Papers

Neural network loss landscapes have many symmetries, which are preserved by gradient flow but broken by finite-stepsize stochastic gradient descent (SGD). A canonical example of such a symmetry is scale in homogeneous networks: one can scale up the parameters in one layer and down in the next without changing the network output. Previous work has documented cases in which SGD breaks this symmetry in favor of balancing gradient noise or minimizing fluctuations. Here, we show that the solution geometry of undercomplete linear autoencoders instead selects a preferred sign for scale drift: on the PCA solution manifold, SGD favors large decoder weights. This directed scale drift occurs on a slow timescale, and its dynamics admit an analytically-tractable effective description. However, it cannot continue indefinitely: increasing scale eventually drives the dynamics towards a finite-stepsize stability boundary. The resulting solutions are sharper than a balanced baseline in the sense of the maximum eigenvalue of the loss Hessian, but different sharpness measures can move in opposing directions. Thus, undercomplete autoencoders give a concrete illustration of how loss geometry can convert residual gradient noise into directed motion along a manifold of functionally-equivalent solutions.

Contrastive representation learning is increasingly used to recover low-dimensional structure from neural recordings, but its output is typically validated by decoding accuracy rather than by the geometry of the manifold it produces. We apply CEBRA to EEG recorded from dyads in conversation, and analyze the resulting embedding, which training constrains to the 2D sphere. Labels describing the dyads, including the absolute difference between partners' autism-quotient scores, decode well above chance (0.77 against a 0.55 majority baseline for binary AQ magnitude; 0.44 against 0.25 for the six-class $|\Delta$AQ$|$ partition). However, the two permutation controls have notable differences in results: permuting labels over a frozen embedding yields p = 0.001, whereas retraining the encoder under each permutation yields p = 0.50. Only the latter tests the label rather than the geometry. Consistent with this, spherical mixture structure and per-class dispersion track identity rather than autism trait differences in dyads; frequency-band and non-oscillatory activity ablation controls do not change the results. However, participant-level model does separate from its identity-aware null (p = 0.0099) while speaker-versus-listener role analysis performs at chance in the same embedding, indicating a manifold organized by individual – and, in contrast with current neurolinguistics models, almost invariant to speaking vs. listening. Based on these results, we suggest that retraining-based nulls should be the default for grouped-data contrastive embeddings.

Understanding how humans reason – and how reasoning responses vary across tasks and individuals – remains a core challenge for modeling and explanation in cognitive science. We investigate the stability of response patterns within reasoners and whether variation in these patterns can be used to predict learning effects. We introduce a formal, geometry-based method to quantify distances between individual reasoning patterns and their internal variability, grounded in heuristic theories. The proposed framework is tested against experimental data via generalized linear mixed-effects models and clustering, where we find that our proposed variation measure interacts with correctness to predict performance gains. Moreover, we find that reasoning patterns are stable over time within the same reasoner. The method is general enough to be applied to other reasoning domains.

Understanding how neural activity represents higher-order cognition and how these representations evolve over time has long been a central pursuit in neuroscience. However, current analytical tools cannot easily distinguish representational plasticity from recording instability in chronic neural recordings. Here, we introduce NeuroLens (Latent Embeddings of Neural Semantics), a self-supervised model based on the Joint-Embedding Predictive Architecture (JEPA) framework that learns denoised, semantically informative latents from chronic neural recordings. An adaptive encoder maps changing neural populations into a common latent space, while a temporal predictor learns structure that supports prediction of future latent states. By predicting in latent space, NeuroLens captures temporally predictable structure and reduces sensitivity to transient, recording-specific variability. Across chronic intracortical data in mice and humans, the learned representations improve decoding of decision-making and semantic task variables. Multi-day pretraining enables generalization to future sessions, rapid few-shot adaptation to unseen neural populations, and more stable decoding over time than state-of-the-art baselines. Together, these results establish NeuroLens as a new paradigm for studying how neural representations change during learning and over long timescales.

Machine learning (ML) has emerged as a promising approach for improving diffusion MRI (dMRI) tractography, a task that remains limited by the intrinsic tension between local diffusion information and global anatomical plausibility. In this work, we systematically evaluate recurrent neural networks (RNNs) and Transformer models for iterative tractography, with particular attention to training strategies, input representations (including convolutional neural network (CNN)-based embeddings and end-of-sequence (EOS) tokens), and hyperparameter selection. We introduce a generation-validation phase enabling supervision at the streamline level during training, allowing supervision despite the mismatch between local loss functions and global streamline quality. Using the ISMRM2015 tractography challenge dataset, our models achieve the highest reported performance to date. Through controlled experiments, we quantify the impact of missing bundles, noisy or imperfect training streamlines, and invalid fibers in the training set. Finally, we demonstrate the applicability of our best-performing models for in vivo data from the Tractoinferno database. Overall, our results highlight both the potential and the limits of sequence-based deep learning models such as Transformers and RNNs for tractography, and emphasize the need for improved phantoms and evaluation methods for in vivo validation. We provide takeaways and recommendations for future researchers training and validating sequence-based supervised methods for tractography.

Neural activity often exhibits multiple timescales that can vary with behavioral states and task conditions. Identifying these timescales from neural recordings is important for better understanding neural computation and function. However, traditional approaches based on autocorrelation fitting are difficult to scale to high-dimensional population recordings and can become unreliable when neural dynamics change with behavior. State-space models have been a powerful framework for modeling high-dimensional neural population activity through latent dynamical systems, but standard formulations and inference methods do not explicitly account for multiple timescales and therefore do not guarantee accurate recovery of the underlying temporal structure. Motivated by these questions, we introduce the Multi-Timescale Switching Linear Dynamical System (MTS-SLDS), a framework for identifying regime-specific latent timescales from continuous or spiking neural observations. MTS-SLDS combines a multi-lag moment initialization, which captures temporal structure across multiple observation lags, with regime-conditioned Laplace-EM inference, which reduces mixing of dynamical statistics across uncertain regimes. Characteristic timescales can then be extracted directly from the eigenvalues of the learned latent transition matrices. In synthetic and neural experiments with Gaussian and Poisson spike observations, MTS-SLDS accurately recovers timescales and switching structure over multiple datasets.

Stimulus-driven auditory attention determines which sound gains priority when multiple sources compete without an explicit listening goal, yet most computational studies focus either on acoustic salience or on decoding predefined attended targets. This study investigates instruction-free auditory competition using the Stimulus-driven Auditory Attention (SAAD) paradigm and develops a neurophysiologically informed framework that integrates stimulus-derived sound priority with trial-specific EEG evidence. Behavioral analysis using a Bradley–Terry model showed that sound priority estimated from previous competitions generalized to unseen sound pairings, improving held-out prediction from an AUC of 0.577 to 0.718. EEG analysis further revealed mid-to-late centro-temporal lateralization associated with the reported selection side, with neural information remaining predictive beyond acoustic asymmetry. Guided by these findings, the proposed model first estimates a latent priority for each competing sound and forms relative stimulus evidence from their difference. A multi-scale EEG pathway with complementary signed and power-based readouts then extracts trial-specific neural evidence, which is incorporated through gated decision-level integration. The framework is evaluated using mirror-constrained and pairing-held-out protocols, together with representative acoustic, EEG, multimodal baselines, and systematic ablations. The results support a computational account in which spontaneous auditory selection reflects the interaction between generalizable stimulus priority and trial-specific neural variability.

Spectroscopy reads structure by sending a structured probe into a system and measuring what comes back. Here we apply this principle to animal decision-making. Harmonic Theory casts choice as motion on an angular landscape over heading, whose Fourier components form an animal's decision spectrum. We show that the arrangement of cues enters that landscape as a structure factor, so it factors like a diffraction amplitude, and symmetry imposes selection rules: a p-fold cue array annihilates every harmonic that is not a multiple of p, and sweeping two cues apart draws a notch through each harmonic. Cue geometry becomes a tunable probe: the experimenter sets which harmonics an experiment can see. We read these spectra from public recordings of flies, schooling fish, larval zebrafish, mice and a swimming alga, and from the fly's neural compass, where symmetric scenes suppress the forbidden harmonics and breaking that symmetry restores them. Measuring the compass form factor favours the theory's smooth bump over its square one, and raising cue contrast narrows that bump in every fly. A width fitted to one harmonic then predicts the next. Cue geometry is a spectrometer for decisions, and symmetry sets what it can see.

Patient-reported pain location is diagnostically decisive for some presentations and nearly uninformative for others. A companion paper argues this is not one gradient of diagnostic utility but three distinct failures of localization. This paper gives those failures their mathematics and shows they are one object: a single Bayesian generative model failing at different nodes - the likelihood, the model class, and group- or context-dependence in that same likelihood. The count is not in dispute: the first three are failures of the inference that produces a felt location, and the fourth, added here, is a failure of reporting it. Anatomical multiplexing is a non-identifiable inverse problem: a rank-deficient referral matrix sends distinct causes to one report. Delocalized amplification is a change of generative model whose dynamics are a neural-field bifurcation, with spatial extent as the order parameter. Referred and atypical displacement is group- or context-dependence in the presentation likelihood - a probability, not a loss - whose consequence sits one layer downstream, in a decision threshold varying with group prevalence and cost. Repeated observation cannot reduce recoverable information about a fixed inferential target, an exact chain-rule identity, while practical value can fall. The fourth node is the report itself. A spatial Bayesian model factorizes mislocalization into referral blur, an anatomical offset and a precision-weighted cognitive override, and reproduces phantom-limb, mirror-box and central-post-stroke reports as regimes of one equation. It is estimable under conditions the paper states as necessary or sufficient rather than assuming them; Sec. 8 gives the design that meets them. One control is reported against the paper's own interest: the transport statistic first used to measure migration scores a stationary, spreading profile as though it had travelled.

End-to-end backpropagation has been the dominant mode of training in deep learning, allowing for the coordination of parameter updates across layers of a neural network. Prior studies have explored alternative – and, in some cases, simpler – training mechanisms, showing that they can sometimes achieve performance similar to backpropagation. However, the architectural conditions under which locally optimized networks, which avoid end-to-end backpropagation of error, can learn representations comparable to those learned through end-to-end training remain unclear. We aim to answer this question in the context of self-supervised learning, an important framework for large-scale pretraining in artificial intelligence. Here, we investigate how network width and depth affect the efficacy of greedy layer-wise and end-to-end self-supervised training in convolutional networks. We find that in wider networks, the benefits of end-to-end backpropagation over greedy layer-wise training shrink: in relatively shallow and very wide networks, we even observed higher performance in models trained with greedy layer-wise training. Subsequent analysis of the representations formed by these networks shows that very wide greedy-trained networks exhibit more favorable representational geometry than do networks trained end-to-end with backpropagation. This work shows that width can compensate for restricted credit assignment and identifies differences in representational geometry as a potential mechanism for their improved performance.

Magnetoencephalography (MEG) is an imaging technique that offers a non-invasive, millisecond-resolution view of human brain activity. The increasing availability of MEG data presents an opportunity to take advantage of a recent advance in artificial intelligence, namely self-supervised foundation models. Existing foundation models for MEG (and electroencephalography) have been built on a transformer architecture. Here, we introduce MEG-Mamba: a generative foundation model for neural activity (source reconstructed, parcellated MEG) built on a Mamba architecture. MEG-Mamba is trained to predict the next token of a discretised MEG signal, conditioned on a brain region and recording session. MEG-Mamba surpasses the generative fidelity of a transformer-based alternative (MEG-GPT) while pre-training in less time (22 vs 400 GPU-hours) and modelling a longer context (4 vs 0.32 s). We evaluate MEG-Mamba by examining the spatio-spectral characteristics of the data it generates and by interpreting its learned embeddings. Furthermore, we demonstrate that lightweight stimulus conditioning (LoRA) can be used to generate realistic task-evoked responses that are not included in pre-training. Our results suggest that Mamba is a promising backbone for scaling neural foundation models.

Pseudo-random number generation often requires trade-offs among quality, power consumption, and bandwidth to produce unpredictable sequences of numbers. The brain, on the other hand, efficiently generates unpredictable output complex network dynamics occurring in a high-dimensional state. This state, which is hypothesized to be chaotic, relies on the balance between excitation and inhibition. Here, we investigated if computational models of these chaotic balanced states can be harnessed for Neuromorphic Pseudo-Random Number Generators (NPRNGs) in low power hardware. We successfully constructed a balanced spiking neural network model consisting of leaky-integrate-and-fire neurons that could be readily implemented in low power FPGAs and used as a NPRNG. The prototyped NPRNG consumed 3.24 mW during operation and produced pseudo-random numbers at 120kbps. In both hardware and software instantiations, NPRNGs produce high-quality random numbers as validated by standard metrics for testing RNG quality.

Network topology describes connections between neurons, whereas synaptic placement specifies how those connections are arranged within individual cells. How these levels of organization correspond remains incompletely understood. Here we show that connectivity between presynaptic neurons is reflected in relative input placement within Drosophila descending neurons (DNs). Across thousands of one-way DN connections in the independently reconstructed MaleCNS and FlyWire brains, inputs from sources also contacting the partner DN lay on average 9.4 and 8.0 {\mu}m nearer that partner's inputs along the receiver's neurites than other non-DN inputs. The same ordering held when source-input positions were fixed and partners with and without recorded source input were compared after standardizing reference-set size. In BANC and MaleCNS, shared input and measured spatial overlap provided complementary predictive information about DN interconnection. Two interconnected DNs sharing an ascending source differed in their identified cord outputs. Together, these findings link a three-neuron topological relationship to relative input placement, revealing a correspondence between network connectivity and the internal spatial organization of a receiving neuron.

We examine how local motif structure and global network topology jointly shape spiking dynamics in stochastic neuronal networks. Using networks of Izhikevich neurons with Erdős-Rényi (ER) and scale-free (SF) background connectivity, we compare motif-embedded networks with synapse-count-matched, non-motif controls under noise- and stimulus-driven protocols. Motif embedding increases noise-induced coherence in both topologies and provides a smaller improvement in signal transmission, while SF-based networks attain greater absolute coherence and transmit signals faster and more reliably than ER-based networks. Motif-rewiring experiments show that this benefit depends on the specific arrangement of intra-motif connections rather than on strong local connectivity alone: randomizing the motif skeleton reduces coherence overall, although rewiring the non-recurrent feed-forward loop and bi-parallel motif can increase it. Targeted ablation shows that the SF advantage depends in part on hub integrity; removing high-degree excitatory neurons impairs coherence and signal transmission more than matched random node or edge removal. Together, these results indicate that local motifs and global hubs make distinct, complementary contributions to coherent spiking dynamics.

Neuronal networks exhibit complex dynamics shaped by connectivity and stochastic input. Empirical studies show that neuronal networks contain recurring subgraphs, or motifs, but the collective influence of different motif types after embedding in large stochastic networks remains less well understood. We construct a spiking network composed of six representative structural classes and examine how intrinsic noise, coupling strength, inter-motif connectivity, network size, and neuronal heterogeneity shape synchronization. Both motif- and network-level coherence peak at intermediate noise intensities, consistent with coherence resonance. Bidirectionally coupled pairs (M2) and the type-2 recurrent feed-forward loop (M3c) consistently exhibit high motif-level coherence. Rewiring these motifs while preserving local synapse number and strength produces some of the largest reductions in network coherence, showing that connection arrangement contributes beyond strong local coupling alone. M2 and M3c also exhibit frequent spike doublets, linking short inter-spike intervals with elevated coherence. Relative to a degree- and weight-matched random control, the motif-structured network reaches greater coherence at weaker noise. Increasing network size enhances coherence until the response begins to saturate, whereas applied-current heterogeneity lowers peak coherence and shifts the optimum toward stronger noise without erasing the relative differences among motif classes. These results show that local connection arrangement remains dynamically consequential after embedding and can shape noise-driven coherence at the population level.

We find that major reported improvements in decoding words from non-invasive brain recordings are largely reproducible without any brain data. In the influential work of d'Ascoli et al. (2025), time series of brain activity from subjects perceiving continuous speech are segmented into fixed-length windows starting at each word. A neural network then generates predictions for all of the words in a sentence together. Neighbouring windows partially overlap, implicitly revealing the interval between words. Since these intervals indicate the duration of the words spoken, and different words tend to have different durations - for example, "the" is much shorter than "supercalifragilisticexpialidocious" - the neural network can improve its predictions of words without relying on the underlying brain activity. Consistent with this, the method reaches 22.0% balanced accuracy on synthetic signals containing no brain information, compared with 22.3% on real brain recordings. To prevent the network from learning this shortcut, we make a single, simple change. Instead of jointly encoding all windows in a sentence, we process each independently. As a result, the neural network achieves better performance by learning underlying word-specific information from brain recordings. This makes two existing strategies become much more effective than before. Both aggregating predictions from distinct neural responses to the same word and using a pretrained LLM as a linguistic prior now substantially improve results. On our perceived speech benchmark, this simple recipe (SimpleB2T) achieves a word error rate of 36.6% with five observations per word, approaching past invasive speech decoding performance, albeit under different conditions. The results in this work expose an important shortcut in brain-to-text decoding and show that removing it leads to a simple and considerably more effective strategy.

How is computation organized and reused across tasks and time in a trained recurrent network? Most analyses emphasize the geometry of neural activity, dynamical motifs, or local perturbation growth. We instead study the network's global first-order perturbation response. The finite-horizon Green's operator maps perturbations at each source along a trajectory to their downstream state-space responses and therefore directly represents perturbation routing. Simple reductions of this operator provide task-to-task and time-to-time views of the same computation, while matrix-free products make these views accessible without constructing the full operator. In a flexible multitask recurrent network, task reductions reveal structured reuse of known computational motifs, while temporal reductions reveal causal pathways and how they emerge during training. Our main point is simple: the Green's operator provides a global response geometry for mapping the organization of learned dynamical computation.

When recalling the orientation of recent stimuli, observers are systematically biased away from the cardinal axes. The prevailing explanation is that this ”anti-cardinal bias” arises because cardinal orientations are encoded with greater neural resources and therefore less noise, consistent with efficient coding of environmentally common features. Under this account, the bias should occur independently for each stimulus; any serial attraction towards previously seen orientations should be independent of absolute orientation. By contrast, we find that serial effects in orientation recall are unexpectedly strong and reflect a hitherto unrecognised hierarchical interaction, consistent with recurrent connectivity across hierarchies of sensory processing in the brain. We found that serial dependence in visual orientation recall is dominated not by attraction to the previous orientation itself, but to a compound feature representing both a tilt and its mirror reflection about the cardinal axes. Consistent with greater spatial abstraction at higher levels, the influence of this compound feature persisted across the visual midline, whereas orientation-specific effects weakened between hemifields. We developed a quantitative model based on the optimal combination of a noisy representation of the current stimulus with a hierarchical memory trace incorporating these compound features. Remarkably, although the coding fidelity was uniform across angles, this model accurately reproduced not only the serial dependence, but also the anti-cardinal bias. Our results suggest that memory traces are not isolated to single features, but reflect compound reactivation of hierarchical representations. Furthermore, they reveal that inference across hierarchical representations can generate systematic perceptual distortions previously ascribed to coding constraints.

Intrinsic motivation plays a central role in adaptive and goal-directed behavior by conferring agents reward-independent objectives and biases useful to act in noisy and uncertain environments. Active Inference addresses the problem of acting in a partially observable environment through a principled framework for belief updating and action selection. A key component of Active Inference is the specification of prior preferences, which shapes behavior by encoding desirable future outcomes. An intrinsic motivation approach called the Maximum Occupancy Principle (MOP) proposes that agents act so as to maximize occupancy over future paths of states and actions, with no preferences or epistemic targets. Despite its simple formulation, MOP gives rise to rich and adaptive behaviors that combine exploratory variability with goal-directed dynamics. In this work, we extend MOP to partially observable environments and introduce a Bellman reformulation of the Expected Free Energy for Active Inference, both incorporating belief-based inference over hidden states as part of the agent state. The Bellman formulation enables tractable offline computation via value iteration over the full belief-state space. We compare the resulting behaviors in a set of minimal experimental settings with uncertain food sources. We find that MOP agents switch between goal-directed (food seeking) behavior and exploration between different food sources, depending on their energy available and their belief state. In contrast, Active Inference agents mostly inhabit regions around a single food source, a strategy having both high pragmatic and epistemic value. We finally compare with Empowerment, which is shown to be qualitatively similar to Active Inference.

Randomised copies of a connectome are the usual baseline for asking whether measured wiring matters, and the answer depends on what the randomisation preserves. We evolved embodied foraging agents whose brains are a compressed adult Drosophila connectome (FlyWire v783; 512 cell-type groups and 1,000 Kenyon cells) alongside agents built on randomised wiring, in pre-registered experiments with ten seeds, four ecologies and 600 generations. Two standard randomisations, a column shuffle and degree-preserving edge swaps, route 10.6 to 10.7 % of olfactory output directly onto descending motor groups, against 0.012 % in the connectome. On the registered primary endpoint, fitness averaged over the run, no difference was detected; at the last common-garden probe the connectome was behind both controls (-0.22 and -0.20 fitness units on seed means). Against controls that keep every sensory-output and motor-input edge and rewire only the interior, the seed-mean difference lay within a +/-0.10 equivalence bound (+0.002 and -0.074, unchanged under a calibration that also matches activity spread), although per ecology the interior column shuffle was ahead by 0.26 in one of four ecologies at ten seeds, a lead that ten further pre-registered seeds did not replicate. Rewiring the connectome so that it acquires the shortcut raised its fitness by 0.44 (10 of 10 seeds) and its dependence on olfaction from 0.15 to 0.99; graded doses raised both in step; at comparable swap counts the full dose was ahead of an interior-only sham by 0.53 (10 of 10 seeds); and a sham that rewired the same boundary edges without creating shortcuts matched the connectome (+0.007) while the full dose was ahead of it by 0.60. What a null preserves at the sensory-motor boundary can decide an evolutionary connectome comparison, and sensory-to-motor path statistics belong next to the degree statistics a null is said to preserve.

Intracortical motor decoders degrade across sessions because the set of recorded units changes and persisting units can alter how their firing relates to behavior. Most existing methods update network weights on each new session or rely on unlabeled activity, which does not directly reveal such changes. We present APST, an Association Profile-conditioned Set-Temporal transformer that adapts to new sessions with all network weights frozen. From a few labeled calibration trials, APST summarizes how each unit's firing relates to behavior in a four-dimensional association profile computed in closed form. The profiles condition a set-attention encoder that accepts any number and order of units, followed by a causal transformer for streaming decoding. On held-out DANDI688 sessions from two monkeys, APST reaches velocity $R^2$ of $0.78$ and $0.81$, versus $0.40$ and $0.58$ for a variant that uses neural activity alone, and matches or exceeds an RNN fine-tuned on the same trials. On FALCON private held-out evaluation, it attains $R^2$ of $0.65$, $0.42$, and $0.44$ on M1, M2, and H1.

A foundational principle of connectionism is that perception, action, and cognition emerge from parallel computations among simple, interconnected units that generate and rely on neural representations. Accordingly, researchers employ multivariate pattern analysis to decode and compare the neural codes of artificial and biological networks, aiming to uncover their functions. However, there is limited analytical understanding of how a network's representation and function relate, despite this being essential to any quantitative notion of underlying function or functional similarity. We address this question using analysable two-layer linear networks and numerical simulations in non-linear networks. We find that function and representation are dissociated, allowing representational similarity without functional similarity and vice versa. Further, we show that neither robustness to input noise nor the level of generalization error constrain representations to the task. In contrast, networks robust to parameter noise have limited representational flexibility and must employ task-specific representations. Our findings suggest that representational alignment reflects computational advantages beyond functional alignment alone, with significant implications for interpreting and comparing the representations of connectionist systems.

Studying the alignment between the internal representations of vision models and the responses of the visual cortex to the same observed visual stimuli has enabled us to better understand human visual processing. However, studies so far have largely overlooked the fact that the human brain not only processes observed visual stimuli, but also predicts upcoming stimuli based on what has been observed. Accordingly, we hypothesize that internal representations for generating future video frames are better aligned with the predictive nature of human visual processing than representations of the observed video itself. To this end, we compare the alignment between human video-watching fMRI responses in the visual cortex and the internal representations from two types of video diffusion models, an autoregressive (AR) model and its non-AR base model. We first conduct a within-model analysis of the AR video diffusion model and show that the representations for future video generation align better with the visual cortex than the representations of the observed video. We then compare the internal representations of the AR model with those of its non-AR base model and again show that the representations for future video generation align better with the visual cortex than the representations for observed video reconstruction by the base model. Specifically, the alignment of observed video reconstruction is concentrated in lower-order visual cortex, whereas that of future video generation is concentrated in higher-order visual cortex. Finally, we show in a human behavioral experiment that humans prefer videos generated by amplifying the contributions of individual layers that align better with the visual cortex.

Connectome models of the fly nerve cord generate walking-like motor rhythms, but oscillation alone does not show that the specific wiring matters. Here we provide, to our knowledge, the first test of which features of motor output depend on the specific wiring. We simulated the leg motor systems of two independent Drosophila connectomes, with synapse counts as fixed weights and glutamatergic synapses treated as inhibitory, and compared each with six families of rewired networks that preserve progressively more of its structure, using pre-registered criteria. We find that rhythm is generic but antagonist coordination is not: many rewired networks were more rhythmic than the real ones, yet the real wiring coordinated antagonistic motor pools more strongly than every rewired network, most of all at the thorax–coxa joint. We trace this specificity to how premotor input is allocated between antagonistic pools. Both connectomes carry Sherrington's reciprocal innervation—neurons that excite one pool inhibit its antagonist—and no rewired network does. Reassigning premotor inputs between the pools abolished coordination even when motor neurons' typical input strength changed little (all pre-registered criteria met in one connectome; same direction in the other). Coordination, not rhythm, therefore reveals whether a connectome model's wiring matters.

Recent advances in musculoskeletal modeling and reinforcement learning have enabled muscle-actuated agents to reproduce increasingly complex human motions. Yet these capabilities remain largely confined to flat ground, in part because motion datasets rarely include aligned terrain geometry and because retargeting terrain interactions to complex musculoskeletal bodies is challenging. We present TERRA, an end-to-end pipeline for terrain-aware retargeting and control of musculoskeletal locomotion. From kinematic trajectories alone, TERRA combines terrain priors, estimated contacts, and negative free-space evidence to recover task-relevant support geometry. TERRA further considers anatomical, tendon-continuity, and contact constraints during retargeting. Using the resulting motion-terrain pairs from five datasets, we successfully train a single muscle-actuated control policy on 9.4 hours of diverse locomotion. Across reconstruction, retargeting, and held-out tracking benchmarks, TERRA improves terrain accuracy, sharply reduces anatomical and interaction violations, and achieves the highest observed completion rate over supported terrain families. Overall, TERRA provides a practical route from scene-less motion data to muscle-actuated locomotion over diverse non-flat terrain. Project website: https://cnai.epfl.ch/terra/

Tree-like branching structures are common in nature, from botanical trees to neurons, blood vessels and respiratory trees. Their branching shape often reflects function, making structural modelling central to understanding how these systems work. Because acquiring real-world 3D data is often expensive or infeasible, realistic generative models are valuable for simulation and data augmentation. Existing morphology-specific models either constrain how topology is generated or rely on hand-tuned, mechanistic procedures. Generic 3D graph generators, by contrast, do not exploit or enforce the structure of trees. We propose Autoregressive Frontier Expansion, a generative framework that constructs trees through an iterative expansion process, simulating the biological growth of real trees. At each step, a flow-matching model parameterised by an SO(2)-equivariant GNN expands the frontier by predicting whether each active branch bifurcates or terminates. We evaluate our method on cortical neurons and botanical trees in unconditional, class-conditioned, and morphology-guided generation. Across both domains, the generated morphologies agree closely with the reference distributions and, in conditional experiments, with the specified targets.

Global Neuronal Workspace Theory (GNWT) is one of the most extensively developed empirical research programmes on access consciousness. By contrast, the Motivated Emotional Mind (MEM) model proposes an embodied, semi-hierarchical associative memory in which representational selection, action, recurrent reconstruction of modality-specific fields, interoception, and valence form a single functional cycle. This article assesses whether MEM mechanisms can reproduce the functions explained by GNWT and whether they yield additional hypotheses concerning phenomenal quality. We first reconstruct the origins, mechanistic core, strengths, and limitations of GNWT, and then present the origins of MEM, the architecture of semblions, and the FFS-RP cycle. In the comparative analysis, we distinguish functional similarity from mechanistic identity and empirical equivalence. We examine 25 publications: twelve canonical studies of GNWT or critiques thereof, and thirteen studies of perception, interoception, and affect. For each study, we juxtapose the GNWT interpretation with a possible MEM reinterpretation. The findings indicate that most classical effects - including threshold dynamics, maintenance, recurrence, long-range connectivity, and flexible use of information - are compatible with both theories. This compatibility weakens the claim that the GNWT interpretation is exclusive, but does not yet provide theory-specific support for MEM. MEM advances a broader hypothesis integrating access, modality-specific reconstruction, bodily state, affect, and action; its potential explanatory advantage nevertheless remains a hypothesis requiring preregistered quantitative tests. We propose a series of interventions designed to dissociate sensory re-entry, global availability, report, and interoceptive valuation.

Teleoperation systems are tuned for channel fidelity, while whether the operator experiences the device as part of the body, the Sense of Embodiment (SoE), is measured only afterwards, by questionnaire. Predictive-processing accounts suggest controlling devices to reduce the mismatch between the operator's predictions and the returned feedback, but those predictions are unobservable, and an objective that only penalizes mismatch is minimized by removing feedback. We formulate an embodiment-aware controller, the Universal Embodiment Engine (UEE), that infers the operator's embodiment and visuo-proprioceptive cue weighting from implicit gaze and pupil signals and task outcome, and chooses bounded device settings under explicit preferences, cast as a discrete Active Inference agent. In simulations with 300 heterogeneous synthetic operators, the UEE found the suitable setting within half a minute for most operators, before identifying their exact type, and came close to an oracle in the second half of the session (embodiment 1.68 against 0.73 for the best fixed setting, on a 0-2 scale). Adapting without reading the operator did no better than fixed control, and model-free bandits did worse, whereas an expected-utility controller with the same inference did exactly as well: the benefit comes from Bayesian inference over the operator with explicit preferences, not from the information-seeking term of Active Inference. A naive prediction-error minimizer withheld feedback, as its objective implies, and lost task success (0.74 vs 0.90). The benefit shrank but persisted for operators outside the controller's model family, grew with the variety of the population, and vanished when the controller trusted an uninformative signal or when cue weighting changed mid-session without being modeled. These failures show what studies with people must establish first: calibrated signals and a model of change.

On a single task, deep networks can learn many solutions, depending on their optimizer, training data, architecture, and hyperparameters. Many of these solutions are mode-connected: rather than isolated points in weight space, they are connected by low-loss regions. Yet how their internal computation varies within these regions is unknown. A parallel line of work has identified the degeneracy of neural representations: many networks reach similar training loss with distinct internal structures. However, it is unclear how these solutions are related in weight space. We unify these subfields and show for the first time that many different internal mechanisms exist within a local mode-connected region in weight space. To do so, we introduce Hessian Null Space Continuation (HNC), a scalable method that uses local curvature to traverse regions of weight space that preserve network function, and can be steered toward solutions with specified properties. In RNNs trained on a memory task, HNC reaches drastically different representations and dynamics with maintained behavior. In ImageNet-trained Vision Transformers, HNC finds representations that differ more from the original network than any independently trained model with a different architecture or objective. In reinforcement-learning agents, HNC uncovers a distinct navigation strategy at comparable return and exposes reward hacking in an AI Safety Gridworld. Finally, HNC measures the local geometry of the solution set, showing how model size and task complexity shape its dimension and functional sensitivity. Our results show that a surprisingly large amount of representational diversity exists near a single trained solution, unseen by standard gradient-based optimization. HNC identifies and quantifies this diversity, opening new possibilities for mechanistic understanding of solution spaces and for model merging, editing, and fine-tuning.

Brain-AI alignment is often interpreted as a sign that model and brain perform similar computations. Whether the aligned units are causally involved in model computation is rarely checked. On an abstract pattern-completion task (AAABAAA $\rightarrow$ B), we compare LLM attention-head representations with human EEG and test how ablating those heads affects task performance. Alignment and causation dissociate: brain-aligned heads contribute to performance, but their removal is substantially less disruptive than removal of heads selected via attribution patching. We compare two head sets that prior interpretability work defines without reference to the brain: concept vectors (CVs), which represent abstract patterns across formats, and function vectors (FVs), selected for their contribution to correct-answer prediction. Brain alignment shows little association with FV scores, while its association with CV scores varies across models. Among brain-aligned heads, we find recurring attention profiles: one emphasizes distinctive elements (novelty heads), the other repeating elements (repetition heads). The novelty family tracks salience and attends to the same elements that humans look at, yet its removal is less damaging than random ablation on average. Repetition heads contribute modestly to performance and are associated with abstract-pattern representation (CVs). Across 17 models spanning 3B-72B parameters, FV-ranked removal is substantially more disruptive than brain-ranked removal. Brain alignment thus captures how the model reads the stimulus, and only faintly captures how it represents the pattern and solves the task.