphometry revealed that BIS scores correlated with greater gray matter volume in the fusiform face area. Further, a higher BIS score was associated with slightly better performance in face recognition. These findings provide novel evidence that trait motivation, particularly behavioral inhibition, is linked to both the structure and function of the face processing system. This underlines the intrinsic coupling between motivational and perceptual systems, blurring the presumed divide between affective and perceptual processes.

Balancing specificity and generalization in object recognition is a significant challenge for biological and artificial visual systems. Here, we investigated how the brain addresses this challenge by examining the relationship between interconnectivity of neural networks, dimensionality of neural space, and levels of abstraction in representing objects, employing combined neurophysiological data from macaques and computational modeling. We found that higher interconnectivity within area TEa of macaques' inferior temporal (IT) cortex was associated with lower dimensionality and greater generalization, while lower interconnectivity within area TEO correlated with higher dimensionality and greater specificity. To establish a causal link, we developed a brain-inspired computational model constrained by empirical wiring length. This structured interconnectivity created optimal dimensionality of the neural space, facilitating efficient energy distribution across the representational manifold embedded within the neural space, balancing specificity and generalization. Our findings underscore the critical role of structured connectivity in enabling robust object recognition through multi-level abstraction.

Gamma-frequency oscillations (~30-160 Hz) are a hallmark of neuronal synchronization, yet the fine-scale temporal arrangement of spikes within individual gamma cycles remains poorly understood. Here, we examine head direction (HD) cells in the mouse anterodorsal thalamic nucleus (ADn)\---|a circuit distinguished by prominent high-gamma activity\---|to uncover the principles governing gamma coordination. We reveal two fundamental mechanisms: (i) a stable anatomical gradient in which neurons firing earlier in the gamma cycle exhibit longer anticipatory time intervals (ATIs), and (ii) a dynamic rate-phase shift whereby spike timing advances as the animal's head aligns with a neuron's preferred direction. Together, these results delineate a spatiotemporally structured framework for gamma synchronization, advancing our understanding of the functional roles and circuit mechanisms underlying gamma rhythms in local brain networks. ### Competing Interest Statement The authors have declared no competing interest.