# q-bio.NC Published articles for q-bio.NC. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Structure is in the zoom: probing neural symmetry through dimensionality scaling DevFeed: [Structure is in the zoom: probing neural symmetry through dimensionality scaling]() Original publisher: [Read original article]() Author: Adil Amin Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Neural Network](), [Deep neural networks](), [hugging face]() Tags: [cond-mat-dis-nn](), [cs-cl](), [cs-lg](), [language](), [neural](), [q-bio-nc](), [scaling](), [symmetry](), [v1]() ### AI overview The authors argue that participation-ratio scaling depends on how neural activity is sampled, so they separate a sampling floor from shifts measured along a declared axis. In mouse V1, three axes yield slopes of 0.25, 0.31, and 0.35; only drift direction carries structure. That direction-aligned shift exceeds a label-permutation null in eight of eight grating recordings and is positive in 32 of 32 animals from a second lab. The analysis attributes distinct roles to even and odd sectors of the code's O(2) symmetry. Applied to five language-model architectures, the split distinguishes token-inherited content from depth-built structure; a nuisance-matched null attributes a reported grammaticality signal of 20 standard deviations to carrier composition. The abstract describes the method as requiring no trained probe and applicable layer by layer to Hugging Face models. ### Source excerpt arXiv:2611.08506v1 Announce Type: cross Abstract: How much of a neural population's measured dimensionality is the code, and how much is the act of measuring it? The slope of the participation ratio against subsample size is reported for cortical populations and for the layers of language models as a number that belongs to the system. It does not. One patch of eleven thousand mouse V1 neurons returns three slopes, 0.25 along random stimulus subsets, 0.31 along drift direction, 0.35 along neuron count, and only the middle one carries structure; the other two are the sampling floor. We split every exponent into that floor and a shift earned along a declared axis, and read the shift against two nulls: permuting the labels, and permuting them only among samples that share a nuisance. On V1 the direction-aligned shift clears the label null in eight of eight grating recordings, is positive in 32 of 32 animals in a second laboratory, and reads both sectors of the code's O(2) symmetry. The even sector sets how the ladder starts: it decides which order of adding classes climbs faster, predicts the opposite order in an equivariant network, and flows under coarse-graining by a closed-form factor that says what orientation columns do to mesoscale signals. The odd sector, carried by the direction-selective neurons, keeps the ladder climbing past half the classes. On language models the same split separates content inherited from the tokens from construction built with depth, on five architectures, and the second null shows that a grammaticality signal of twenty standard deviations is entirely carrier composition. With a floor and a declared axis, dimensionality scaling is an axis-resolved measurement, and where the axis follows a known symmetry it resolves the sector content of the code. The instrument, rung, trains no probe and reads any array or any Hugging Face model layer by layer (https://github.com/adilamin89/structure-in-the-zoom). ## Invariance before structure: two modes of emergent compression in representation learning DevFeed: [Invariance before structure: two modes of emergent compression in representation learning]() Original publisher: [Read original article]() Author: Hardik Rajpal, Dan Goodman Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Representation learning](), [Machine learning](), [Neural Network]() Tags: [compression](), [cs-lg](), [cs-ne](), [invariance](), [learning](), [q-bio-nc](), [structure]() ### AI overview The proposed Emergent Adaptive Representation Learning (EARL) framework separates two kinds of compression in learned representations. State-space compression means the representation predicts its own future beyond the predictive power of individual network units; parameter-space compression means networks with different parameters converge on the same representation. In a recurrent reservoir with a trained bottleneck, parameter-space compression increased within the first few training epochs, partly independently of the task. State-space compression increased about 30 epochs later, only as the network began learning the task's temporal structure. The results suggest substrate invariance and task-specific structure emerge in distinct phases. ### Source excerpt arXiv:2611.07952v1 Announce Type: new Abstract: Learned representations show striking invariance across biological and artificial neural networks, with different substrates and variations driven by drift, heterogeneity and initialisation. To understand this invariance, we propose Emergent Adaptive Representation Learning (EARL), a framework that distinguishes two modes of emergence in learned representations: state-space compression, where a learned representation gains predictive power over its own future beyond what the network's units can predict individually, and parameter-space compression, where networks with different micro-level parameters converge to the same representation. We formalise both as Fisher sensitivities and introduce ensemble-level information-theoretic measures for each. In a recurrent reservoir with a trained bottleneck, parameter-space compression rises within the first few epochs of training and partly independently of the task, whereas state-space compression rises about 30 epochs later and only when the network begins to learn the temporal structure of the task. These results suggest that invariance to the substrate and task-specific emergent structure are acquired in distinct phases of learning. ## Characterizing universal object representations across vision models DevFeed: [Characterizing universal object representations across vision models]() Original publisher: [Read original article]() Author: Florian P. Mahner, Johannes Roth, Ka Chun Lam, Michael F. Bonner, Francisco Pereira, Martin N. Hebart Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Artificial Intelligence]() Tags: [architecture](), [convergence](), [cs-cv](), [cs-lg](), [deep-neural-networks](), [dimension](), [dimensions](), [interpretability](), [it](), [models](), [object](), [performance](), [q-bio-nc](), [semantic](), [similarity](), [vision]() ### AI overview Across 162 vision models, the study decomposes object similarity structure into non-negative dimensions and measures how often each dimension recurs. Recurring, universal dimensions are more interpretable and more strongly tied to conceptual image properties than model-specific ones. Differences in architecture, objective, training data, model size, and performance do not explain which universal dimensions emerge. Models with more such dimensions better predict macaque IT activity and human similarity judgments, suggesting a link between representational universality and biological vision. ### Source excerpt arXiv:2605.13675v2 Announce Type: replace-cross Abstract: Deep neural networks trained with different architectures, objectives, and datasets have been reported to converge on similar visual representations. However, what remains unknown is which visual properties models actually converge on and which factors may underlie this convergence. To address this, we decompose the object similarity structure of 162 diverse vision models into a small set of non-negative dimensions. To determine universal versus model-specific dimensions, we then estimate how often each dimension reappears across models. In contrast to model-specific dimensions, universal dimensions are more interpretable and more strongly driven by conceptual image properties, indicating the relevance of interpretability and semantic content as implicit factors driving universality across models. Differences in architecture, objective function, training data, model size, and model performance do not explain the emergence of universal dimensions. However, models with more universal dimensions also better predict macaque IT activity and human similarity judgments, suggesting that universality reflects representations relevant to biological vision. These findings have important implications for understanding the emergent representations underlying deep neural network models and their alignment with biological vision. ## BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks DevFeed: [BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks]() Original publisher: [Read original article]() Author: Carlo Alberto Barbano (MIND), Demian Wassermann (MIND) Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Machine learning](), [AI Foundation Models](), [Open Data]() Tags: [abstract](), [approach](), [arxiv](), [beats](), [behaviour](), [brain](), [can](), [code](), [connectivity](), [context](), [continuous](), [cs-ai](), [cs-lg](), [data](), [datasets](), [developing](), [diagnosis](), [domain](), [easier](), [foundation](), [networks](), [prediction](), [q-bio-nc]() ### AI overview BrainPFN predicts behavioral scores from functional MRI using a single forward pass over a labeled context, avoiding separate per-score or per-cohort fitting. It is meta-trained using a domain-specific prior built from about 7,000 real functional connectivity matrices and one billion synthetic training samples. The abstract reports that across four independent cohorts it outperformed Kernel Ridge Regression (KRR) in small-sample prediction and led recent brain foundation models trained at four to nine times larger scale. Its training pool is smaller and publicly released with code and checkpoints; these results are limited to the reported cohorts and comparisons. ### Source excerpt arXiv:2611.06934v1 Announce Type: cross Abstract: The standard approach to predicting behavioral phenotypes from functional MRI is represented by Kernel Ridge Regression (KRR) applied on functional connectivity (FC) matrices. However, this approach is limited: separate models must be fit and tuned for every score and cohort, they do not transfer across datasets, and accuracy degrades sharply when only a handful of subjects are available, as is typical of clinical or cognitive studies. Despite an increasing interest in developing brain foundation models, none has consistently outperformed KRR on behaviour prediction; most reported gains come from easier targets such as age, sex, or diagnosis rather than continuous behavioural scores. We introduce BrainPFN, a Prior Data Fitted Network for brain behaviour regression. BrainPFN is meta-trained on a domain-specific prior built from ~7k real FC matrices, from which we sample one billion synthetic training samples. At inference, a labelled context is simply passed through the network in a single forward pass, with no per-score or per-cohort refitting required. Across four independent cohorts, BrainPFN beats KRR on small sample prediction, and leads recent foundation models pre-trained at four to nine times larger scale. Our meta-training pool, aggregated from open repositories, is a fraction of the size of the corpora behind competing models, most of which sit behind restrictive data use agreements such as UK Biobank. We release this pool publicly alongside code and checkpoints, showing that a modest, openly available pool can match or beat models trained at Biobank scale. ## Capturing Membrane Dynamics and Spike Timing of Human Neurons at Scale using Neural Operators DevFeed: [Capturing Membrane Dynamics and Spike Timing of Human Neurons at Scale using Neural Operators]() Original publisher: [Read original article]() Author: Luca Ghafourpour, Cindy Zhao, Valentin Duruisseaux, Bahareh Tolooshams, Philip H. Wong, Costas A. Anastassiou, Anima Anandkumar Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Neural Network](), [Artificial Intelligence]() Tags: [brain](), [cells](), [cs-lg](), [distributions](), [experimental](), [features](), [framework](), [generation](), [input](), [learning](), [major](), [membrane](), [metrics](), [model](), [modeling](), [models](), [multiple](), [neural](), [operators](), [q-bio-nc](), [scale](), [timing]() ### AI overview The framework scales neuron modeling by learning a shared neural operator across modeled neurons of the same cell type, conditioned on biologically interpretable electrophysiological features, rather than building a separate surrogate for each neuron. Extending the single-neuron NOBLE setting to heterogeneous cells requires richer conditioning and a training objective that preserves diverse spiking dynamics. The approach generates ensembles of somatic voltage responses while retaining variation across cells and model configurations. The abstract reports reproduction of subthreshold dynamics, action-potential waveforms, firing-rate responses and spike timing across major cortical inhibitory neuron types, with predicted feature distributions agreeing with detailed simulator outputs and human recordings. Evidence is limited to the supplied abstract. ### Source excerpt arXiv:2611.06933v1 Announce Type: cross Abstract: Characterizing the properties of neuronal cell types is central to understanding brain function. Yet, biophysically realistic neuron models remain difficult to scale to large populations while preserving the richness of cellular dynamics and within-cell-type variability. We present a cell-type-specific neural-operator framework for scalable neuronal population modeling. Rather than training separate surrogate models for individual neurons, the proposed framework learns a shared operator across multiple modeled neurons from the same cell type, conditioned on biologically interpretable electrophysiological features. Relative to the original single-neuron \texttt{NOBLE} setting, learning across heterogeneous neurons requires a richer electrophysiological conditioning representation and a training objective designed to preserve diverse spiking dynamics. This enables efficient generation of large ensembles of somatic voltage responses while preserving variability across cells and model configurations. We evaluate the approach using membrane dynamics metrics, action potential waveform metrics, spike timing accuracy, input-output firing relationships, and electrophysiological feature distributions. Our framework reproduces subthreshold membrane dynamics, spike waveforms, and firing-rate responses across major cortical inhibitory neuron types, while accurately preserving spike timing despite the sensitivity of threshold-crossing dynamics. The predicted electrophysiological feature distributions show strong agreement with detailed simulator outputs and human experimental recordings, including high explained variability in properties such as inter-spike intervals, firing frequency, action-potential shape, and post-spike recovery. ## Oscillatory Phase Coding Emerges in Adaptive RNNs for Flexible Working Memory DevFeed: [Oscillatory Phase Coding Emerges in Adaptive RNNs for Flexible Working Memory]() Original publisher: [Read original article]() Author: Kehan Lu, Tianhao Chu, Aiersi Tuerhong, Yiwen Sun, Huan Luo, Fang Fang, Si Wu Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Artificial Intelligence]() Tags: [ablation](), [abstract](), [adaptive](), [amplitude](), [analysis](), [arxiv](), [cells](), [circuit](), [coding](), [connectivity](), [content](), [cs-ai](), [delay](), [develop](), [elements](), [encoding](), [flexibility](), [flexible](), [found](), [functions](), [memory](), [q-bio-nc]() ### AI overview In trained recurrent neural networks performing sequence working-memory tasks with variable delays, oscillatory phase coding emerged spontaneously: oscillation amplitude represented item identity, while phase represented serial order. Ablations identified adaptation in neural dynamics, variable retention delays, and flexible readout as necessary elements. The networks developed distinct content and time cells, with phase-organized connections among time cells and structured connectivity within digit-selective content groups. A proposed circuit model reproduced the coding through interactions between content units and an adaptive time ring. These findings concern trained RNNs and a proposed model; the excerpt does not establish that the same mechanisms operate in biological neural systems. ### Source excerpt arXiv:2611.06932v1 Announce Type: cross Abstract: Oscillatory phase coding (OPC) is widely observed in neural systems and has been implicated in cognitive functions such as memory and sequence processing. However, its computational role and neural substrates remain largely unclear. Here, we studied OPC by training recurrent neural networks (RNNs) with adaptation to perform sequence working memory tasks with variable delays. After training, we found that OPC emerges in the RNN spontaneously: the memorized items recur at distinct phases of a shared rhythm, with oscillation amplitude encoding item identity and phase encoding serial order. Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information. Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately. Recurrent connectivity analysis revealed phase-organized connections among time cells and structured connections within digit-selective content-cell groups. Inspired by these findings, we proposed a neural circuit model which reproduces OPC through interactions between content units and an adaptive time ring. Together, this study reveals how OPC arises from network structures and task demands, offers insights into the implementation and computational mechanisms of OPC in biological systems, and inspires us to develop computational models for human-like cognitive functions. ## Do sEMG Foundation Models Improve Cross-Subject Generalisation? DevFeed: [Do sEMG Foundation Models Improve Cross-Subject Generalisation?]() Original publisher: [Read original article]() Author: William Raftery, Dario Farina Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [AI Foundation Models](), [generalization in machine learning](), [Fine-tuning](), [machine learning overfitting]() Tags: [classification](), [computer-science](), [cs-lg](), [datasets](), [evaluation](), [evaluations](), [fine-tuning](), [foundation-model](), [foundation-models](), [machine-learning](), [ml](), [overfitting](), [performance](), [pretraining](), [q-bio-nc](), [recording]() ### AI overview The evaluation of three sEMG foundation models across three gesture-classification datasets finds that pretraining improves results when training and test samples come from the same recording session, but that benefit largely disappears in inter-subject evaluation. The study uses three fine-tuning strategies to isolate pretraining's contribution. On datasets with few subjects, pretraining provides little or no benefit, and classical machine-learning baselines match or outperform the foundation models. ### Source excerpt arXiv:2611.08199v1 Announce Type: new Abstract: sEMG Foundation Models aim to learn general purpose representations over diverse subjects and recording conditions that boost performance on a variety of downstream tasks. However, existing evaluations make it difficult to discern whether pretraining alone is sufficient to combat the cross-subject overfitting which makes machine learning on sEMG so challenging. We conduct a systematic evaluation of three sEMG foundation models across three gesture classification datasets under both intra-session and inter-subject splits, using three fine-tuning strategies designed to isolate the contribution of pretraining. We find that pretraining boosts performance when train and test samples are drawn from the same recording session, but this benefit largely disappears under inter-subject evaluation. In low subject count datasets, pretraining provides little to no benefit and classical ML baselines match or exceed foundation model performance. ## Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information DevFeed: [Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information]() Original publisher: [Read original article]() Author: Linqian Wu, Diaoxuan He Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Frontier AI](), [benchmark overfitting machine learning]() Tags: [ai](), [computer-science](), [consciousness](), [cs-ai](), [cs-lg](), [experiments](), [integrated-information-theory](), [interpretability](), [networks](), [q-bio-nc](), [recurrent]() ### AI overview In stochastic networks of four and five units, integrated information φ can be hard to infer from behavior alone or from partial recordings, yet some claims can be certified from finite experiments. The study reports recurrent and feed-forward networks with identical input-output laws at their response delay but φ values of 0.28 and 0 bits; internal recordings distinguish them within 4-53 trials. Omitted units can impose precision floors of 0.30-1.02 bits regardless of sample size, and some parity networks conceal hidden-unit fidelities behind a product symmetry that one intervention breaks. φ is also discontinuous at many partition-selection edges: tiny mechanism changes altered it by 25-100% at 99% of observed edges, while about 1% of surrogate models fitting interventional data nearly perfectly misreported it by over 20%. The authors show that one-sided lower bounds remain certifiable across such edges, including φ > 1.004 bits over a 64-dimensional mechanism region using about 10⁷ controlled transitions with false-certificate probability at most α. ### Source excerpt arXiv:2611.08469v1 Announce Type: cross Abstract: Interpretability, AI auditing and assessments of machine consciousness increasingly rest on mechanism-level quantities: functionals of a system's causal transition law that describe how it computes, not what it outputs. Experiments see that law only through finite, partial and noisy interventions. We ask when such a quantity is identifiable from an experiment's access, continuous in the mechanism and certifiable from finite data, and answer exactly for a deliberately demanding case - the integrated information $\varphi$ of integrated information theory - in stochastic networks of four and five units with known ground truth. Behavior can be silent: a recurrent and a feed-forward network with identical input-output laws at their response delay carry 0.28 and 0 bits, yet internal recordings separate them within 4-53 trials. Partial access can be just as blind: in our examples, units left out of a one-step recording impose precision floors of 0.30-1.02 bits that no sample size lowers, and in all-but-self parity networks, recordings of any length see several hidden units only through the product of their fidelities - a symmetry that one intervention broke in exact checks. The measure has sharp edges: because $\varphi$ selects a partition, and the cause and effect it specifies, by one criterion and reports another, infinitesimal changes of mechanism that cross an edge change it by 25-100% at 99% of the edges we found in random networks, and about 1% of surrogate models that fit interventional data almost perfectly misreport it by more than 20%. One-sided claims survive: the same data certify a lower bound on both sides of an edge, and we prove $\varphi>1.004$ bits throughout a 64-dimensional region of mechanisms and certify it from about $10^7$ controlled transitions with false-certificate probability at most $\alpha$. Measurability is thus a design question with computable answers. ## Effective pruning of task-trained recurrent neural networks using noisy fluctuations and connection rescaling DevFeed: [Effective pruning of task-trained recurrent neural networks using noisy fluctuations and connection rescaling]() Original publisher: [Read original article]() Author: Sanjith Senthil, Rishidev Chaudhuri Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Post-training optimization](), [Neural Network]() Tags: [cs-lg](), [cs-ne](), [networks](), [neural](), [neural-networks](), [performance](), [pruning](), [q-bio-nc](), [recurrent](), [recurrent-neural-networks]() ### AI overview Noise-prune preserves task performance in task-trained recurrent neural networks by sampling which connections to retain based on noisy activity fluctuations, then strengthening survivors to maintain average connection strength. The abstract reports that this approach outperforms magnitude-only pruning and exceeds a non-local strategy using second-order information. Sampling and rescaling are both essential to its performance and can augment other pruning strategies. The best empirical rescaling was lower than the level predicted by the original theory, prompting a revised version that outperformed the earlier one. Evidence here is limited to the supplied abstract. ### Source excerpt arXiv:2608.05464v2 Announce Type: replace-cross Abstract: Pruning network connections is key to brain function but, despite its importance, there exist few biologically-plausible pruning rules with demonstrated good performance. We evaluate noise-prune, a recently introduced unsupervised local pruning rule for recurrent networks that uses noisy fluctuations to determine the importance of connections. Noise-prune has previously only been empirically tested on random networks without a specific computational function. We show that noise-prune preserves task-performance in task-trained recurrent neural networks, greatly outperforming a strategy that only uses connection magnitude and exceeding a non-local strategy that uses second-order information. Rather than deterministically removing connections that fall below a threshold, noise-prune samples connections to preserve based on their importance and strengthens retained connections to preserve average strength. We show that this sampling and rescaling is essential to good performance, and can augment other pruning strategies. We also find that the optimal empirical degree of rescaling is lower than predicted by the original theoretical argument, and thus develop a new version of noise-prune that outperforms the previous version. Our results demonstrate that a biologically-motivated pruning strategy based on local activity statistics can effectively preserve task performance in functional recurrent networks by combining activity-dependent connection selection with compensatory strengthening of surviving connections. ## Emergent Generalization by Representation Learning in Artificial Neural Networks DevFeed: [Emergent Generalization by Representation Learning in Artificial Neural Networks]() Original publisher: [Read original article]() Author: Hardik Rajpal, Dan Goodman Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Representation learning](), [Neural Network](), [Machine learning](), [Artificial Intelligence](), [Embeddings]() Tags: [artificial-neural-networks](), [complexity](), [cs-it](), [cs-lg](), [cs-ne](), [distributed](), [distribution](), [functional](), [information](), [interpretability](), [learn](), [learning](), [magnitude](), [math-it](), [network](), [neural-networks](), [neuron](), [neuroscience](), [nlin-cd](), [performance](), [prediction](), [q-bio-nc](), [recurrent](), [reduction](), [structure](), [time]() ### AI overview In a time-series prediction task, an explicit information bottleneck that forces a recurrent neural network to learn a low-dimensional representation was necessary for rotational and out-of-distribution generalization. A measure of causal emergence followed a non-monotonic path--decreasing, reaching a minimum, then rising--even as prediction loss fell steadily; its trajectory scaled with task complexity, and the amount of emergent structure predicted generalization performance. Rat hippocampal CA1 and medial prefrontal activity during an alternating-maze task showed analogous dynamics that tracked behavior. The findings suggest compact, distributed representations can support generalization, though the supplied abstract does not establish the mechanism beyond these experiments. ### Source excerpt arXiv:2607.10430v2 Announce Type: replace-cross Abstract: Dimensionality reduction has proven powerful for identifying neural manifolds, which are low-dimensional structures underlying high-dimensional neural activity. These low-dimensional representations have improved the interpretability of population-level coding. Yet whether such low-dimensional representations are biologically relevant and confer functional advantages in learning systems, or merely reflect neuron-level activity, remains contested in neuroscience. We show that an explicit information bottleneck forcing a recurrent neural network to learn a low-dimensional representation is necessary for rotational and out-of-distribution generalisation in a time-series prediction task. Using information-theoretic measures of causal emergence, we characterise the dynamics of this representation across the memorisation-to-generalisation transition, finding a non-monotonic trajectory which shows an initial decrease, a minimum, and a subsequent rise to a maximum, even as prediction loss falls monotonically. This trajectory scales with task complexity, and the magnitude of emergent structure reliably predicts generalisation performance. Analysis of hippocampal CA1 and medial prefrontal activity in rats learning an alternating maze task reveals analogous non-monotonic emergence dynamics that track behavioural performance. Together, these findings indicate that the ability of neural networks to learn compact, distributed and emergent representations confers a functional advantage for generalisation, supporting a causal role for learned representations in cognition. ## Time-Division Multiplexing: The Operational Rhythm and Temporal Organization Layer of Digital Life DevFeed: [Time-Division Multiplexing: The Operational Rhythm and Temporal Organization Layer of Digital Life]() Original publisher: [Read original article]() Author: Haibin Tong, Luming Guo, Nana Cui, Jiang Yu Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [scheduling]() Tags: [clocks](), [consciousness](), [coupling](), [cs-ai](), [experiments](), [exploration](), [layer](), [multiplexing](), [noise](), [prototype](), [q-bio-nc](), [qwen3](), [scheduling](), [temporal]() ### AI overview The paper proposes a temporal scheduling layer for continual agents that coordinates urgent online work with offline replay and durable updates. It uses phase-tagged micro-epochs, an observable duty cycle, separate content and resource controls, and three clocks to distinguish exploring candidates from committing changes. The authors describe busy/idle execution contracts and a projected duty-cycle update, with deterministic descent guarantees and an additional noise term needed for a stochastic guarantee. Five synthetic experiments probe descent, objective alignment, gain asymmetry, resource coupling, and lost replay logs. In one constructed urgent, high-load subset, decoupled scheduling achieved 26.6% success versus 11.1%; the best tested gain ratio in the gain/noise model was 1.55. A frozen-backbone Qwen3-8B prototype tests orchestration only: its proxy outcomes are not task-benchmark evidence. The results support a testable design under specified conditions, not claims of consciousness or general-purpose superiority. ### Source excerpt arXiv:2611.06603v1 Announce Type: new Abstract: Continual agents must handle urgent interaction while retaining alternatives, replaying errors, and committing safe long-term changes. A previous digital-life framework distinguishes online (busy) and offline (idle) channels, but leaves open whether these are separate programs or phases of a common process and how their operations should be ordered. Motivated by recent observations of rapidly alternating perceptual representations in primate cortex, we develop a temporal organization layer with phase-tagged micro-epochs, an observable duty cycle, independent content and resource controls, and three nested clocks separating candidate exploration from durable commitment. We give a projected duty-cycle update, state its deterministic descent guarantee and the additional noise term required for a stochastic guarantee, and specify coupled busy/idle execution contracts. Five synthetic experiments (64 seeds each) examine descent and noise, objective misalignment, gain asymmetry, resource coupling, and dropped replay logs. In the gain/noise model the best tested gain ratio is 1.55; decoupled scheduling yields 26.6% versus 11.1% success on a constructed urgent/high-load subset. A frozen-backbone Qwen3-8B prototype further checks orchestration, but its proxy outcomes are not task-benchmark evidence. The results establish a testable scheduling design and identify its conditions of validity, not machine consciousness or general-purpose agent superiority. ## Mind Control of LLMs via Closed-Loop Stimulation: Out-of-Distribution Robustness DevFeed: [Mind Control of LLMs via Closed-Loop Stimulation: Out-of-Distribution Robustness]() Original publisher: [Read original article]() Author: Kasra Fallah, Rudramani Singha, Hannah Ghanei, Paul Sajda, James Anderson, Erfan Zabeh Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Large Language Model](), [LLM evaluation / benchmarking]() Tags: [control](), [cs-ai](), [cs-sy](), [distribution](), [eess-sy](), [feedback](), [llms](), [q-bio-nc](), [robustness](), [steering]() ### AI overview The work frames LLM steering as state-dependent feedback control: interventions respond to the network's internal state rather than applying a fixed perturbation. Its H∞-based controller is designed to limit sensitivity to disturbances and mismatches between tractable dynamical models and nonlinear transformers. The abstract reports improved performance over existing baselines across multiple steering benchmarks, with especially pronounced gains under out-of-distribution conditions and varying length complexities. It also introduces robust steerability (S_rob), a proposed pre-deployment measure of tolerance to dynamics mismatch that varies by task and systematically with model scale. The supplied text is an abstract and does not provide implementation details or quantitative results. ### Source excerpt arXiv:2611.06418v1 Announce Type: cross Abstract: Brains and large language models need not be considered equivalent systems to reveal an interpretable commonality: both exhibit high-dimensional, nonlinear dynamics whose behavior can only be modeled imperfectly. This creates a shared challenge whenever we seek to intervene in their dynamics. A controller may either rely on a model identified around a particular operating regime and risk failure as the system moves away from that regime, or explicitly account for uncertainty and a changing internal state. Neurostimulation has confronted an analogous problem for decades. The limitations of fixed, open-loop stimulation have motivated closed-loop and state-dependent approaches that condition intervention on the ongoing state of the neural system. Inspired by this principle, we formulate LLM steering as a feedback-control problem in which the network's internal state, rather than a fixed intervention alone, determines how the system is perturbed. We further account for the mismatch between tractable dynamical models and the nonlinear transformer they approximate through robust control. As one principled realization of this framework, we develop an $H_\infty$-based steering controller that limits sensitivity to disturbances and linearization mismatches. Across multiple steering benchmarks, the controller improves performance over existing baselines, with particularly pronounced gains under out-of-distribution conditions and varying length complexities. Beyond controller performance, we introduce robust steerability, $S_{\mathrm{rob}}$, a pre-deployment measure of tolerance to dynamics mismatch; across model-task pairs, it exhibits task-specific structure and varies systematically with model scale. Beyond robust LLM steering, our measure asks not merely whether control is possible, but how much uncertainty it can survive; a principle that may span artificial and biological neural systems. ## Do agents need to sleep? DevFeed: [Do agents need to sleep?]() Original publisher: [Read original article]() Author: Ido Haber, Giulio Tononi Published: 2026-11-05T05:00:00Z Content type: opinion Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [long-running](), [AI Context Management](), [llamaindex]() Tags: [agents](), [capacity](), [continuous](), [cs-ai](), [digital](), [engineers](), [environment](), [errors](), [interaction](), [learning](), [llm](), [memory](), [plasticity](), [precision](), [prediction](), [problems](), [q-bio-nc](), [retrieval](), [semantic](), [semantic-drift](), [sessions](), [sleep]() ### AI overview The paper argues that persistent LLM agents face memory problems associated with unchecked plasticity: semantic drift, poorer retrieval, stored errors, and capacity saturation. It maps the three phases of the selectionist "ringing, resonance and renormalization" account of sleep to operations feasible in digital memory architectures. From this analogy, it predicts that selectionist memory consolidation will outperform instructionist alternatives as memory stores grow large. This is presented as a falsifiable prediction; the supplied abstract does not report comparative test results. ### Source excerpt arXiv:2611.02216v2 Announce Type: cross Abstract: LLM-based agents, which increasingly operate persistently across sessions and accumulate memory through interaction, already exhibit the familiar symptoms of unchecked plasticity: semantic drift, degraded retrieval precision, propagation of stored errors, and saturation of memory capacity. Drawing on the selectionist "ringing, resonance and renormalization" (RRR) account of sleep, we propose that each of its three phases maps onto an operation feasible in digital memory architectures. Engineers building stateful agents, meanwhile, are independently rediscovering the problems sleep evolved to solve: what to forget, and when to do the forgetting. From this we derive the falsifiable prediction that selectionist consolidation will outperform instructionist alternatives as memory stores grow large. The core claim concerns the nature of continuous learning itself: plasticity has a price, and for any system that learns continuously under finite capacity while coupled to its environment, that price is sleep. ## When Model Complexity Matters in Cross-Subject Naturalistic fMRI Prediction DevFeed: [When Model Complexity Matters in Cross-Subject Naturalistic fMRI Prediction]() Original publisher: [Read original article]() Author: Junhao Zhao, Jacob Kang, Colin Kurniawan, David Simberg, Nan Xu Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Machine learning](), [Neural Network](), [Representation learning]() Tags: [blood](), [capacity](), [complexity](), [consistency](), [cs-ai](), [features](), [fixed](), [history](), [linear](), [model](), [pixels](), [prediction](), [protocol](), [q-bio-nc](), [reference](), [simulations](), [stimulus](), [temporal]() ### AI overview Across three naturalistic fMRI datasets, pretrained visual features generally predict held-out subjects' BOLD responses better than raw pixels, and the advantage grows over longer recursive rollouts for linear autoregression. Measured BOLD history adds substantial predictive value, though its gains usually diminish as rollouts lengthen. Nonlinear stimulus-only predictors can outperform linear readouts, while linear autoregression is a strong baseline when recent BOLD history is available. The study evaluates models offline on held-out subjects viewing largely familiar stimuli, without refitting; its results support comparing stimulus representations and downstream model complexity separately under matched conditions. Cross-subject response consistency tracks prediction accuracy, and less consistent regions tend to benefit more from measured history. Simulations suggest slower hemodynamic responses can help sustain temporal prediction correlations over longer rollouts. ### Source excerpt arXiv:2611.04041v1 Announce Type: cross Abstract: Increasingly complex neural models predict blood-oxygen-level-dependent (BOLD) responses to naturalistic stimuli, but the predictive value of this added complexity remains unclear. We ask when pretrained stimulus representations help and how much downstream complexity is needed, with and without recent measured BOLD history. We compare linear diagnostic references and nonlinear predictors across HCP movie fMRI (154 subjects), BOLD Moments (10 subjects), and Algonauts 2025 (4 subjects). Under a shared offline protocol, models are fitted on training subjects and evaluated on held-out subjects viewing largely familiar stimuli, without refitting. Pretrained visual features generally improve prediction over raw pixels; for linear autoregression, this advantage becomes more pronounced over longer rollouts (recursive predictions without additional measured BOLD). Measured BOLD history provides substantial gains that generally diminish {as rollouts lengthen}. Nonlinear stimulus-only predictors can improve on linear readouts, whereas linear autoregression provides a strong history-conditioned baseline. Controls examine context length, regularization, fitting procedures, and model capacity. Regional response consistency across subjects is associated with held-out prediction accuracy, while less consistent regions tend to gain more from measured history. In paired simulations with fixed linear neural dynamics, slower hemodynamic responses generally sustain high temporal prediction correlations over longer rollouts, illustrating how the observation process can contribute to predictive persistence. Together, these findings support evaluating stimulus representations and downstream complexity separately, using history-conditioned linear prediction as a diagnostic reference under matched inputs and measurement availability. ## Curriculum unlearning in humans and neural networks DevFeed: [Curriculum unlearning in humans and neural networks]() Original publisher: [Read original article]() Author: Jirko Rubruck, Alice Zhang, Andrew Saxe, Christopher Summerfield Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Machine learning](), [Neural Network](), [Representation learning](), [Artificial Intelligence]() Tags: [bias](), [compression](), [cs-lg](), [machine-learning](), [neural-networks](), [q-bio-nc](), [simple](), [tasks]() ### AI overview The abstract reports that simple-to-complex curricula help humans and neural networks overcome a bias toward simple rules that underfit the data. In neural networks, this bias accompanies compression of internal representations; curricula are reported to support representations and policies with dimensionality suited to the task. The authors also report that learners pre-exposed to simple tasks, and those with more compressed initial representations, benefit more from curricula. They interpret this as curricula helping learners "unlearn" simple rules. ### Source excerpt arXiv:2611.04565v1 Announce Type: new Abstract: During learning, humans often benefit from curricula that teach simple rules before exceptions. However, we lack a computational explanation for this phenomenon. Here, using a canonical behavioural task and modelling with neural networks, we develop and test a theory for why simple-to-complex curricula succeed. We show that during learning, both humans and neural networks show simplicity biases, a tendency to learn simple rules that are underfit to the data. In neural networks this bias is accompanied by compression of neural representations. Simple-to-complex curricula help both types of learners overcome this compression, developing representations (and policy) with appropriate dimensionality for the task. We test and confirm a further prediction from this theory: learners that have been pre-exposed to simple tasks and with more compressed initial representations benefit more from curricula. These findings suggest that curricula help with ``unlearning'' of simple rules. ## S1CogBench: What System One Decision Models Can and Cannot Do on Cognitive Experiments DevFeed: [S1CogBench: What System One Decision Models Can and Cannot Do on Cognitive Experiments]() Original publisher: [Read original article]() Author: Dengzhe Hou, Kazunori D Yamada Published: 2026-11-03T05:00:00Z Content type: news Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Artificial Intelligence](), [Benchmark](), [llm-reasoning]() Tags: [benchmark](), [cs-ai](), [cs-cl](), [decision-models](), [experiments](), [feedback](), [language-models](), [memory](), [q-bio-nc](), [tasks]() ### AI overview S1CogBench evaluates whether fast System One decision models respond to evidence and changing task demands, beyond simply producing valid choices. Across 24 cognitive tasks, the authors evaluate five open models and Jev, keeping benchmark conditions and scoring distinct from disclosed adaptations. They report wide variation in reliability; strong task performance does not necessarily mean human-like predictions; and aggregate scores can conceal how decisions change with conditions. The reported speed advantage comes mainly from answering without generating text. The benchmark tests responses to evidence, feedback, memory load, and future opportunities, but the supplied abstract does not give model-specific results or detailed findings for individual tasks. ### Source excerpt arXiv:2611.00911v1 Announce Type: cross Abstract: Language models are no longer used only to generate text. Software also asks them for decisions, in routing, guardrails, and automated gates. System One decision models, also called typed decision models, such as Jev and Laya, return one fast decision over the legal options, leaving no text to inspect. However, a fast, well-formed decision does not show that the model used the evidence. Cognitive experiments can test this, because they vary task demands and measure how behavior changes, yet existing cognitive benchmarks evaluate language and multimodal models. We introduce S1CogBench to ask what these models can and cannot do on 24 cognitive tasks, evaluating five open models and Jev and keeping conditions and scoring apart from disclosed adaptations. We find that decision models vary widely in reliability, acting well does not imply predicting humans, aggregate scores hide how decisions respond to conditions, and speed comes mainly from answering without generation. To our knowledge, S1CogBench is the first benchmark of System One decision models on established cognitive experiments, and it gives a shared test of whether their decisions respond to evidence, feedback, memory load, and future opportunities before software relies on them. ## AVID: A 5T fMRI Dataset for Benchmarking Auditory-induced Visual Mental Imagery Decoding DevFeed: [AVID: A 5T fMRI Dataset for Benchmarking Auditory-induced Visual Mental Imagery Decoding]() Original publisher: [Read original article]() Author: Shiqi Shen, Shurui Li, Xilin Zhang, Yuanning Li Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Decoding](), [Medical imaging](), [Image processing](), [Image]() Tags: [benchmark](), [benchmarking](), [brain](), [cs-ai](), [cs-cv](), [dataset](), [decoding](), [evaluation](), [model-development](), [q-bio-nc]() ### AI overview AVID introduces a dedicated dataset for decoding auditory-induced visual mental imagery from fMRI, designed to reduce visual-prompt and short-term-memory confounds. It contains about 67 hours of 5.0T recordings from nine participants, split between complex scenes described in naturalistic Mandarin and simple scenes with explicitly controlled visual attributes. Its standardized evaluation separates source-image consistency from semantic alignment. Baseline reconstruction pipelines remain limited, though their strongest results retain recognizable scene-level meaning from spoken descriptions. The dataset and protocol provide a training-scale resource and benchmark for imagery-focused model development. ### Source excerpt arXiv:2611.02593v1 Announce Type: cross Abstract: Visual mental imagery (VMI) decoding uses brain activity to reconstruct internally generated visual scenes, providing a potential pathway for externalizing internal experiences and developing neuroprosthetic communication interfaces. However, progress is limited by the lack of dedicated, training-scale VMI datasets. Current VMI paradigms often rely on visual prompts or short-term memory images, introducing visual and mnemonic confounds that make it difficult to disentangle neural activity associated with endogenous image construction from signals driven by prior or concurrent visual input. Furthermore, existing VMI datasets offer limited descriptive control over individual visual attributes, making it difficult to systematically assess fine grained visual details. To address these gaps, we introduce AVID, the first dedicated training-scale 5.0 Tesla functional magnetic resonance imaging (fMRI) resource for auditory-induced VMI decoding. AVID comprises approximately 67 hours of dense recordings from nine participants. The dataset is structured into two complementary splits: a scene-level complex split using naturalistic Mandarin descriptions and a controlled simple split with explicitly specified visual attributes. We further formalize a standardized evaluation protocol that decouples source-image consistency from semantic alignment, allowing for a more nuanced assessment of imagery reconstruction. Baseline qualitative and quantitative results show that current reconstruction pipelines remain limited on AVID, although the most successful reconstructions preserve recognizable scene level semantics from spoken descriptions. By combining dedicated training-scale neural recordings and standardized evaluation resources, AVID provides both a data foundation for imagery-specific model development and a benchmark for evaluating VMI decoding with substantially reduced sensory contamination. ## Prior Action Control Causally Alters Language-Model Plasticity DevFeed: [Prior Action Control Causally Alters Language-Model Plasticity]() Original publisher: [Read original article]() Author: Marc-Andre Schulz Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [rlvr](), [Large Language Model](), [post-training](), [LLM evaluation / benchmarking]() Tags: [control](), [cs-lg](), [language-model](), [learning](), [model-evaluation](), [plasticity](), [q-bio-nc](), [reward](), [training]() ### AI overview In the reported experiments, language models whose training rewards depended on their own answers learned a later task faster than yoked copies receiving the same rewards regardless of their answers. The advantage appeared across starting models and also when paired models practiced identical, preselected answers in the same order with the same number of rewards, supporting an effect of reward control beyond differences in practice. Models trained without a reward signal learned about as fast as yoked models, suggesting the observed gap was mainly a benefit of control. The advantage could also be forecast from model weights before the new task, though conventional evaluation need not reveal it. The supplied text is an abstract and does not give detailed methods or effect sizes. ### Source excerpt arXiv:2611.02362v1 Announce Type: new Abstract: Two copies of one language model start from identical weights and receive the same rewards during training. For one copy the rewards depend on its own answers. The other copy is yoked: it receives the first copy's rewards no matter what it answers, a design taken from the learned-helplessness experiments of animal learning. We ask whether the copy whose rewards depended on its answers, the copy with control, later learns a new task faster, that is, whether control changes the model's plasticity. In Experiment 1, copies of Qwen3-0.6B generated their own answers in a two-turn game and afterwards learned a new task. The copies with control learned the new task faster in every pair, and the advantage held on different starting models as well as on the larger Qwen3-1.7B. Copies with no reward signal, trained only on the updates that every copy shared, learned about as fast as the yoked copies. The advantage is therefore mostly a benefit of control, not damage from lacking it. Control had also changed how the copies answered, so the twins practiced different answers. That practice could carry part of the advantage, and Experiment 2 removed the difference. Both copies of a pair trained on the same answers, fixed in advance, in the same order, and received the same number of rewards. Only one copy's rewards followed the answers, and that copy again learned a new task faster. The advantage of the copies with control over their yoked twins could be forecast before the new task began from each copy's weights. In sum, control over rewards during earlier training left models faster at learning a new task later on, and a conventional model evaluation need not show it. ## The (Butter)fly Dream: Linear Separability of World-Model Rollouts in a Connectome-Constrained Visual Model DevFeed: [The (Butter)fly Dream: Linear Separability of World-Model Rollouts in a Connectome-Constrained Visual Model]() Original publisher: [Read original article]() Author: Junichiro Niimi Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [World models](), [Computer vision](), [model architecture](), [few-shot](), [machine learning overfitting]() Tags: [about](), [abstract](), [all](), [arxiv](), [brain](), [butterfly](), [can](), [column](), [cross](), [cross-validation](), [cs-lg](), [defects](), [detection](), [dream](), [dreams](), [fly](), [linear](), [model](), [published](), [q-bio-nc](), [world](), [world-model]() ### AI overview A connectome-constrained model of the fly visual system, trained only to estimate optic flow, can distinguish generated visual sequences from real footage using a linear classifier on its cell-type response statistics. Across 50 trained models, all nine continuations from three world models and three temperatures were separable from MPI-Sintel sequences (AUC 0.85-0.99), including unseen sequences; the same networks before training performed substantially worse. This separability did not track pixel error. In graded tests that shuffled the order of real frames, one adjacent swap among 14 generated frames yielded AUC 0.73 for trained responses, versus 0.54 before training and 0.56 for pixel statistics. A test over the first k generated frames helped order world models when full-sequence AUCs saturated. These findings are limited to the described model, task, footage, and probes. ### Source excerpt arXiv:2611.01866v1 Announce Type: new Abstract: Zhuangzi dreamt he was a butterfly and, on waking, could not tell which was the dream. We ask a version of his question of a model of the fly brain: placed inside a world generated by a world model, can a fixed brain tell that the world is a dream? We treat it as a question about what a connectome-constrained model trained on one task makes available beyond that task: do its responses make the defects of a dream linearly decodable? The observer is a published model of the fly visual system trained only for optic-flow estimation; reality is the MPI-Sintel movie on the model's 721-column retina; dreams are continuations of five real frames by three small world models (a persistence random walk, an MDN-RNN and an RSSM) at three temperatures. The judge is a classifier two-sample test on temporal statistics of cell-type responses, with sequence-level cross-validation. In the trained observer's responses all nine dreams are separable from reality (AUC 0.85-0.99, mean over 50 trained models), in those of the same network before training much less so (0.58-0.75 for the learned world models; paired difference 0.20-0.37), equally on sequences the observer never saw, and in an order unrelated to pixel error. Because these AUCs are near ceiling, we add graded probes made of each window's own real frames with only their order broken. Swapping one adjacent pair of the 14 generated frames gives 0.73 on the trained responses against 0.54 before training and 0.56 for pixel statistics; the trained curve reaches 0.75 at 1.1 swapped pairs of seven, the pixel statistics at 4.7, and the untrained network not at all. A detection horizon, the test on the first k generated frames, orders the world models where the full-length AUC saturates. ## Adversarial robustness does not entail human-aligned categorization DevFeed: [Adversarial robustness does not entail human-aligned categorization]() Original publisher: [Read original article]() Author: Itay Reznik, Raz Lapid, Moshe Sipper, Tal Golan Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [ai alignment](), [Machine learning](), [Adversarial attacks](), [ImageNet](), [human feedback]() Tags: [adversarial](), [alignment](), [classification](), [classifiers](), [computer-vision](), [confidence](), [cs-cv](), [cs-lg](), [experimental](), [model](), [prediction](), [q-bio-nc](), [recognition](), [robustness](), [training](), [vision]() ### AI overview Adversarial robustness did not ensure that ImageNet classifiers categorized controversial images in line with people. Researchers generated 1,008 stimuli to provoke high-confidence disagreements among robust models and compared four classifiers with two-choice judgments from 60 participants. On stimuli designed to pit a model against another model, classifier predictions matched individual human choices only 48.6-51.3% of the time, while other participants' majority choices predicted them at 85.0-86.8%. For stimuli optimized on other model pairs, all four classifiers performed above chance but below human-majority predictions; 17 additional RobustBench models showed the same gap, with 67.3% average model-participant agreement versus 81.3% participant-participant agreement. Confidence-aware analysis found weak positive correspondence for the two convolutional models. The results suggest robustness and human alignment should be evaluated separately. ### Source excerpt arXiv:2611.00715v1 Announce Type: cross Abstract: Adversarial training can improve some forms of correspondence between artificial and biological vision, but whether robust classifiers acquire human-aligned classification behavior remains unclear. We generated 1,008 controversial stimuli by optimizing latent inputs to a pretrained GAN to elicit high-confidence disagreements between pairs of robust ImageNet classifiers. We compared four adversarially robust ImageNet classifiers - ResNet-101, ViT-B, ConvNeXt-L, and Swin-L - with two-alternative category judgments from 60 participants. We first evaluated each classifier on stimuli optimized to elicit disagreement with another model. Across classifiers, accuracy in predicting individual participants' single-trial category choices was near chance (48.6-51.3%), whereas majority judgments from other participants predicted responses with 85.0-86.8% accuracy. An analysis incorporating model prediction strength and participants' reported confidence found weak positive model-human correspondence for the two convolutional models. On stimuli optimized for disagreement between other model pairs, all four robust classifiers predicted human responses above chance but less accurately than predictions based on other participants' majority judgments; the same pattern held for 17 additional RobustBench models not used in stimulus generation. Across these 17 models, mean model-participant agreement was 67.3%, versus 81.3% mean participant-participant agreement on the same images. These results identify model-specific failures of human alignment and a model-human mismatch that persists across robust classifiers and substantially exceeds disagreement among human observers. Our findings show that adversarial robustness alone does not guarantee human-aligned categorization and suggest that robustness and human alignment should be evaluated as distinct properties of visual recognition systems. ## Geometric and algebraic inference in humans and neural networks DevFeed: [Geometric and algebraic inference in humans and neural networks]() Original publisher: [Read original article]() Author: Jacob Russin, Joonhwa Kim, Ellie Pavlick, Michael J. Frank Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [llm-reasoning](), [Neural Network](), [Large Language Model](), [Machine learning]() Tags: [attention](), [context](), [cs-ai](), [cs-cl](), [cs-lg](), [data](), [experiments](), [flexibility](), [inference](), [large-language-models-llms](), [machine-learning](), [networks](), [neural-networks](), [q-bio-nc](), [reasoning]() ### AI overview The abstract distinguishes two inference strategies: geometric inference builds representational structure from familiar relations, making later inference inexpensive; algebraic inference selectively reasons over new relations in context, at additional inference-time cost. In transitive inference, experiments report opposing signatures of these strategies in humans, neural networks trained from scratch, and large language models: geometric inference is predicted to dominate for familiar relations, while algebraic inference supports novel, context-provided ones. Neural-network interventions link the strategies to representational geometry and selective attention, respectively, and reportedly reproduce effects of analogous disruptions in humans. The authors say this framework also connects findings on content effects in human and LLM reasoning and data-distribution effects in machine learning. This overview reflects the supplied abstract. ### Source excerpt arXiv:2611.00387v1 Announce Type: cross Abstract: Humans possess a remarkable ability to draw inferences from limited data. While some inferences are made intuitively from past experience, others require online reasoning based on novel information given in context. Here, we formalize these strategies, which we term geometric and algebraic inference, and study their computational tradeoffs in humans and neural networks. The geometric strategy gradually shapes representational geometry to reflect underlying relational structure, making subsequent inference cheap. The algebraic strategy flexibly attends to and reasons over novel relations given in context, but costs additional computation at inference time. We identify opposing signatures of these two strategies in transitive inference, a domain widely studied across species, and characterize the conditions under which each is observed. Our framework predicts that the geometric strategy should be dominant when relations are familiar, whereas the algebraic should emerge to support de novo inferences drawn in context. Multiple experiments demonstrate these opposing signatures in humans, neural networks trained from scratch, and large language models (LLMs). Causal interventions in the neural networks reveal that the geometric and algebraic strategies depend on representational geometry and selective attention mechanisms, respectively, and reproduce the effects of analogous neurobiological disruptions in humans. Our framework unifies these effects with disparate phenomena, including content effects observed in human and LLM reasoning and data-distribution effects observed in machine learning. Our findings reconcile competing views of cognitive flexibility that emphasize either representational geometry or algebraic manipulation by showing that each can have unique strengths, and can characterize different aspects of both human and neural network behavior. ## Lattice-Based Pruning of Recurrent Neural Networks via Poset Modeling: A Corrected Theory and a Matched-Sparsity Evaluation DevFeed: [Lattice-Based Pruning of Recurrent Neural Networks via Poset Modeling: A Corrected Theory and a Matched-Sparsity Evaluation]() Original publisher: [Read original article]() Author: Rakesh Sengupta Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Model Optimizer](), [Neural Network](), [model architecture]() Tags: [abstract](), [all](), [arxiv](), [based](), [better](), [bias](), [cohen](), [column](), [components](), [costs](), [cs-lg](), [evaluation](), [lattice](), [modeling](), [networks](), [neural](), [neural-networks](), [q-bio-nc](), [recurrent](), [recurrent-neural-networks](), [sparsity](), [theory]() ### AI overview The corrected theory shows that retaining meet-irreducible ideals preserves every element of a finite poset, while the earlier critical-set rule selects only elements comparable to all others and finds none in a dense recurrent layer. The paper instead condenses thresholded dependency graphs into strongly connected components to define an order score and acyclic pruning rule; acyclicity also lets local stability decompose over modules in the studied Cohen-Grossberg networks and LSTMs. In matched-sparsity tests on a 128-unit LSTM trained on two sequential-MNIST formats, the earlier rules removed all recurrent weights or performed like random column pruning. The order score did not outperform degree, PageRank, or norm, and trained weights showed little directional bias. Enforcing acyclicity from initialization reduced performance by 2.1 and 10.5 percentage points at 50% and 90% sparsity, respectively. These results suggest that, in the networks tested, order structure must be imposed rather than learned. ### Source excerpt arXiv:2502.16525v2 Announce Type: replace-cross Abstract: Magnitude pruning of recurrent neural networks (RNNs) ignores how information flows between neurons. An earlier version proposed to model RNN dependencies as a partially ordered set (poset) and to retain the meet-irreducible elements of its dependency lattice. By Birkhoff's representation theorem, the meet-irreducible ideals of a finite poset are in bijection with its elements, so retaining them retains everything; the critical-set criterion used earlier selects precisely the elements comparable to all others, of which a dense recurrent layer has none. Condensing a thresholded dependency graph into strongly connected components yields a well-defined dependency order, a graded order score and an acyclic pruning rule; for Cohen-Grossberg networks and LSTMs, acyclic pruning makes local stability decompose over modules. For a 128-unit LSTM pruned at matched sparsity on row-wise and chunked sequential MNIST (three seeds, equal training budgets), the original rules remove every recurrent weight or equal random column pruning; the order score is no better than degree, PageRank or norm; trained weights show little directional bias; and imposing acyclicity from initialisation costs 2.1 and 10.5 percentage points at 50% and 90% sparsity. In the networks studied, order structure must be imposed rather than discovered. ## Latent Prediction Yields Cross-Subject Alignable Representations in Sensory-Evoked fMRI DevFeed: [Latent Prediction Yields Cross-Subject Alignable Representations in Sensory-Evoked fMRI]() Original publisher: [Read original article]() Author: Serwan Jassim, Victoria Bosch, Daniel Anthes, Elia Bruni, Tim Kietzmann Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Machine learning](), [Representation learning](), [Embeddings](), [Decoding](), [Medical imaging]() Tags: [abstract](), [alignment](), [approach](), [architecture](), [arxiv](), [brain](), [can](), [central](), [cross](), [cs-ai](), [cs-lg](), [data](), [dataset](), [datasets](), [decoding](), [domain](), [embedding](), [embeddings](), [fine](), [fine-tuning](), [prediction](), [q-bio-nc](), [yields]() ### AI overview The study adapts a joint embedding predictive architecture (JEPA) to learn participant-specific fMRI representations from brain data alone, without stimulus images or derived representations. On the Natural Scenes Dataset, the embeddings support brain captioning and image reconstruction at competitive performance with parameter-rich supervised pipelines that use image and language data. Independently trained participant encoders can also be aligned without images or shared stimuli; after alignment, a decoding head trained on one participant transfers to another without fine-tuning and retains most within-subject performance on both tasks. The authors present this as a step toward pooling fMRI data across participants and datasets. ### Source excerpt arXiv:2610.40280v1 Announce Type: cross Abstract: Decoding sensory-evoked functional magnetic resonance imaging (fMRI) data is a central analytical tool in cognitive neuroscience. Here we ask whether self-supervised learning (SSL), which has been scarcely explored in this domain, can yield representations that support strong downstream performance. To do so, we adapt a joint embedding predictive architecture (JEPA) to train participant-specific encoders on brain data alone. The encoders never see the stimulus images or any representation derived from them. Applied to the Natural Scenes Dataset (NSD), we evaluate our learned brain embeddings on the downstream tasks of brain captioning and image reconstruction, as well as cross-participant generalization. Our self-supervised approach achieves competitive performance with parameter-rich supervised pipelines that are trained end-to-end with access to image and language data. We further find that encoders trained independently for different participants can be aligned fully unsupervised, using neither images nor stimuli shared between them. After alignment, a neural decoding head trained on one participant can be applied to another without fine-tuning and retains most of its within-subject performance on image reconstruction and brain captioning. By removing the need for images and shared stimuli, our approach is a step towards pooling fMRI data across participants and, eventually, across datasets. ## Data Amount Heterogeneity Improves Performance in Federated Learning for Neuroimaging DevFeed: [Data Amount Heterogeneity Improves Performance in Federated Learning for Neuroimaging]() Original publisher: [Read original article]() Author: Paul Kurian, Tamoghna Chattopadhyay, Abhijith Shaji, Greg Ver Steeg, Paul M. Thompson, Jose Luis Ambite Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Federated Learning](), [Machine learning](), [Training AI Models]() Tags: [classification](), [convergence](), [cs-cv](), [cs-lg](), [data](), [deep-learning](), [encryption](), [enigma](), [federated-learning](), [homomorphic-encryption](), [imaging](), [performance](), [q-bio-nc]() ### AI overview Federated neuroimaging models performed better when data amounts varied across sites--one large site alongside smaller ones--than when sites held equal amounts, especially in low-data settings. The heterogeneous setup often converged faster and achieved higher final accuracy, approaching centralized training, while data remained at each site and encrypted model parameters were aggregated under homomorphic encryption. The authors say this arrangement makes better use of small-data sites, a common situation in large consortia such as ENIGMA. ### Source excerpt arXiv:2610.38172v1 Announce Type: new Abstract: Federated Learning (FL) allows for secure analysis of distributed data without data sharing. We and others have successfully applied FL to neuroimaging tasks, such as Alzheimer's disease classification from structural magnetic resonance imaging (MRI). FL trains a deep learning model locally at each site, encrypts the local model parameters and sends them to a federation controller that aggregates the local models under homomorphic encryption, the resulting community model is sent back to the sites for further training, and the process repeats for a number of rounds. FL often achieves learning performance (e.g., accuracy) close to centralized training, but no data leaves the sites and model parameters are protected through encryption and other measures. In this paper, we show that federations composed of sites with heterogeneous amounts of data (1 large site plus small sites) perform better than homogeneous (same data per site) federations, particularly in low data regimes, while reaching performance close to centralized training. Having a big site often results in faster convergence and higher final accuracy. Heterogeneous federations better exploit sites with small amounts of data, a common situation in large consortia such as ENIGMA. [Next page]()