Multilayer Perceptrons

Hybrid models that combine physics-based and data-driven components have shown strong potential for achieving accuracy and interpretability in control applications. While recent methods have made progress in incorporating physical consistency, challenges remain in scalability, robustness to noise, and control of model complexity. This paper proposes a Physics-Encoded Modular Hybrid Layer (PE-MHL) framework, in which a baseline physics-based model is incrementally refined through the addition of new sub-models, where each new component adds complexity while preserving what previous components have already learned. We establish a theoretical guarantee for this construction: with a least-squares initialization of each new sub-model, the training error is monotonically non-increasing in the number of sub-models and provably converges. Empirical evaluations on a nonlinear NARX benchmark and the Quanser Aero 2 platform demonstrate that PE-MHL outperforms equivalently sized monolithic networks in both accuracy and generalization, while also providing more stable training dynamics and better preservation of underlying data structures.

We introduce Humanoid-GPT, a GPT-style Transformer with causal attention trained on a billion-scale motion corpus for whole-body control. Unlike prior shallow MLP trackers constrained by scarce data and an agility-generalization trade-off, Humanoid-GPT is pre-trained on a 2B-frame retargeted corpus that unifies all major mocap datasets with large-scale in-house recordings. Scaling both data and model capacity yields a single generative Transformer that tracks highly dynamic behaviors while achieving unprecedented zero-shot generalization to unseen motions and control tasks. Extensive experiments and scaling analyses show that our model establishes a new performance frontier, demonstrating robust zero-shot generalization to unseen tasks while simultaneously tracking highly dynamic and complex motions.

3D representations are fundamental to scene rendering, understanding, and interaction. Recent approaches, such as 3D Gaussian Splatting and Neural Radiance Fields, achieve impressive photorealistic novel-view synthesis, but lack the ability to easily decompose scene elements into a few primitives, requiring additional segmentation or grouping for object-level manipulation. We present MLP-Splatting, a method that enables scene decomposition via a few expressive light-field primitives while providing photorealistic novel-view synthesis. MLP-Splatting models each primitive as an independent compact MLP with localized spatial support that predicts radiance and opacity. In contrast to low-level Gaussian primitives or a single global radiance field, our neural primitives provide greater expressive capacity while remaining spatially localized. Rendering is performed through efficient sparse volumetric compositing over ray-primitive interactions. Our primitives are supervised using RGB supervision alone, which yields primitives that represent local scene regions often corresponding to objects or object parts, enabling interactive object-level editing without segmentation masks by selecting a handful of primitives. Our method, augmented with optional semantic feature distillation, enables open-vocabulary scene interaction and open-set instant segmentation. Compared to state-of-the-art methods, we achieve substantially lower memory usage (1/15$\times$) and faster rendering (3$\times$), as we show in our experiments compared to semantic 3DGS methods. Project Page: https://shinjeongkim.com/mlp-splatting

Language models do not simply choose an answer at the output layer. In a 9,000-trajectory MMLU study across Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3, the score of the answer moves across depth in structured ways. We describe each trajectory with three quantities: the current answer margin, the next-layer change in that margin, and the distance from a decision flip. The main empirical picture is that correctness and stability are different: the largest group is unstable-correct, not stable-correct. A traced subset then asks what moves the margin. In stable-correct cases, the average attention scalar points in the correct direction, while the average MLP scalar does not; span deletion shows that removing answer-supporting text hurts the margin and removing distractor-like text helps it. The result is not a full circuit explanation. It is a reproducible way to see which answers are settled, which remain fragile, and which measured sources move them.

Large language models are able to compose skills in order to perform complex tasks, many of which might not have been seen during training. The details of how exactly this composition occurs remain elusive. In this paper, we study a mechanism for compositional generalization in transformers by considering a simple controlled setting involving variable assignment and modular addition. By partitioning our training data into disjoint sets, we observe that small transformers are able to generalize to previously unseen combinations of variables and numbers. Our mechanistic analysis shows that the same ``modular addition'' MLP module is used whether the inputs are given directly or indirectly through a separate variable assignment mechanism. We also analyze the training dynamics from an empirical lens, which reveals three phases of learning: first, modular addition is learned, then the structure required for variable assignment, and finally a refinement phase where the model generalizes to some hard sequences not seen in training. Finally, we provide a theoretical framework to explain how compositionality emerges from training dynamics. These results suggest that compositional generalization can be a natural consequence of the compositionality of internal mechanisms in~transformers.

A central aim of deep learning theory is to characterize how neural networks make predictions in the regime of simultaneously large model and training set size. Since the limits of diverging number of model parameters and dataset size do not commute it is not clear a priori what limits exist. In this work, we shed new light on these questions by studying Bayesian inference in deep non-linear MLPs in the regime where the number of training samples ($P$), the input dimension ($N_0$), the hidden layer width ($N$), and the number of hidden layers ($L$) can all be large. We build on the Neural Covariance SDE (Li et al., 2022) to analyze predictive posteriors in the regime where $LP/N\inΘ(1)$, playing the role of an effective network depth. Our framework covers both smooth and ReLU activation functions and applies to arbitrary temperature. We find to first order in $LP/N$ a simple criterion for which data generating processes benefit from depth in the sense that larger $LP/N$ increases the Bayesian model evidence. We also give a novel derivation of a prior result from the physics literature that at least to first order in $LP/N$, the Bayesian predictive posterior is remarkably simple and is simply equivalent to that of a data-dependent kernel method.

Tabular data in knowledge-rich domains often carries a latent prior in the form of Boolean implication relationships (BIRs) between pairs of features. We mine such relationships with a sparse-exception binomial test. The mined implications form a typed directed graph, equivalent to a propositional rule base of 2-literal clauses. We encode this graph as the connectivity of a layered neural network, called BIRDNet, in which each hidden unit corresponds to one mined rule and binds only to its two features. We show two consequences of this design: First, the architecture is sparse by construction: at most $2/d$ of the weights in each BIR layer are active, where $d$ is the input dimension. Second, the model is interpretable: every trained unit keeps a stable symbolic identity, so rules can be read off the network without surrogate models. Unlike most neurosymbolic models, BIRDNet does not consume an external rule base; its structural prior is mined from the data. We evaluate BIRDNet on six transcriptomic and proteomic benchmarks. Our results show that BIRDNet stays within 0.02 AUROC of the strongest dense baseline, at a small accuracy cost, while using up to $96\times$ fewer active parameters than an architecture-matched dense MLP. First-layer rules recover known biological signatures across multiple cancer subtypes and tissue types, including canonical amplicons, lineage-defining co-expression modules, and immune-infiltration markers. Data and code are available at: https://github.com/MAHI-Group/BIRDNet.

In-context learning \ -- performing tasks based on examples given in the prompt \ -- is an important capability that has emerged in large language models and has received significant attention in both theory and practice. Existing theoretical work often focuses on settings where the learning uses information purely from the prompt. However, many practical instances of in-context learning require the model to retrieve factual knowledge stored in the model's parameters, with the context serving to identify which knowledge is relevant. In this work, we study how in-context learning leverages factual knowledge recall. We formalize this behavior by introducing the \emph{in-context factual recall (IC-recall)} task, where a transformer is provided a context of (subject, answer) pairs generated from a hidden relation, along with a query subject, and must both infer this hidden relation and retrieve the corresponding answer. Factual knowledge is modeled by the transformer having access to a simple pre-constructed MLP associative memory storing (subject, relation, answer) triplets. We analyze the supervised fine-tuning dynamics of a one-layer transformer on IC-recall data and prove that the model successfully performs IC-recall by converging to a particular pairwise attention pattern. This fine-tuning stage requires a very small number of samples \ -- only polylogarithmic in the number of stored knowledge triplets. Experiments verify our theoretical predictions and show that the pairwise attention pattern emerges even when the MLP layer is pretrained instead of constructed.

The Mixture of Experts MoE architecture is highly promising for resource constrained on device deployments yet training these models from scratch incurs prohibitive costs Current methods attempt to alleviate this by upcycling dense models into MoEs however they often introduce parameter redundancy that degrades inference efficiency Alternatively standard layer pruning mitigates redundancy but inevitably compromises model accuracy To resolve this dilemma we propose Dense2MoE a novel framework that unifies pruning and upcycling through Layer Fusion UpCycling LF UC Guided by hardware Roofline theory Dense2MoE systematically overcomes the inference memory wall by pruning bandwidth heavy attention modules from redundant layers while repurposing their Multi Layer Perceptrons MLPs into MoE experts This structural innovation preserves the models core capabilities and strictly limits active parameters via selective token routing With a modest continual pre training budget Dense2MoE efficiently converts publicly available dense LLMs into on device ready MoE models Extensive experiments demonstrate that Dense2MoE significantly advances the Pareto frontier for on device inference latency versus model accuracy outperforming dense baselines state of the art compression and standard upcycling methods

Length generalization remains a persistent challenge for neural networks: recurrent models tend to suffer from positional biases, while transformers are constrained by fixed computational depth. Regular languages provide a frequently used testbed for evaluating length generalization, as label prediction can be checked for any sequence length. We propose MLP-LDRU, a type of Log-Depth Recurrent Unit, which captures a class of associativity-biased operators designed to approximate recurrence through parallel reduction. We evaluate MLP-LDRU on 21 regular-language tasks, consisting of standard benchmarks and new prefix languages, where it achieves 100% out-of-distribution accuracy on 18 tasks and at least 99.9% on the remaining 3 when increasing max training length, outperforming comparable recurrent and attention-based models. We further evaluate MLP-LDRU beyond regular languages on ListOps and NLP classification benchmarks, where it performs competitively.

We here study whether training biases can make hidden neurons specialize in minimal one-hidden-layer MLPs, and whether such specialization improves prototype-based reconstruction of the training dataset from the learned weights. We consider Gaussianactivation MLPs of width equal to dataset size and compare three structural losses that respectively encourage coverage of the training samples, separation between neuron-induced prototypes, and low overlap of hidden responses, against the standard fitting baseline. Experiments on uniformly sampled one-dimensional datasets show a stable pattern from N = 3 to N = 100 across 480 controlled runs. Coverage regularization gives the lowest mean reconstruction error at every tested size and raises the prototype-usage specialization ratio relative to the standard baseline, while separation has mixed effects and overlap penalties are systematically harmful. We show that the harm is not an optimization failure: overlap-active approaches fit the data as well as overlap-free ones but route the optimizer to a degenerate equilibrium in which prototype centers are pushed outside the convex hull of the training inputs. Coverage cannot reward this expulsion and acts as an attractor: separation admits it only at large temperature and overlap admits it at the nominal hyperparameter choice. A direct τ-sweep on the separation-only mask and a prototype-position visualization at N = 100 confirm the mechanism. The findings yield a simple design principle for prototype-recoverability-aware training: every repulsive structural loss must be compensated by a compatible attractor, or it will collapse the latent geometry it was meant to refine.

Multimodal Attributed Graph Learning (MAGL) integrates intrinsic node attributes with structural topology via graph aggregation. However, as pretrained encoders evolve into Large Foundation Models (LFMs), the landscape of MAGL fundamentally shifts: under high-confidence LFM priors, mandatory aggregation introduces topological noise that overwhelms discriminative signals, triggering a counter-intuitive performance inversion where sophisticated MAGL architectures underperform simple topology-agnostic MLPs. Through systematic empirical and theoretical analysis, we identify that this inversion stems from a fundamental aggregation dilemma characterized by two concurrent pathologies: (1) Representational Pathology (SNR Degradation) - mandatory aggregation dilutes robust intrinsic features with topological noise, causing the noise penalty to outweigh its collaborative benefit; and (2) Optimization Pathology (Gradient Starvation) - topological aggregation attenuates gradient flow, while a shared task loss causes dominant modalities to prematurely suppress weaker ones. To resolve this dilemma, we propose SUPRA (Shared-Unique Prior-Retaining Architecture), a decoupled dual-pathway paradigm. SUPRA processes modality-specific features through topology-agnostic MLPs while capturing structural synergy via a lightweight shared GNN, with auxiliary deep supervision counteracting gradient starvation. Extensive evaluations demonstrate that SUPRA achieves state-of-the-art performance while requiring 3.5x lower peak GPU memory and up to 4.4x faster training time than Multimodal Graph Transformers.

Finite-width fully connected neural networks with Gaussian-initialized weights deviate from their infinite-width Gaussian limit, exhibiting non-vanishing higher-order cumulants. We approximate these deviations, for a neural network evaluated in a finite number of inputs, using multidimensional Edgeworth expansions of arbitrary order $4m-1$, with $m\in\mathbb{N}$. Assuming that the corresponding Gaussian limit has an invertible covariance matrix and that the activation function is polynomially bounded, we establish a bound of order $n^{-m}$ on the total variation distance between the law of the true network output and its Edgeworth approximation, with matching lower bounds. As an application, we quantify the error in Bayesian posterior distributions when the prior is replaced by its Edgeworth expansion. Our results are more general and also apply to sequences of conditionally Gaussian vectors converging to a Gaussian vector with invertible covariance.

The maximum softmax probability (MSP) represents a default approach when evaluating uncertainty quantification for language model generation with structured output. Although cheap, it is often miscalibrated. Methods that probe the model's internal activations feed raw hidden states into opaque classifiers, reading activations as static snapshots and leaving implicit the layer-wise trajectory by which a representation is formed. Yet, similar endpoints can arise from very different paths, and how evidence accumulates, reinforces, or reverses across depth might reveal uncertainty that final probabilities obscure. We extract eleven scale-invariant geometric features, tracing the cumulative path of per-layer MLP updates, and feed them to a sparse linear probe. The probe outperforms MSP under selective abstention, with gains scaling with baseline miscalibration up to 21 AURC points. Because every feature has a closed-form geometric meaning, the probe's coefficients trace how and where along depth errors take shape -- which layers commit prematurely, which contradict the running state, where trajectories drift away from their endpoint.

A striking geometric disparity has long persisted in the practice of deep learning. While modern neural network architectures naturally exhibit rich symmetry and equivariance properties, popular optimizers such as Adam and its variants operate inherently coordinate-wise, rendering them unable to respect the equivariance structures of the parameter space. We address this disparity by introducing a symmetry-compatible principle for optimizer design: the gradient update rule should be equivariant under the symmetry group acting on the corresponding weight block. Following this principle, we first provide a unified perspective on bi-orthogonally equivariant updates for general matrix layers, as employed by stochastic spectral descent, Muon, Scion, and polar gradient methods. More importantly, by moving from orthogonal groups to permutation and shared-shift symmetries, we derive symmetry-compatible optimizers for parameter blocks whose symmetries differ from those of general matrix layers: embedding and LM head matrices, SwiGLU MLP projections, and MoE router matrices. These constructions include one-sided spectral, row-norm, hybrid row-norm/spectral, row-aware, column-aware, centered row-norm, and left-spectral updates. They yield an end-to-end layerwise optimizer stack in which each major matrix-valued parameter class is assigned an update whose equivariance matches its symmetry group. We corroborate this principle through pre-training experiments on dense and sparse MoE language models, including Qwen3-0.6B-style, Gemma 3 1B-style, OLMoE-1B-7B-style, and downsized gpt-oss architectures. Across these experiments, symmetry-compatible update rules consistently improve final validation loss, reduce expert load imbalance in sparse MoE models, and in several cases control final vocabulary-logit growth, improve router stability, and overall training stability over the corresponding AdamW updates.

Federated Learning (FL) has emerged as a promising paradigm for collaborative model training across distributed edge devices while preserving data privacy especially with the huge increase amount of data due to the adoption of technologies which contributes to the growing number of IoT devices. Storing this amount of data centrally is challenging due to issues like limited communication, privacy, and regulations. FL can be Centralized (CFL), Decentralized (DFL), and Semi-decentralized (SDFL). Choosing the right FL architecture depends on the application's needs. However, very few research studies have experimentally compared these three types of architectures to not only understand the respective strengths and limitations, but also trade-offs between different performance indicators. This paper overcome this lack of analysis, conducting experimental analyses using the Fedstellar simulator, MNIST dataset, and MLP classifier.

Model growth from a given checkpoint aims to accelerate training of a larger model, offering potential resource savings. Despite recent interest, warmstarting has seen limited practical adoption in large-scale training. We attribute this to two underexplored factors: (1) an overemphasis on preserving the smaller model's performance at initialization, which constrains operator design for new architectures, and (2) insufficient analysis of how growth interacts with hyperparameters and scaling behavior, compounded by inconsistent growth factors across the literature. We show that preserving the base model's initial post-growth performance is not necessary for strong final performance, and that simple, architecture-agnostic growth strategies can outperform more complex warmstarting operators. Crucially, we empirically identify an upper bound on the growth factor $g$ beyond which training from scratch is more efficient. We observe this across multiple ablation setups. Notably, this limit is also present, but unreported, in prior published results. Across our experiments on dense MLPs and dense language models, we find that a $2\times$ growth factor is the most reliable in yielding convergence speedups, with gains most pronounced under 20 tokens/parameter budgets and diminishing as budget increases. We fit scaling laws over these observations to provide predictive guidance for practitioners deciding when and how much to grow. Together, our analysis provides practical guidelines and empirical limits for model growth.

Many features in pretrained Transformers span multiple layers: they emerge through stages of inference, persist in the residual stream, or are built jointly by parallel MLPs. Crosscoders (namely, sparse dictionaries trained jointly across layers) aim to recover these cross-layer features in a single shared latent space. We show that standard crosscoders largely fail at this purpose. Although their decoder weight norms spread evenly across layers, a functional coherence metric we introduce reveals that each latent's activation is effectively driven by only one or two layers on average. While functionally coherent latents act as human-interpretable concept detectors (e.g., US states and cities), the layer-localized latents that crosscoders predominantly learn collapse onto surface-level patterns such as digit detectors. We trace this failure to two structural limitations: unconstrained cross-layer parameterization and unregularized cross-layer dependence. We address both by introducing fmxcoders, which (i) replace the encoder and decoder with low-rank tensor factorizations that draw every latent's per-layer weights from a shared cross-layer basis, and (ii) apply stochastic layer masking, a denoising regularizer along the layer axis that penalizes latents whose contribution collapses when a single layer is masked. Across GPT2-Small, Pythia-410M, Pythia-1.4B, and Gemma2-2B, fmxcoders lift mean probing F1 by 10-30 points, surpassing per-layer SAE baselines that standard crosscoders fail to reach, reduce reconstruction MSE by 25-50%, and roughly double mean functional coherence. An LLM-as-a-judge evaluation further shows that fmxcoders recover 3-13$\times$ more semantically coherent latents than standard crosscoders across all four base LLMs.

We introduce a use of the $M$-cover (or $M$-layer) transform for machine learning. The method replicates a model $M$ times, but instead of coupling the copies through parameter averaging or an explicit attractive force, as in replicated SGD or Elastic SGD, it rewires the contexts in which local learning messages are computed. Each local loss is evaluated on a routed model whose parameters are drawn from different copies according to permutations sampled from a structured mixing kernel $Q$. Training then uses the original local update rule, while the resulting learning messages are redistributed across the copies through these routed computational paths. Thus $Q$ defines a topology for message transport and controls the long-loop structure of the lifted factor graph. We formulate this construction for perceptrons, committee machines, and multilayer perceptrons, showing that the same principle applies from discrete models to differentiable neural networks. The resulting framework provides a mechanism for improving generalization through structured message sharing rather than replica collapse or parameter-space coupling.

Large language models fail at counting repeated tokens despite strong performance on broader reasoning benchmarks. These failures are commonly attributed to limitations in internal count tracking. We show this attribution is wrong. Linear probes on the residual stream decode the correct count with near-perfect accuracy at every post-embedding layer, across all model depths. This holds even at the exact layers where the wrong answer crystallizes while the model simultaneously outputs an incorrect count. Attention patterns show no evidence of collapse over repeated tokens and tokenization artifacts account for none of the failure. Instead, a format-triggered multi-layer perceptron (MLP) block overwrites the correctly-encoded count with a fixed wrong answer at roughly 88--93,% network depth. This prior fires for repeated word-tokens in space-separated list format and is absent for repeated digit-tokens. It is suppressed by comma-separated delimiters in larger models but persists in smaller ones. The finding holds across Llama-3.2 (1B and 3B) and Qwen2.5 (1.5B, 3B and 7B) at consistent relative depth. Counting failure is a failure of routing not of representation and the two require different interventions.

Large language models confidently produce outdated answers, and no existing method can detect them. We show this is not an engineering failure but a structural one: temporal drift, whether a stored fact has changed since training, is encoded as a direction in the residual stream geometrically orthogonal to both correctness and uncertainty. Any method operating on correctness or uncertainty signals is therefore blind to drift by construction. We verify this across six instruction-tuned models. A linear probe trained directly on drift labels achieves AUROC $0.83$--$0.95$; methods based on token entropy, semantic entropy, CCS, and SAPLMA all remain near chance ($0.49$--$0.57$). Five tests confirm the geometric orthogonality: weight cosines ($|\cos| \leq 0.14$), score correlations ($|r| \leq 0.20$), bidirectional null-space projection ($|Δ| \leq 0.008$), iterative null-space projection with $k{=}10$, and difference-of-means dissociation. Mechanistically, the MLP retrieval circuit produces identical dynamics for stale recall and confabulation ($r > 0.81$, six models), explaining why output confidence cannot separate them. A cross-cutoff experiment holds inputs constant and varies only the model: the probe fires on the model whose training predates the fact's transition and stays silent otherwise ($P(A{>}B) = 0.975$--$0.998$, twelve model pairs), confirming it reads model-internal knowledge state rather than input properties. Our code and datasets will be publicly released.

This paper presents a compact three-class sentiment analysis study for Indonesian social media text. The task is formulated with positive, negative, and neutral outputs derived from a fine-grained emotion dataset. The proposed practical baseline combines TF--IDF text features, three lightweight numeric metadata features, and a balanced multinomial Logistic Regression classifier. For comparison, the study also includes a neural baseline using a two-layer multilayer perceptron (MLP) over the same hybrid feature representation. The dataset originally contains 732 rows and 191 fine-grained emotion labels; after cleaning, deduplication, and label remapping, 707 samples remain with an imbalanced distribution of 459 positive, 188 negative, and 60 neutral instances. Experimental results show that the Logistic Regression deployment model reaches 0.8028 accuracy, 0.8003 weighted F1, and 0.7276 macro F1, while project documentation reports a higher-accuracy but non-production MLP baseline. These findings indicate that careful preprocessing, interpretable feature engineering, and class balancing remain competitive for small Indonesian sentiment datasets, whereas the neural baseline is better treated as a comparative experiment than as the default deployment model.

Multivariate time series forecasting remains a challenge due to the complexity of local temporal dynamics and global dependencies across multiple variables. In this paper, we propose \textbf{N}eighboring \textbf{P}atching \textbf{Mixer} (\textbf{NPMixer}), a hierarchical architecture featuring a Learnable Stationary Wavelet Transform that adaptively learns filter coefficients to decompose signals into trend and detail components in a data-dependent manner. Our framework introduces a Neighboring Mixer Block that captures local temporal dynamics through a series of hierarchical MLP layers operating on non-overlapping patches. Specifically, the mixer block utilizes MLPs to learn temporal patterns within and across these patches, expanding the receptive field to capture multi-scale dependencies. A Channel-Mixing Encoder is applied to high-frequency components to learn channel correlations while preserving the stability of the underlying global trend. Extensive experiments on seven benchmark datasets demonstrate that NPMixer consistently outperforms state-of-the-art models, achieving better performance in 20 out of 28 ($71.4\%$) evaluated experimental setups for MSE.

Final outputs hide when a checkpoint commits to its next-token prediction. We introduce the convergence gap, a model-diffing diagnostic that decodes each layer's next-token distribution and measures its distance to the model's own final distribution. Across six paired pretrained and instruction-tuned checkpoints in native prompting regimes, instruction-tuned checkpoints remain farther from their final predictions later into the stack. The effect persists under endpoint-matched raw and tuned readouts, endpoint-free same-history checks, and fixed-history template replay. Matched-prefix interventions identify late MLP windows as the largest tested leverage point: late IT grafts into PT hosts increase late KL by +0.34 nats, while PT-late swaps into IT hosts reduce it by -0.51 nats; matched random late perturbations give only +0.003 versus +0.327 for the true late graft. A preselected Gemma case study provides behavior-facing plausibility for the same late swap, without serving as a benchmark claim. These results identify a robust predictiondynamics signature of post-training: released instruction-following checkpoints tend to settle later, and late MLP computation is the strongest tested bidirectional handle on that delay under matched histories.

We show that, in a precise sense, a broad class of feedforward neural networks learn (have finite sample complexity) in the PAC model: every fixed finite feedforward architecture whose layers are definable in an o-minimal structure has finite sample complexity in the agnostic PAC setting, even with unbounded parameters. This covers standard fixed-size MLPs, CNNs, GNNs, and transformers with fixed sequence length, together with the operations and layers typically used in such architectures, including linear projections, residual connections, attention mechanisms, pooling layers, normalization layers, and admissible positional encodings. Hence, distribution-free learnability for modern non-recurrent architectures is not an exceptional property of particular activations or architecture-specific VC arguments, but a consequence of tame feedforward computation. Our results reposition finite-sample PAC learnability as a baseline rather than a differentiator: they shift the focus of architectural comparison toward inductive biases, symmetries and geometric priors, scalability, and optimization behaviour.

Sequence models face a fundamental tradeoff between memory capacity and computational efficiency. Transformers achieve expressive context modeling at quadratic cost, while linear attention and state-space models run in linear time by compressing context into a fixed-size hidden state, inherently limiting recall. Log-linear attention navigates this tradeoff by organizing memory across a Fenwick tree hierarchy, growing its hidden state logarithmically with sequence length at log-linear compute cost. However, its memory decay parameter λ is fixed and independent of the input, assigning uniform weights across all hierarchy levels regardless of the content, which introduces unnecessary rigidity. We propose learning λ directly from the input via a lightweight two-layer MLP, producing per-token, per-level decay that adapts to content rather than position. A softplus activation lets each Fenwick tree level scale independently, avoiding the inter-level competition that softmax introduces. This modification preserves log-linear complexity exactly and adds negligible parameter overhead. We evaluate on associative recall, selective copying, and language modeling, finding that input-dependent decay consistently outperforms the baseline, with the largest gains in long-range memory settings where baseline λ degrades or collapses entirely.

In grokking, a model first fits the training data while test accuracy remains low, and only later begins to generalize. We ask whether this transition can be localized from observed training trajectories before the test accuracy rises, and formulate grokking transition localization as a diagnostic problem with an explicit threshold/FPR/lead-time trade-off. Task-dependent observables are summarized as empirical distributions, mapped to Wasserstein/quantile coordinates, and analyzed by Hankel dynamic mode decomposition (DMD); the resulting reconstruction residual, together with spectrum and effective rank, forms the diagnostic output. On held-out modular-addition Transformer runs, the residual achieves AUROC $\approx$ 0.93 for grokking-vs-non-grokking discrimination at the run level; under a fixed sustained-threshold operating rule, true-positive alarms can precede onset, with lead time reported jointly with false-alarm rate and uncertainty intervals. Perturbation experiments show that, in the tested $wd=1$ pool, high-residual windows exhibit about $3\times$ larger short-horizon perturbation deviation than low-residual windows. In a same-data norm-window control, perturbation sensitivity aligns with the residual ordering rather than total-parameter-norm ordering, suggesting that the residual is not merely a total-norm proxy at the window level in the studied $wd=1$ dynamics. Norm signals remain strong run-level regime indicators, and log-probability performs best among the observables tested under the current protocol. We position the residual as a window-level monitoring and localization signal in the studied modular-arithmetic Transformer settings, not a universal early-warning predictor or an intervention rule.

By far the most common way to estimate an expected loss in machine learning is to draw samples, compute the loss on each one, and take the empirical average. However, sampling is not necessarily optimal. Given an MLP at initialization, we show how to estimate its expected output over Gaussian inputs without running samples through the network at all. Instead, we produce approximate representations of the distributions of activations at each layer, leveraging tools such as cumulants and Hermite expansions. We show both theoretically and empirically that for sufficiently wide networks, our estimator achieves a target mean squared error using substantially fewer FLOPs than Monte Carlo sampling. We find moreover that our methods perform particularly well at estimating the probabilities of rare events, and additionally demonstrate how they can be used for model training. Together, these findings suggest a path to producing models with a greatly reduced probability of catastrophic tail risks.

SwiGLU is the dominant MLP activation in modern large language models, yet its intermediate tensor materialization costs 9-37% of MLP execution time. We present two complementary CUTLASS-based SM90 kernels that fuse SwiGLU into GeMM at the tile level. Kernel-1 overlaps Swish computation on the Gate accumulator with Up-tile loading using the Pingpong warp-specialized schedule; Kernel-2 interleaves SwiGLU with tile stores via a custom Epilogue Visitor Tree. Evaluated on Qwen-2.5 models (0.5B-72B) on NVIDIA H100, our kernels achieve up to 2.47x speedup over PyTorch, shifting workloads from memory-bound to compute-bound and reaching 79.5% peak BF16 utilization. We demonstrate that torch.compile cannot replicate this fusion (3-7x slower than our kernels), validating the need for hand-crafted tile-level design. Our fused kernels are also numerically superior, achieving zero mismatches compared to 4.5-11% for cuBLAS.

Communication Dynamics Neural Networks (CDNNs) apply the circulant-spectral machinery of the Communication Dynamics framework to neural-network layer design. We introduce CDLinear, a block-circulant linear layer with block size B = 2l + 1 that uses 1/B the parameters of a dense layer with the same input and output dimensions. The construction gives an explicit Fourier-domain diagnostic for optimization: for mean-squared loss, the weight Hessian is diagonalized by the discrete Fourier transform, with eigenvalues determined directly by the Fourier spectrum of the input blocks. Under input pre-whitening, the population Hessian condition number is exactly 1, and the empirical condition number is bounded by 1 + O(sqrt(B/N)) for N samples. We implement CDLinear in pure NumPy with hand-derived backward passes and verify gradients by finite differences. On the 8x8 MNIST digits benchmark, across three random seeds, a CDLinear MLP with B = 4 reaches 97.50% +/- 0.23% test accuracy using 2,380 parameters, compared with 98.15% +/- 0.47% for a dense baseline using 8,970 parameters. This gives a 3.8x parameter reduction at a 0.65% accuracy cost. The CD-MLP's mean Hessian condition number is 1.9e4, about 310x smaller than the dense baseline's 5.9e6. We position CDLinear as a special case of structured matrix neural-network layers, with the main contributions being a closed-form Hessian-spectrum diagnostic, a principled discrete sequence of block multiplicities, and an explicit conditioning analysis. We also release a reference PyTorch implementation integrating CDLinear into a DeepSeek-V3-style mixture-of-experts transformer for future large-scale benchmarks.

Long-term time series forecasting requires models that simultaneously capture rapid oscillations, medium-range periodicities, and slowly evolving macro-trends from a fixed look-back window. Existing lightweight MLP-based models typically operate on a single temporal resolution, limiting their ability to explicitly model patterns at multiple scales. We propose MSMixer, a channel-independent multi-scale MLP architecture that addresses this limitation through three complementary innovations: (i) three parallel scale branches at down-sample factors {1x, 4x, 16x} with independent MLP blocks, (ii) a learnable softmax gate that dynamically weighs branch outputs, and (iii) a DLinear complementary shortcut that provides full-window trend and seasonality context. MSMixer contains only 112K parameters at H=96 and runs at O(T) complexity. Evaluated on four ETT benchmarks with standard chronological splits and three random seeds, MSMixer achieves the lowest average MSE (0.357) among lightweight models, outperforming DLinear (0.386, -7.4%) and NLinear (0.365, -2.1%), winning 12 of 16 configurations. Against five Transformer-based baselines from the literature, MSMixer achieves best or second-best MSE in 9 of 16 configurations while using 5x fewer parameters than PatchTST. Ablation and sensitivity analyses confirm the complementary contributions of the multi-scale branches and the DLinear shortcut.

Existing research largely attributes the global sequence modeling capability of Transformers to the explicit computation of attention weights, a process that inherently incurs quadratic computational complexity. In this work, we offer a novel perspective: we demonstrate that attention can be mathematically reframed as a Multi-Layer Perceptron (MLP) equipped with dynamically predicted parameters. Through this lens, we explain attention's global modeling power not as explicit token-wise aggregation, but as an implicit process where dynamically generated parameters act as a compressed representation of the global context. Inspired by this insight, we investigate a fundamental question: can we achieve Transformer-level sequence global modeling entirely through dynamic parameterization while maintaining linear complexity, effectively replacing explicit attention? To explore this, we design various dynamic parameter prediction strategies and integrate them into standard network layers. Extensive empirical studies on vision models demonstrate that dynamic parameterization can indeed serve as a highly effective, linear-complexity alternative to explicit attention, opening new pathways for efficient sequence modeling. Code is available at https://github.com/LeapLabTHU/WeightFormer.

Large language models encode vast factual knowledge that can become outdated or incorrect after deployment, yet retraining is prohibitively costly. This motivates lifelong model editing, which updates targeted behavior while preserving the rest of the model. Existing editors, both parameter-modifying and parameter-preserving, degrade severely as edits accumulate and struggle to generalize across paraphrases. We propose HoReN, a codebook-based parameter-preserving editor that wraps a single MLP layer with a discrete key-value memory. HoReN treats each codebook entry as both a knowledge key and a Hopfield stored pattern, retrieves edits by angular similarity on the unit hypersphere, and refines queries through damped Hopfield dynamics so paraphrases converge to the correct memory basin while unrelated inputs remain stable. HoReN achieves strong editing performance with consistent gains across diverse benchmarks spanning standard ZsRE, structured WikiBigEdit, and unstructured UnKE evaluations. Moreover, HoReN scales to 50K sequential edits on ZsRE with stable overall performance above 0.93, while prior editors collapse or degrade severely before reaching 10K. Our code is available at https://github.com/ha11ucin8/HoReN.

Does structure in representations imply structure in computation? We study how Llama-3.1-8B reasons over cyclic concepts (e.g., "what month is six months after August?"). Even though Llama-3.1-8B's representations for these concepts are circularly structured, we find that instead of directly computing modular addition in the period of the cyclic concept (e.g., 12 for months), the model re-uses a generic addition mechanism across tasks that operates independently of concept-specific geometry. First, it computes the sum of its two inputs using base-10 addition (six + August=14). Then, it maps this sum back to cyclic concept space (14->February). We show that Llama-3.1-8B uses task-agnostic Fourier features to compute these sums--in fact, these features have periods that respect standard base-10 addition, e.g., 2, 5, and 10, rather than the cyclic concept period (e.g., 12 for months). Furthermore, we identify a sparse set of 28 MLP neurons re-used across all tasks (approximately 0.2% of the MLP at layer 18) that can be partitioned into disjoint clusters, each computing the sum for a Fourier feature with a different period. Our work highlights how an interplay between causal abstraction and feature geometry can deepen our mechanistic understanding of LMs.

We prove pathwise convergence of the layerwise evolution of tokens in a finite-depth, finite-width transformer model with MultiLayer Perceptron (MLP) blocks to a continuous-time stochastic interacting particle system. We also identify the stochastic partial differential equation describing the evolution of the tokens' distribution in this limit and prove propagation of chaos when the number of such tokens is large. The bounds we establish are quantitative and the limits we consider commute. We further prove that the limiting stochastic model displays synchronization by noise and establish exponential dissipation of the interaction energy on average, provided that the common noise is sufficiently coercive relative to the deterministic self-attention drift. We finally characterize the activation functions satisfying the former condition.

This paper presents an application of the biologically realistic JASTAP neural network model to classification tasks. The JASTAP neural network model is presented as an alternative to the basic multi-layer perceptron model. An evolutionary procedure previously applied to the simultaneous solution of feature selection and neural network training on standard multi-layer perceptrons is extended with JASTAP model. Preliminary results on IRIS standard data set give evidence that this extension allows the use of smaller neural networks that can handle noisier data without any degradation in classification accuracy.

Rapid advances in Large Language Models (LLMs) create new opportunities by enabling efficient exploration of broad, complex design spaces. This is particularly valuable in computer architecture, where performance depends on microarchitectural designs and policies drawn from vast combinatorial spaces. We introduce Agentic Architect, an agentic AI framework for computer architecture design exploration and optimization that combines LLM-driven code evolution with cycle-accurate simulation. The human architect specifies the optimization target, seed design, scoring function, simulator interface, and benchmark split, while the LLM explores implementations within these constraints. Across cache replacement, data prefetching, and branch prediction, Agentic Architect matches or exceeds state-of-the-art designs. Our best evolved cache replacement design achieves a 1.062x geomean IPC speedup over LRU, 0.6% over Mockingjay (1.056x). Our evolved branch predictor achieves a 1.100x geomean IPC speedup over Bimodal, 1.5% over its Hashed Perceptron seed (1.085x). Finally, our evolved prefetcher achieves a 1.76x geomean IPC speedup over no prefetching, 17% over its VA/AMPM Lite seed (1.59x) and 21% over SMS (1.55x). Our analysis surfaces several findings about agentic AI-driven microarchitecture design. Across evolved designs, components often correspond to known techniques; the novelty lies in how they are coordinated. The architect's role is shifting, but the human remains central. Seed quality bounds what search can achieve: evolution can refine and extend an existing mechanism, but cannot compensate for a weak foundation. Likewise, objectives, constraints, and prompt guidance affect reliability and generalization. Overall, Agentic Architect is the first end-to-end open-source framework for agentic AI architecture exploration and optimization.

We study when a skip connection around a single-hidden-layer MLP can be absorbed into a residual-free MLP of the same width. We first show that for any architecture whose skip branch is an invertible linear map (including Hyper-Connections and their manifold-constrained variants), the problem reduces to the identity skip case. For homogeneous activations of degree $k \neq 1$, such as ReLU$^2$ and ReGLU, absorption is unconditionally impossible by a degree argument. For gated activations whose gate is differentiable at the origin with $g(0) = 0$, including SwiGLU and GeGLU, a linearization argument gives the same conclusion. These impossibility results extend to arbitrary depth: a composition of $L$ residual blocks using such activations cannot be replicated by any composition of $L$ residual-free blocks of the same width. For ungated ReLU and GELU, the situation is richer. For generic weight matrices, absorption holds at the single-block level if and only if there exists an index set $S$ of size at least $d$ such that $W_{\mathrm{down}}[:,S]\,W_{\mathrm{up}}[S,:] = -I_d$. This condition is non-generic (it fails with probability one under continuous weight distributions), so skip-connected and residual-free MLPs of the same width represent generically disjoint function classes. Whether this disjointness persists for deep compositions of ReLU or GELU blocks remains open.

Most reinforcement-learning (RL) controllers used in continuous control are architecturally centralized: observations are compressed into a single latent state from which both value estimates and actions are produced. Biological control systems are often organized differently. Insects, in particular, coordinate navigation, heading stabilization, memory, and context-dependent action selection through distributed circuits rather than a single monolithic controller. Motivated by this contrast, we study an RL policy architecture that decomposes control into interacting modules for sensory encoding, heading representation, sparse associative memory, recurrent command generation, and local motor control, with a learned arbitration mechanism that allocates motor authority across modules. The model is evaluated on a two-dimensional navigation task that require simultaneous food seeking, obstacle avoidance, and predator escape. In a six-seed predator-navigation experiment trained with Proximal Policy Optimization (PPO) for 75 updates, the modular policy achieves the strongest final mean performance among the tested controllers, with final episodic return $-2798.8\pm964.4$ versus $-3778.0\pm628.1$ for a centralized gated recurrent unit (GRU) and $-4727.5\pm772.5$ for a centralized multilayer perceptron (MLP). The modular policy also attains the lowest final value loss and stable PPO optimization statistics while driving module-assignment entropy to $0.0457\pm0.0244$, indicating highly selective control allocation. These results suggest that distributed control can serve as a useful inductive bias for RL problems involving dynamically competing behavioral objectives.

While Central Pattern Generators (CPGs) and Multi-Layer Perceptrons (MLP) are widely used paradigms in robot control, few systematic studies have been performed on the relative merits of large parameter spaces. In contexts where input and output spaces are small and performance is bounded, having more parameters to optimize may actively hinder the learning process instead of empowering it. To empirically measure this, we submit a given robot morphology, with limited proprioceptive capabilities, to controller optimization under two bio-inspired paradigms (CPGs and MLPs) with evolutionary- and reinforcement- trainer protocols. By varying parameter spaces across multiple reward functions, we observe that shallow MLPs and densely connected CPGs result in better performance when compared to deeper MLPs or Actor-Critic architectures. To account for the relationship between said performance and the number of parameters, we introduce a Parameter Impact metric which demonstrates that the additional parameters required by the reinforcement technique do not translate into better performance, thus favouring evolutionary strategies.

In [97,99,100], an fl-RDT framework is introduced to characterize \emph{statistical computational gaps} (SCGs). Studying \emph{symmetric binary perceptrons} (SBPs), [100] obtained an \emph{algorithmic} threshold estimate $α_a\approx α_c^{(7)}\approx 1.6093$ at the 7th lifting level (for $κ=1$ margin), closely approaching $1.58$ local entropy (LE) prediction [18]. In this paper, we further connect parametric RDT to overlap gap properties (OGPs), another key geometric feature of the solution space. Specifically, for any positive integer $s$, we consider $s$-level ultrametric OGPs ($ult_s$-OGPs) and rigorously upper-bound the associated constraint densities $α_{ult_s}$. To achieve this, we develop an analytical union-bounding program consisting of combinatorial and probabilistic components. By casting the combinatorial part as a convex problem and the probabilistic part as a nested integration, we conduct numerical evaluations and obtain that the tightest bounds at the first two levels, $\barα_{ult_1} \approx 1.6578$ and $\barα_{ult_2} \approx 1.6219$, closely approach the 3rd and 4th lifting level parametric RDT estimates, $α_c^{(3)} \approx 1.6576$ and $α_c^{(4)} \approx 1.6218$. We also observe excellent agreement across other key parameters, including overlap values and the relative sizes of ultrametric clusters. Based on these observations, we propose several conjectures linking $ult$-OGP and parametric RDT. Specifically, we conjecture that algorithmic threshold $α_a=\lim_{s\rightarrow\infty} α_{ult_s} = \lim_{s\rightarrow\infty} \barα{ult_s} = \lim_{r\rightarrow\infty} α_{c}^{(r)}$, and $α_{ult_s} \leq α_{c}^{(s+2)}$ (with possible equality for some (maybe even all) $s$). Finally, we discuss the potential existence of a full isomorphism connecting all key parameters of $ult$-OGP and parametric RDT.

Robotic parkour provides a compelling benchmark for advancing locomotion over highly challenging terrain, including large discontinuities such as elevated steps. Recent approaches have demonstrated impressive capabilities, including dynamic climbing and jumping, but typically rely on sequential multilayer perceptron (MLP) architectures with densely activated layers. In contrast, sparsely gated mixture-of-experts (MoE) architectures have emerged in the large language model domain as an effective paradigm for improving scalability and performance by activating only a subset of parameters at inference time. In this work, we investigate the application of sparsely gated MoE architectures to vision-based robotic parkour. We compare control policies based on standard MLPs and MoE architectures under a controlled setting where the number of active parameters at inference time is matched. Experimental results on a real Unitree Go2 quadruped robot demonstrate clear performance gains, with the MoE policy achieving double the number of successful trials in traversing large obstacles compared to a standard MLP baseline. We further show that achieving comparable performance with a standard MLP requires scaling its parameter count to match that of the total MoE model, resulting in a 14.3% increase in computation time. These results highlight that sparsely gated MoE architectures provide a favorable trade-off between performance and computational efficiency, enabling improved scaling of control policies for vision-based robotic parkour. An anonymized link to the codebase is https://osf.io/v2kqj/files/github?view_only=7977dee10c0a44769184498eaba72e44.

Standard transformer architectures apply the same number of layers to every token regardless of contextual difficulty. We present Token-Selective Attention (TSA), a learned per-token gate on residual updates between consecutive transformer blocks. Each gate is a lightweight two-layer multi-layer perceptron (MLP) that produces a continuous halting probability, making the mechanism end-to-end differentiable with 1.7% parameter overhead and no changes to the base architecture. Notably, TSA learns difficulty-proportional routing without any explicit depth pressure: even at $λ=0$ (no depth regularisation), the task-loss gradient alone drives the router to skip 20% of token-layer operations. On character-level language modeling, TSA saved 14-23% of token-layer operations (TLOps) across Tiny-Shakespeare and enwik8 at <0.5% quality loss. At matched efficiency, TSA achieved 0.7% lower validation loss than early exit, and the learned routing transfers directly to inference-time sparse execution for real wall-clock speedup.

Large language models (LLMs) have demonstrated impressive capabilities, yet their internal mechanisms for handling reasoning-intensive tasks remain underexplored. To advance the understanding of model-internal processing mechanisms, we present an investigation of how LLMs perform arithmetic operations by examining internal mechanisms during task execution. Using early decoding, we trace how next-token predictions are constructed across layers. Our experiments reveal that while the models recognize arithmetic tasks early, correct result generation occurs only in the final layers. Notably, models proficient in arithmetic exhibit a clear division of labor between attention and MLP modules, where attention propagates input information and MLP modules aggregate it. This division is absent in less proficient models. Furthermore, successful models appear to process more challenging arithmetic tasks functionally, suggesting reasoning capabilities beyond factual recall.

MLP is a heavily used backbone in modern deep learning (DL) architectures for supervised learning on tabular data, and AdamW is the go-to optimizer used to train tabular DL models. Unlike architecture design, however, the choice of optimizer for tabular DL has not been examined systematically, despite new optimizers showing promise in other domains. To fill this gap, we benchmark 15 optimizers on 17 tabular datasets for training MLP-based models in the standard supervised learning setting under a shared experiment protocol. Our main finding is that the Muon optimizer consistently outperforms AdamW, and thus should be considered a strong and practical choice for practitioners and researchers, if the associated training efficiency overhead is affordable. Additionally, we find exponential moving average of model weights to be a simple yet effective technique that improves AdamW on vanilla MLPs, though its effect is less consistent across model variants.

Generative Agents, owing to their precise modeling and simulation capabilities of human behavior, have become a pivotal tool in the field of Artificial Intelligence in Education (AIEd) for uncovering complex cognitive processes of learners. However, existing educational agents predominantly rely on static personas to simulate student learning behaviors, neglecting the decisive role of deep cognitive capabilities in learning outcomes during practice interactions. Furthermore, they struggle to characterize the dynamic fluidity of knowledge internalization, transfer, and cognitive state transitions. To overcome this bottleneck, this paper proposes a human-like educational agent capable of simulating student cognitive evolution: CogEvolution. Specifically, we first construct a cognitive depth perceptron based on the Interactive, Constructive, Active, Passive (ICAP) taxonomy from cognitive psychology, achieving precise quantification of learner cognitive engagement. Subsequently, we propose a memory retrieval method based on Item Response Theory (IRT) to simulate the connection and assimilation of new and prior knowledge. Finally, we design a dynamic cognitive update mechanism based on evolutionary algorithms to simulate the real-time integration of student learning behaviors and cognitive evolution processes. Comprehensive evaluations demonstrate that CogEvolution not only significantly outperforms baseline models in behavioral fidelity and learning curve fitting but also uniquely reproduces plausible and robust cognitive evolutionary paths consistent with educational psychology expectations, providing a novel paradigm for constructing highly interpretable educational agents.

Transformers commonly exhibit an attention sink: disproportionately high attention to the first position. We study this behavior in GPT-2-style models with learned query biases and absolute positional embeddings. Combining structural analysis with causal interventions, validated across natural-language, mathematical, and code inputs, we find that the sink arises from the interaction among (i) a learned query bias, (ii) the first-layer MLP transformation of the positional encoding, and (iii) structure in the key projection. Crucially, each component we identify is individually dispensable: architectures omitting each of them robustly exhibit sinks. This indicates that attention sinks may arise through distinct circuits across architectures. These findings inform mitigation of sinks, and motivate broader investigation into why sinks emerge.

We study how depth, finite precision, state dimension, and chain-of-thought (CoT) affect the expressive power of multi-layer state-space models (SSMs). For the explicit-table $K$-function-composition problem, a canonical benchmark for sequential information propagation, we prove that any $L$-layer SSM solving $(L+3)$-function composition must satisfy $d^2p=Ω(N/L^3)$, where $d$ is the state dimension and $p$ is the per-scalar precision. Conversely, $K$-function composition is solved exactly by a $(K+1)$-layer generalized SSM with $d=1$ and $p=Θ(\log N)$. This gives a worst-case depth hierarchy for this formal problem family. We then distinguish post-input reasoning, in which all thought tokens are generated after the input, from input-interleaved reasoning, in which thought tokens may be inserted while the input stream is being read. Post-input reasoning does not circumvent our communication-based lower-bound pipeline, whereas input-interleaved reasoning admits bidirectional simulations with general deterministic one-pass streaming algorithms at the granularity of persistent memory. Finally, width and precision are not interchangeable under exact step-preserving simulation in the base affine-state model, but become interchangeable through the streaming-memory characterization once input-interleaved reasoning is allowed.

We present THEIA, a 2.75M-parameter modular neural architecture that learns the complete Kleene three-valued logic (K3) truth table from task data without external symbolic inference or hand-encoded K3 gate primitives. Across 5 seeds it passes all 39 K3 rules at >99% per-rule accuracy. K3 learnability is not the central finding: Transformer baselines also pass all 39 rules, and flat MLPs match THEIA on Phase-1 accuracy within 0.04pp. The contributions are two properties of the learned system. (1) Uncertainty-verdict asymmetric propagation. THEIA preserves Has-Unknown at every upstream boundary (80.0/91.1/90.8/99.7% across Arith/Order/Set/Logic vs. ~52% majority) while final-verdict decodability stays at or below a 73.4% U-vs-non-U oracle reference under linear and nonlinear probes. Activation patching on non-absorbent T->U cases flips 4,898/4,898 OR and 4,719/4,719 AND pairs across 5 seeds, ruling out residual shortcuts. (2) Reliability spectrum under discretized end-to-end training, on tasks decomposable along the engine boundaries. A mod-3 sequential composition task generalizes from 5- to 500-step evaluation at 99.96+-0.04% (5 seeds). Under identical Gumbel-softmax training, flat MLPs collapse to chance by 50 steps; a 2x2 ResMLP grid reaches >=99% on only 3/20 (config, seed) trials; a pre-LN Transformer reaches 99.24+-0.34%. Straight-through discretization prevents 0.999^500 compounding; the architectural separator is sustaining Phase-1 accuracy under Phase-3 training, where flat MLPs fail. Auxiliary: under per-architecture development defaults (not optimizer-controlled), THEIA reaches 12/12 Kleene coverage 6.5x faster than a parameter-comparable 8L Transformer; this narrows to ~3.6x with Transformer-standard tuning and 4.93x with the same recipe on both. Ratios are config-specific, not asymptotic.

We present ViewSplat, a view-adaptive 3D Gaussian splatting network for novel view synthesis from unposed images. While recent feed-forward 3D Gaussian splatting has significantly accelerated 3D scene reconstruction by bypassing per-scene optimization, a fundamental fidelity gap remains. We attribute this gap to the limited capacity of single-step feed-forward networks to regress static Gaussian primitives that satisfy all viewpoints. To address this limitation, we shift the paradigm from static primitive regression to view-adaptive splatting. Instead of a rigid Gaussian representation, our pipeline learns a view-adaptive latent representation. Specifically, ViewSplat initially predicts base Gaussian primitives alongside the weights of scene-conditioned View MLPs. During rendering, these MLPs take target-view coordinates as input and predict view-dependent residual updates for each Gaussian attribute (i.e., 3D position, scale, rotation, opacity, and color). This mechanism, which we term view-adaptive splatting, allows each primitive to rectify initial estimation errors, effectively capturing high-fidelity appearances. Extensive experiments demonstrate that ViewSplat achieves state-of-the-art fidelity while maintaining fast inference and real-time rendering; our large backbone variant runs at 15 FPS during inference and 90 FPS during rendering. Our project page is available at https://cvlab-uos.github.io/ViewSplat.

Machine learning models can achieve high predictive accuracy in hydrological applications but often lack physical interpretability. The Mass-Conserving Perceptron (MCP) provides a physics-aware artificial intelligence (AI) framework that enforces conservation principles while allowing hydrological process relationships to be learned from data. In this study, we investigate how progressively embedding physically meaningful representations of hydrological processes within a single MCP storage unit improves predictive skill and interpretability in rainfall-runoff modeling. Starting from a minimal MCP formulation, we sequentially introduce bounded soil storage, state-dependent conductivity, variable porosity, infiltration capacity, surface ponding, vertical drainage, and nonlinear water-table dynamics. The resulting hierarchy of process-aware MCP models is evaluated across 15 catchments spanning five hydroclimatic regions of the continental United States using daily streamflow prediction as the target. Results show that progressively augmenting the internal physical structure of the MCP unit generally improves predictive performance. The influence of these process representations is strongly hydroclimate dependent: vertical drainage substantially improves model skill in arid and snow-dominated basins but reduces performance in rainfall-dominated regions, while surface ponding has comparatively small effects. The best-performing MCP configurations approach the predictive skill of a Long Short-Term Memory benchmark while maintaining explicit physical interpretability. These results demonstrate that embedding hydrological process constraints within AI architectures provides a promising pathway toward interpretable and process-aware rainfall-runoff modeling.

Multi--task learning seeks to improve the generalization error by leveraging the common information shared by multiple related tasks. One challenge in multi--task learning is identifying formulations capable of uncovering the common information shared between different but related tasks. This paper provides a precise asymptotic analysis of a popular multi--task formulation associated with misspecified perceptron learning models. The main contribution of this paper is to precisely determine the reasons behind the benefits gained from combining multiple related tasks. Specifically, we show that combining multiple tasks is asymptotically equivalent to a traditional formulation with additional regularization terms that help improve the generalization performance. Another contribution is to empirically study the impact of combining tasks on the generalization error. In particular, we empirically show that the combination of multiple tasks postpones the double descent phenomenon and can mitigate it asymptotically.

Mixed-integer linear programming (MILP) is a foundational framework for combinatorial optimization across science and engineering, but remains hard to solve at scale due to NP-hardness. Recent learning-based methods typically model MILP instances as variable-constraint bipartite graphs and use Graph Neural Networks (GNNs) for representation learning, yet their locality limits representation power. We propose an attention-driven neural backbone that adopts an element-centric view of variables and constraints, with dual attention performing parallel intra-type self-attention and inter-type cross-attention. Across three representative tasks at the instance, element, and solving-state levels, our model consistently outperforms conventional GNN-based architectures, highlighting attention-based, element-centric modeling as a powerful foundation for learning-enhanced combinatorial optimization.

The Mamba model has gained significant attention for its computational advantages over Transformer-based models, while achieving comparable performance across a wide range of language tasks. Like Transformers, Mamba exhibits in-context learning (ICL) capabilities, i.e., making predictions for new tasks based on a prompt containing input-label pairs and a query, without requiring fine-tuning. Despite its empirical success, the theoretical understanding of Mamba remains limited, largely due to the nonlinearity introduced by its gating mechanism. To the best of our knowledge, this paper presents the first theoretical analysis of the training dynamics of a one-layer Mamba model, which consists of a linear attention component followed by a nonlinear gating layer, and its ICL generalization on unseen binary classification tasks, even when the prompt includes additive outliers. Our analysis shows that Mamba leverages the linear attention layer to select informative context examples and uses the nonlinear gating layer to suppress the influence of outliers. By establishing and comparing to the analysis of linear Transformers under the same setting, we show that although Mamba may require more training iterations to converge, it maintains accurate predictions even when the proportion of outliers exceeds the threshold that a linear Transformer can tolerate. These theoretical findings are supported by empirical experiments.

Large language models (LLMs) are prone to generating factually incorrect content, motivating methods for assessing truthfulness from internal model signals. While supervised probing approaches can be effective, they require labeled data and classifier training. Recent training-free methods avoid parameter optimization but rely on coarse activation statistics that provide limited insight into how truthfulness-related signals arise within the model. We present a training-free approach that operates at the level of individual multi-layer perceptron (MLP) value vectors. Through a systematic analysis, we find that although most value vectors show no meaningful signal, a sparse subset exhibits stable and directionally consistent correlations with content truthfulness. Leveraging this observation, we propose \textbf{TruthV}, a simple inference method that aggregates preferences expressed by these value vectors. TruthV requires only a small support set to identify relevant vectors and introduces no additional model parameters or classifier weights. We evaluate TruthV across model scales from 2B to 13B and multiple benchmarks, including question answering, natural language understanding, and hallucination evaluation. TruthV consistently outperforms existing training-free baselines, demonstrating that truthfulness-related variation in LLMs is captured in a sparse and structured manner at the level of MLP value vectors.

While CNNs and ViTs dominate vision architectures, all-MLP models offer a structurally simpler alternative whose patch-independent processing is naturally suited to exploiting temporal redundancy in video. We present evMLP, an all-MLP architecture that processes image patches independently, enabling an event-driven local update mechanism for video processing: by defining inter-frame changes as "events" and processing only the patches where events occur, evMLP avoids redundant computation on unchanged regions. Because each patch is processed independently, skipping an unchanged patch leaves all other outputs unaffected; at an event threshold of zero, the mechanism produces outputs identical to the dense baseline rather than an approximation. On ImageNet, evMLP achieves 73.5% top-1 accuracy at 1.03 GMACs (rising to 77.0% with knowledge distillation and an extended training schedule). On multiple video datasets, the event-driven mechanism reduces computational cost by 8.4%-26.8% while maintaining output consistency with the dense baseline. Wall-clock measurements confirm that these savings translate into actual speedup under compute-bound conditions, and that stream-level parallelism is the effective deployment strategy for multi-core systems. The code and pre-trained models are available at https://github.com/i-evi/evMLP.

Time series segmentation assigns a label to each part of a sequence. We formulate dense univariate segmentation as node classification on a graph whose nodes are the original time points. A local window provides node features without setting output granularity. We evaluate the approach on a TSSB-derived inductive benchmark built from disjoint UCR training and test instances. Under one fixed Graph Attention Network (GAT), visibility-based transformations achieve the highest mean ranks among eleven graph constructions. WDPVG, directed NVG, and weighted NVG form a statistically indistinguishable top group after Holm correction. On the 59-dataset Time Series Segmentation Benchmark, WDPVG+GAT reaches a weighted F1 of $0.916$, below seq2point at $0.951$ and statistically indistinguishable from same-feature MLP, random-forest, and 1-NN controls, because at this downsampled resolution each segment is short and the fixed $81$-sample window already spans most of it. At native resolution, where each segment is longer than that window, WDPVG+GAT is less sensitive to feature-window width and remains above the same-feature MLP at every tested window. The graph's advantage over these point-wise classifiers comes from context beyond the local window, which the visibility graph's long-range edges can supply, rather than from better features within it. In a separate resolution sweep, it is statistically tied with a window-searched seq2point while using about $70\times$ fewer parameters and $46\times$ less measured peak memory, although seq2point moves ahead after downsampling. This identifies a practical operating regime for finely sampled series when model size and repeated window tuning matter.

Dense prediction is a fundamental requirement for many medical vision tasks such as medical image restoration, registration, and segmentation. The most popular vision model, Convolutional Neural Networks (CNNs), has reached bottlenecks due to the intrinsic locality of convolution operations. Recently, transformers have been widely adopted for dense prediction for their capability to capture long-range visual dependence. However, due to the high computational complexity and large memory consumption of self-attention operations, transformers are usually used at downsampled feature resolutions. Such usage cannot effectively leverage the tissue-level textural information available only at the full image resolution. This textural information is crucial for medical dense prediction as it can differentiate the subtle human anatomy in medical images. In this study, we hypothesize that Multi-layer Perceptrons (MLPs) are superior alternatives to transformers in medical dense prediction where tissue-level details dominate the performance, as MLPs enable long-range dependence at the full image resolution. To validate our hypothesis, we develop a full-resolution hierarchical MLP framework that uses MLPs beginning from the full image resolution. We evaluate this framework with various MLP blocks on a wide range of medical dense prediction tasks including restoration, registration, and segmentation. Extensive experiments on six public well-benchmarked datasets show that, by simply using MLPs at full resolution, our framework outperforms its CNN and transformer counterparts and achieves state-of-the-art performance on various medical dense prediction tasks.