Overview

Sequence modeling asks how information in an ordered stream—such as words, sensor readings, user actions, or DNA—can be used to predict what comes next or interpret the sequence as a whole. Common approaches include recurrent networks such as LSTMs, Transformer attention, state-space models such as Mamba, and Markov or hidden Markov models; hybrids and diffusion methods are also used for long or structured sequences. Applications range from language and time-series forecasting to recommendation, biological analysis, and action recognition, with evaluation often testing prediction quality, long-range recall, and generalization to new lengths or domains. A central tension is balancing expressive memory and reliable long-context behavior against computation, latency, and memory costs, while ensuring benchmarks measure genuine sequence understanding rather than leakage or repetition.

Latest papers

268 papers18 papers in the last four weeks, down 31% on the four weeks before. 0.1% of all new papers.

Nov 16, 2026

Halting in Recurrent Sequence Models

Recurrent sequence models, such as recurrent neural networks (RNNs) and state space models (SSMs), are often trained to emit an end-of-sequence (EOS) token when their output sequence is complete. During training, EOS is treated as another token without a special loss term or incentive. Although this training paradigm is widely used, it is not understood how models decide to output EOS. We study a minimal version of the EOS halting problem in which networks are trained to emit a given input sequence an instructed number of times, followed by EOS. We train tanh RNNs and linear SSMs on this task, finding that the RNNs encode the number of tokens left to copy in a linear subspace, of which two principal components capture most of the variation. Hidden states with the same remaining count converge to a shared representation regardless of their instructed repeat count. By modulating the hidden state using an affine transformation at the boundary between copied sequences, we shift the RNN's state to the representation shared by states with one fewer copy remaining. The unchanged network emits EOS one copy earlier while reproducing the shortened target sequence in 94.4% of new sequences, revealing that this representation causally determines the stopping time. In the SSMs, the difference between trajectories with different repeat counts is separated across two-dimensional rotation planes whose contributions to the EOS comparison cancel during copying and align near the stopping time. Rotating the planes independently removes most of the intended stopping while preserving the copied content, which shows that the stopping time depends on the planes' relative phases. These results identify two candidate primitives of EOS halting in recurrent models: a shared encoding of the number of remaining tokens in tanh RNNs and phase alignment across fixed rotations in linear SSMs.
Nov 15, 2026

Deep Sequence Model for Genome Wide Discovery of Coding and Regulatory Element Signatures

Interpreting deep neural networks for genomic sequence classification remains challenging despite strong predictive performance. We develop a lightweight CNN with post-training gradient-based analysis to identify which sequence positions drive coding versus intergenomic classification. Applied to the standardized demo_coding_vs_intergenomic_seqs dataset, our model achieves 91.87% validation accuracy while revealing interpretable patterns: nine hot spots with consistently high importance, clear class-specific separation at positions 20-100 (coding) and 150-190 (intergenomic), and a strong mean-variance correlation (r = 0.530) indicating robust discriminative features. Gradient-based importance analysis shows that the model implicitly learns biologically meaningful sequence distinctions without explicit annotations. This work demonstrates that neural network interpretability and accuracy can coexist, providing a framework for understanding genomic sequence classification and enabling biology-driven hypothesis generation.
Nov 13, 2026

FlexCast: Flexible Context Learning for Time Series Foundation Models

Time series foundation models (TSFMs) based on autoregressive (AR) next-token prediction entangle two factors: variable-context conditioning and prediction-unit factorization with autoregressive rollout. Strong AR performance therefore does not reveal whether generalization stems from the flexible training distribution, the local prediction target, the left-to-right inference process, or their combination. We propose FlexCast, a flexible context-target learning framework that disentangles these factors. FlexCast reformulates forecasting as variable-length masked prediction over time series patches: during pretraining, it dynamically samples visible context and masked target patches, exposing the model to diverse context-target configurations without sequential decoding. At inference time, the same pretrained model supports one-step forecasting, confidence-based iterative refinement, and next-token-style generation through a unified mask-and-predict interface, enabling controlled comparison of inference schedules under identical architecture, data, and objective. Evaluated on BLAST and multiple zero-shot benchmarks, FlexCast achieves accuracy competitive with or better than strong TSFM baselines while substantially reducing inference latency. Ablations confirm that context diversity and contiguous target masking drive performance, while left-to-right decoding does not yield consistent gains once flexible context-target relationships have been learned, suggesting that the key benefit of AR-style TSFMs lies in flexible training rather than sequential inference. The code is available at https://github.com/BenchCouncil/FlexCast.
Nov 11, 2026

Riordan Arrays as Structured Operators for Artificial Intelligence

Riordan arrays have traditionally been studied as infinite lower-triangular matrices that encode generating-function transformations and combinatorial relationships. This note proposes a complementary direction: using Riordan arrays as structured operators in artificial intelligence and data analysis. For a data sequence with generating function A(z)A(z), a Riordan array (g(z),f(z))(g(z),f(z)) produces the transformation A(z)⟼g(z)A(f(z)).A(z)\longmapsto g(z)A(f(z)). This yields a parameterized family of causal linear operators that combine convolution, substitution, and multiscale mixing. The resulting structure may support parameter-efficient neural-network layers, variable-length sequence models, interpretable recurrent systems, invertible architectures, and models for hierarchical or recursively generated data. We explain what the infinite-dimensional viewpoint does and does not provide, outline several possible architectures, identify theoretical and numerical challenges, and propose an initial experimental program. The aim is not to claim that Riordan arrays are appropriate for all data, but to identify settings in which their algebraic structure supplies a useful inductive bias.
Nov 9, 2026

INTENT: Out-of-Sample, Out-of-Time, and Cold-Start Evaluation on User Event Sequences

Banks now train foundation models on ordered sequences of customer events such as transactions, app sessions, and authorizations. Deployed, such a model predicts forward in time, often for customers and merchants absent from training, yet each is evaluated on private data, so no two can be compared. Public benchmarks for such sequences rarely test a period the model has not seen; recommendation studies usually hold out each user's last events instead. We present INTENT (INTEraction, Novelty, and Transfer evaluation for next-action prediction), a public benchmark that scores a model in seven cells defined by whether the user, the period, and the target item appeared in training. An audit of 59 public dataset families retains five datasets; the most common defect is the lack of a calendar timestamp that orders events. No date lies within the time span of all five datasets, so each is split at its own cutoff, chosen by one rule fixed before any model is evaluated. Summed over the datasets, every cell holds at least 111,769 users. In reference results the cell determines the best model: a transformer leads on seen items after the cutoff but scores zero on new items, a variant that embeds item metadata helps mainly on MicroLens, and repeating the user's last item leads on new items in four of five datasets. This parameter-free control also shows that on three datasets most of the score lost after the cutoff reflects changed behavior rather than the model. The pipeline, the cutoff rule, and the per-cell user counts are publicly available.
Nov 8, 2026

Query- and Candidate-Aware Long User Sequence Modeling for Sponsored Product Ranking in Search

User sequence modeling is widely used in recommendation and advertising. Search ranking presents an additional challenge: the current query expresses immediate intent, while a candidate product determines which parts of a user's long-term history are relevant. We present a production sequence framework for sponsored-product ranking that augments recent behavior with two retrieval channels over long-term history: candidate-nearest-neighbor retrieval and query-nearest-neighbor retrieval. The selected actions are encoded together with dedicated query and candidate context tokens, while candidate-aware early fusion injects the candidate into every history token. Offline ablations show that dual retrieval improves major ranking metrics relative to the production baseline. In an online A/B test on sponsored-product search traffic, the model increased CTR, attributed conversion rate, and orders per impression.
Nov 6, 2026

Implicit Bias in State Space Models and Linear Autoregressive Training

Linear State Space Models (SSMs) and linear autoregressive sequence models reuse the same parameters across a sample-dependent number of steps. As a result, different examples impose constraints through different predictors, even though all predictors share the same weights. We study this phenomenon through a general model of multiple non-homogeneous polynomial predictors with shared parameters, in a classification setting. Under standard interpolation and directional-convergence assumptions, together with a trajectory-dependent effective-degree condition, we show that gradient descent on the exponential loss converges in direction to a KKT point of an effective-degree max-margin problem. In this problem, only examples with minimal effective degree impose hard margin constraints; examples with faster-growing margins are asymptotically inactive except for feasibility. Our result provides a first-order characterization of the implicit bias in linear SSMs and linear autoregressive models. It exposes a mechanism of length bias: a longer input sequence or additional autoregressive steps may increase the example's effective degree, and the examples with the smallest effective degree dominate the limiting classifier. Our result highlights how non-homogeneity and parameter sharing alter classical homogeneous implicit-bias results.
Nov 5, 2026

Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout

Linear attention compresses context into a fixed-size key-value state for linear-time training and constant-memory inference. Gated delta-rule models decay this state and erase its prediction for an incoming key before writing a new value; Gated DeltaNet-2 (GDN-2) generalizes this with channel-wise decay and decoupled, full-rank erase and write gates. However, we show that state updates are far less sensitive to erasure than to writes: key activation and unit normalization concentrate key energy onto a sparse subset of channels, whereas value writes scale all channels, so the erase gate can be parameterized at lower rank. Furthermore, slow-decaying state rows accumulate much larger norms than fast rows, dominating unweighted readout and forcing stronger write-time erasure to suppress interference. Guided by this analysis, we introduce Muir, which fuses all gates into a sensitivity-allocated low-rank projection, re-weights state rows with a matched-filter readout gate, and reinvests the saved parameters in a wider feed-forward network. Because standard chunkwise algorithms become numerically unstable at the large chunk sizes needed to saturate wide matrix units, we design a parallel update for stable, efficient training. Across 166M-1.34B parameters, Muir lowers perplexity by 1.8-2.0% over GDN-2 and 2.5-3.6% over Transformer, raises commonsense accuracy by up to 1.9 points, and improves extractive recall over GDN-2. At 1.34B, Muir outperforms both baselines on multi-hop and multi-needle recall, extrapolates 2,560x from its 4K training length to 10M-token streams without perplexity drift, and maintains near-100% passkey retrieval to 1M tokens with static readout. Muir achieves these gains while matching GDN-2 throughput, reaching its 100B-token validation loss 1.2x faster, and training 2.7x faster than Transformer at 32K.
Nov 5, 2026

MUSS: A Multi-scale and Sequence-based Model for Single-cell Gene Regulation

Modeling gene regulation in single cells requires jointly representing DNA sequence, chromatin accessibility, and gene expression across regulatory scales. Sequence-based models capture DNA-level regulatory grammar but often lack cell-level resolution, whereas many single-cell omics models represent genes and peaks as fixed-vocabulary tokens without their underlying sequence. Here, we present MUSS, a sequence-based multi-scale model for single-cell gene regulation that preserves both DNA-level and cell-level resolution. MUSS derives peak- and gene-level representations from DNA using a frozen pre-trained sequence encoder and integrates them through peak-peak self-attention and distance-biased peak-gene interactions. Peak masked language modeling followed by training on paired scATAC-seq and scRNA-seq aligns these representations with cell-specific regulatory states. Across gene expression prediction, enhancer-gene linking, motif-guided regulatory prioritization, and zero-shot cell clustering, MUSS achieves strong performance against sequence-based and single-cell omics baselines. It further predicts expression for genes and peaks unseen during task training. These results support sequence-grounded, multi-scale representations for decoding cell-type-specific regulatory programs. Code, datasets, and model checkpoints are available at https://github.com/Lotfollahi-lab/MUSS_neurips26.
Nov 3, 2026

Near-Zero Token Embeddings Silently Stall Training in Gated Sequence Models

Among 291 surveyed open-weight language model repositories (280 distinct embedding tables), 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below 10−610^{-6}, compared with 0% of weight-tied models. Such near-zero rows can arise during pretraining when unseen rare or template tokens receive zero input gradients while AdamW weight decay shrinks untied embeddings. In tested Transformer configurations, non-initial positions or a BOS token prevent this failure mode. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), a near-zero token can amplify input-embedding gradients by more than 30 orders of magnitude (up to 35). At zero input, current-token readout factors (SiLU gates or linear queries) vanish, suppressing historical-state readout. Residual RMS values can then fall below the square root of the RMSNorm floor across consecutive layers, causing Pre-RMSNorm to amplify backward gradients by up to 10310^3 per layer. When a gradient entry exceeds 1.84×10191.84 \times 10^{19}, FP32 global-norm clipping in PyTorch, Hugging Face Trainer, and FSDP2 can overflow to infinity and zero the current batch's gradients. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discards updates on approximately 20% of steps and leaves triggered rows frozen. We propose two remedies. For released checkpoints, GroupClip decouples input-embedding clipping using a 64-bit or rescaled norm; its rescaled implementation eliminates update loss and heals triggered rows in one step with negligible measured end-to-end overhead. For pretraining, GateFloor (SiLU(z)+δ\mathrm{SiLU}(z) + δ, δ∈[0.01,0.1]δ\in [0.01, 0.1]) keeps the gate positive at zero input, reducing near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters, with perplexity comparable to the baseline.
Nov 2, 2026

Control Foundation Models across Changing Block Diagrams: Generalizing Time-Series Prediction to Dynamic Graphs

In the prediction and control of physical systems, combining records from different experiments and operations is expected to supply information that no single experiment provides on its own. Time-series foundation models, which learn a shared predictor from diverse time series, offer a framework for treating such records as a common training resource. Moreover, query-based prediction models, which switch between known and target signals within the same time series, make it possible to handle response prediction and control-input generation in a unified manner. However, in control-experiment data, the connections between signals change with the experimental stage, such as system identification, filtering, and feedback control, which calls for a formulation that handles signal values and connection structures dynamically and in a unified way. We formulate a control foundation model that defines a data space of dynamic graphs containing connections, signal values, and missing values, together with an operation on this space that specifies the known signals and the prediction targets. This allows response prediction and control-input generation to be described as a common masked prediction problem, even for experimental data with different connections and recorded signals. The Block Diagram Transformer (BDT), which implements this formulation, combines attention along the time axis with a graph neural network that follows the block-diagram connections, and can be initialized with the pretrained parameters of existing time-series foundation models. Trained on control-experiment data that mix multiple experimental settings, the proposed model stabilized an unstable system through feedback control, which is difficult to achieve by fine-tuning existing time-series foundation models. Ongoing work applying this approach to a wide range of target systems is also yielding promising results.
Nov 1, 2026

CoGSeq-LTV: Scalable Missing-Aware Graph-Sequence Modeling for Multi-Business Transaction Analytics

Large-scale super-apps generate heterogeneous transaction logs in which users interact with multiple businesses, item identifiers are only partially available, and future value targets are sparse, heavy-tailed, and cumulative over time. We study multi-business, multi-horizon value prediction as a data-intensive analytics problem over such logs, rather than as a single-table task. (i) a batch-induced user-item graph that provides collaborative signals without materializing a global graph and remains applicable when item identifiers are missing through reserved unknown-item handling, (ii) business-aware sequence encoders that preserve recency, inter-event gaps, payment context, and goods-table availability, and (iii) cross-view contrastive alignment that transfers user-level intent across businesses while retaining business-specific temporal patterns. To handle zero-inflated and long-tailed cumulative targets, CoGSeq-LTV adopts a zero-inflated mixture log-normal head with MAE-aligned point supervision and validation-calibrated zero-aware inference. Experiments on the BiBiSheng transaction-log dataset show that CoGSeq-LTV improves over representative LTV baselines across businesses and horizons, including under a later out-of-time evaluation window.
Oct 29, 2026

When can Optimally Trained Transformers Achieve Near-Lossless Token Prediction for non-Markovian Data?

There exists little mathematical theory explaining why Transformer architectures are empirically successful even for non-Markovian data. To this end, this paper studies next-token prediction via stochastic optimal control theoretically trained Transformers for such sources, where we consider either stationary and ergodic non-Markovian sources or those which are hidden Markov models of some finite order (possibly non-stationary). In either case, we first obtain error bounds via finite models with finite window approximations. For stationary ergodic sources, the resulting error is bounded by an entropy gap; for NNth order Markovian sources, it decays at an exponential rate α⌊M/N⌋α^{\lfloor M/N\rfloor}, where MM is the window length and αα is determined by filter stability (and the Dobrushin coefficients of the transition and measurement kernels). We then construct a Transformer of depth M+1M+1, one head per layer, realizing the truncated filter, and we optimally train it with stochastic control developed in earlier work, which rigorously returns a near optimal Transformer. In particular, we show that for a non-Markovian source, either under stationarity and ergodicity for general models or under filter stability for hidden Markov models, an optimally trained Transformer has arbitrarily small performance error with increasing context length, data, and approximation refinement. Numerical studies accompany the theory.
Oct 29, 2026

Observability-Aligned Evidential Masked Pretraining for Reliable VRU Motion Prediction

Vulnerable road user (VRU) prediction must operate under partial observability and distribution shift, so reliable uncertainty matters alongside accuracy. Masked pretraining is attractive in this setting because it naturally exposes the model to missing content during self-supervision, but existing time-series masked autoencoding methods are typically deterministic and do not constrain how confidence should change with observability. We introduce eCoMAE, an evidential masked pretraining method for multivariate time series that explicitly aligns predictive evidence with what is observed. eCoMAE combines block-wise time-step masking with a Normal-Inverse-Gamma decoder and three observability-aware mechanisms: uncertainty-weighted mean consistency on positions visible in both a masked view and a reference pass, evidence suppression on masked positions, and representation consistency across stochastic masks. On IMPTC benchmark for VRU intention prediction under leave-one-environment-out weather shift, eCoMAE improves the strongest baseline from 0.7971 to 0.8068 in OOD accuracy and from 0.0718 to 0.0648 in OOD AURC. Additional evaluation on pedestrian trajectory forecasting across IMPTC, Waymo, nuScenes, and inD further yields the best fine-tuning MSE on all four datasets. These results show that observability-aligned evidential pretraining improves reliability under shift and remains effective beyond the primary benchmark.
Oct 28, 2026

DES-JEPA: Learning a Shared Vocabulary for High-Dimensional Discrete Event Sequences

Temporal foundation models such as Chronos and TimesFM generalize across domains through a shared numerical observation space. In contrast, discrete event sequences (DESs), including operational event logs or patient clinical-event trajectories, have domain-specific event vocabularies whose labels lack a consistent correspondence across domains. Beyond this lack of consistency, the number of distinct events in real-world settings is high enough to challenge both classical temporal point process modeling and current in-context foundation inference models such as FIM-PP. Critically, current foundation models for DESs are limited to a small, fixed number of event types, restricting their applicability to realistic, high-dimensional event streams. To address this, we propose two complementary modules: (i) DES-JEPA, a joint-embedding predictive architecture that learns a shared discrete vocabulary from synthetic stochastic processes with unaligned event vocabularies; and (ii) DES-PFN, a discrete event sequence prior-fitted network that reuses the learned vocabulary for in-context next-event prediction. Without task-specific fine-tuning, this framework achieves competitive zero-shot performance on eight unseen real-world datasets and scales up to 700 distinct event types.
Oct 27, 2026

GTFormer: A Syntax-Free Graph Network Alternative to Transformers for 8192×\times Length Extrapolation in Linear Time

Self-attention mechanisms have established the Transformer as the dominant backbone across natural language processing, source code intelligence, and automated theorem proving. However, all-to-all pairwise token comparison enforces a quadratic O(N2)\mathcal{O}(N^2) compute and memory barrier alongside severe softmax attention dilution as sequence length NN scales beyond the training horizon. Inspired by recent advances in algebraic graph learning (WPNet), this work introduces GTFormer, a syntax-free hierarchical Graph Neural Network (GNN) that replaces flat self-attention for general long-context sequence modeling. GTFormer organizes raw token streams into a bounded-degree hierarchy combining a balanced binary reduction tree (⌈log⁡2N⌉\lceil \log_2 N \rceil depth), horizontal boundary-bridging shift edges, random expander shortcuts, multi-channel convex highway gating, and a top-down broadcast pass. Across eight long-context code, log, and multi-hop retrieval benchmarks trained exclusively on sequences of length N≤512N \le 512, GTFormer extrapolates zero-shot to N=4,194,304N = 4{,}194{,}304 (8192×8192\times) on a single commodity GPU in linear O(N)\mathcal{O}(N) memory, maintaining near-perfect accuracy where standard and RoPE Transformers suffer from out-of-distribution attention collapse and out-of-memory exhaustion.
Oct 27, 2026

TRIDENT-Seq: Profile-Matched Prefix-Future Contrast for Label-Efficient Payment Risk Modeling

Ant International accumulates large volumes of transaction history across users and payment businesses. Confirmed fraud and chargeback outcomes, by contrast, are scarce, delayed, and specific to each business. A single encoder pretrained on unlabeled histories could serve all businesses, but only if pretraining captures the behavior that separates risk. Coarse activity differences can make contrastive pretraining on transaction logs too easy: transaction volume, amount, and timing can distinguish a history's continuation from arbitrary candidates without modeling how payment resources are used. We present TRIDENT-Seq, a transaction foundation model trained with Prefix-Future Contrast (PFC). PFC matches each prefix to a summary of its subsequent behavior among candidates with similar activity profiles; activity prediction and dispersion regularization act on the encoder output. Across three production payment businesses, the pretrained encoder improves average precision (AP) and recall at a 0.5% intervention budget over architecture-matched scratch training in all six business-budget settings. With a fixed validation set, it exceeds full-label scratch training in mean AP and recall using 10% of the training labels on Wallet and 25% on Shopping and SME. Pretraining only on another business's histories still helps. Replacing profile-matched candidates with random ones cuts recall from 24.62% to 5.64%. In Ant International's deployment evaluation, recorded amount-based chargeback and overall loss rates are 37.7% and 37.9% lower than those of the previous XGBoost-based system under the same intervention budget.
Oct 26, 2026

Layerwise Repacking: Variable-Length Hybrid-Attention Training without an Extra Redistribution Pass

Variable-length training creates conflicting parallelism requirements across the layers of a hybrid-attention model. The longest samples require wide sequence-parallel (SP) groups of 16 or 32 ranks for full attention, whose cost grows quadratically with sequence length. Sparse-attention layers, dominated by near-linear MoE work, can instead balance most evaluated workloads within node-local SP8 groups. Reusing the wide groups for these layers pushes their all-to-all communication across nodes, where the observed effective rate is 7.0x lower. Yet switching to smaller groups changes which ranks own each sample; a separate redistribution at every full-sparse boundary would introduce another data transfer. Layerwise Repacking keeps each SP16/32 parent for full attention and repacks its samples into node-local sparse children at SP{1, 2, 4, 8}, where hidden states persist between full layers. The layout change rides on the two existing full-layer Ulysses sites: at each, one variable-size all-to-all performs both the head exchange and the sample repacking, so no redistribution pass is added. Because the balance the children can reach depends on which samples share a parent, an auxiliary parent solver rebalances parent membership before the children are formed. On all 22 steps with SP16/32 parents among the first 200 training steps of a 1F3S hybrid-attention model (one full-attention layer per three sparse-attention layers), repacking alone reduces the mean end-to-end time by 5.95% and raises model FLOPs utilization (MFU) by 3.75 percentage points over the baseline with repacking disabled; with the improved solver, end-to-end time falls by 7.61% and MFU rises by 4.83 percentage points.
Oct 26, 2026

LEAP: a Latent Embedding Anomaly Patch-Transformer for time-series anomaly detection

Time-series anomaly detection is dominated by reconstruction-based neural detectors, whose objective is structurally flawed. Since the target is also the input, an expressive model can learn to reconstruct anomalies too, leaving uninformative residuals. Recent large-scale benchmarks show that, on curated data, elaborate deep detectors rarely beat simple statistical baselines, while current leaders rely on representation learning. We propose LEAP (Latent Embedding Anomaly Patch-Transformer), which combines masked prediction with latent representation-based anomaly evidence, without relying on reconstruction or attention-based scoring. The model removes patches from a window and predicts them from the remaining temporal and cross-channel context, so prediction errors measure how well learned dynamics explain hidden segments. Deterministic complementary group masking ensures every time step is scored while hidden, with identical visible fractions during training and inference. LEAP adds three label-free evidence sources: latent-embedding novelty calibrated by local density, local temporal contrast, and per-channel novelty pooled over channels selected by six label-free criteria. Under the TSB-AD protocol, the full model achieves a VUS-PR of 0.5179±0.00220.5179 \pm 0.0022 on 180180 held-out multivariate series. On TSB-AD-U, a univariate variant combining multi-scale masked-prediction scoring with latent novelty, without contrast or channel-pool novelty, achieves 0.6194±0.00160.6194 \pm 0.0016 on 350350 held-out series. Both results average three random seeds. A five-stage ablation on TSB-AD-M, supported by paired Wilcoxon signed-rank tests, shows statistically significant gains from each successive component.
Oct 17, 2026

HiPPO Functional Memory: Coordinate-Queryable Attention for Irregular Time Series

Irregular time series require models to answer temporal queries at arbitrary coordinates, yet existing approaches typically retain observation-dependent representations that must be repeatedly resolved for each query. We introduce HiPPO Functional Memory (HiPPO-FM), a functional-memory attention operator that compiles an irregular history into a reusable coordinate-addressable state. Instead of storing observation-indexed tokens or latent trajectories, HiPPO-FM writes observations into a fixed-rank functional space and performs direct coordinate evaluation through matched functional geometry. Kernel-smoothed HiPPO coordinates and Gram-based readout define a shared write/read mechanism, while observed-support self-attention enables content-dependent representations before compression. We characterize the induced finite-dimensional query space and derive an error decomposition separating functional geometry approximation from learned-memory mismatch. Across continuous-coordinate reconstruction, irregular representation learning, real-world extrapolation, and event-sequence prediction, HiPPO-FM consistently improves strong irregular time-series baselines, achieving strong reconstruction accuracy on Spiral, competitive performance across 20 UEA datasets, improved real-world extrapolation, and superior results on several irregular event benchmarks. After encoding, the compact functional state can be reused across different query sets without reprocessing the observation history.
Oct 14, 2026

Empirical Boundaries of Reservoir Computing for Few-Shot Sequence Pattern Tasks

Efficient sequence modeling remains a fundamental challenge in machine learning. Despite the success of Transformers, recurrent neural networks (RNNs) have recently re-emerged as efficient alternatives. Particularly, Echo State Networks (ESNs) offer a compelling framework of randomized RNNs that have been extensively studied for time-series forecasting and dynamical systems. This paper presents an empirical analysis of distributional, architectural, and optimization variants of ESNs evaluated on few-shot sequence-to-sequence (seq2seq) pattern tasks. These synthetic tasks utilize discrete symbol sequences, characterized by delays, inactive phases, and triggered markers to isolate and evaluate memory and pattern recognition capabilities. The minimality of these tasks is crucial to understand the strengths and limitations of machine learning models. Furthermore, when the input signal utilizes a one-hot encoding, ESN dynamics are governed strictly by their weight distributions, offering an ideal lens through which to study reservoir computing. Our findings map the performance boundaries of ESNs relative to state-of-the-art (SOTA) models. We show that preserving a linear state regime is vital for memory tasks, such as pattern completion or simple copy. We propose a data-dependent modulation mechanism for nonlinear input injection into the reservoir (G-ESN). G-ESN functions as a statistical filter capable of solving tasks that conventional ESNs cannot (the adding problem and bracket matching). Finally, for Deep ESNs, we evaluate a shared readout mechanism across layers that outperforms conventional concatenated readouts on structural tasks, including selective copy, sorting problem, and bracket matching.
Oct 13, 2026

Learning Source Structure: Bayesian Prediction with Transformers

A transformer trained on sequences whose source parameters are drawn afresh each time can learn only what the sequences share: the source family and its prior. The Bayesian predictor that knows both, which we call informed Bayes, is the unique minimizer of the training loss and is computable from the family alone; we use it as a quantitative, falsifiable reference. Where earlier work compared transformers with Bayes on sources whose contexts can be learned independently, we build Markov sources with up to 64 symbols in which observations in one context inform prediction in another: product sources with known or hidden encodings, transition rows sharing an unknown base law and depth, and a mixture of three families. Our main result is that informed Bayes predicts, from the family and prior alone, how a trained transformer behaves in four tests that go beyond the learning curve: the loss that remains under a source outside the training family (predicted 2.237 bits, measured 2.309), the doubling of the extra loss when two independent components change rather than one (0.345 predicted, 0.350 measured), the benefit of other contexts when a token appears for the first time, and the depth at which copying an earlier continuation starts to pay. On the learning curves themselves, small transformers match informed Bayes when the shared structure is visible in the tokens, and hidden structure needs more layers and training, with gaps below 0.01 bits per token for the largest models. We explain why the imitation is possible and where it must fail: informed Bayes decomposes into sufficient statistics that attention selects and accumulates and an estimator, fixed by the prior, that the feed-forward network learns; training pays the complexity of informed Bayes under the architecture's prior, and finite-memory universal prediction bounds how closely a bounded network can reach it.
Oct 12, 2026

The Language of Looking: Motion Tokens Reveal Domain-Dependent Local Structure in Gaze Sequences

Human gaze is the primary high-bandwidth interface through which we parse the semantics of our environment. While often modeled as a sequence, the degree to which gaze exhibits transferable local structure across domains remains underexplored. Using motion tokenization as a diagnostic lens across four diverse datasets, we investigate the transferability of local sequence structure. We demonstrate that while gaze can share a reusable motion-token vocabulary, identical motion tokens are organized into different local transition patterns depending on the dataset and task context. Our results show the limitations of naive data-pooling under low-semantic motion-token representations and provide a framework for evaluating behavioral sequences through information-theoretic diagnostics.
Oct 12, 2026

FON: A Draft-Revise Dual-Organ Language Model without Sequence-Level Quadratic Attention

Self-attention assigns two functions to one mechanism-where to read and how to mix-at the price of O(L^2) training, an O(T) decode state, and position signals bound to the training length. Existing attention-free alternatives all optimize the single function of generation. We propose FON (Fusion-Organ Network), which differentiates one weight stack into a drafting organ (multi-tap learnable delays + content-adaptive recursion; causal, O(S) training, O(1) decode state, no sequence-level quadratic attention) and a revision organ that re-processes the written prefix through multi-scale ripple kernels injected via weak gates, never generating. At 30M parameters, iso-parameter and iso-token budget, FON reaches perplexity 131.03 on WikiText-103, beating Transformer (197.73) and Mamba2 (152.53); the advantage persists under all depth-aligned baselines and at the 70M tier (-20.8%). At 8x length extrapolation FON degrades by ~8-9% without positional-encoding engineering, between ALiBi and RoPE. The revision channel provides a structural pathway for in-domain infilling that causal architectures lack in principle (+19% over on-demand coupling), at a perplexity cost indistinguishable within measurement precision. We report all costs and boundaries: impaired long-range retrieval, synthetic-task disadvantages, weaker out-of-domain code infilling than a same-recipe Transformer, and a 6.5-12.3x wall-clock gap.
Oct 11, 2026

What DNA Foundation Models Learn Beyond Sequence Composition

Do genomic foundation models (FMs) provide predictive information beyond simple sequence composition? We evaluate frozen representations from five DNA FMs against normalized k-mer frequencies across 57 genomic classification datasets and two exploratory gene-expression regression tasks. Simple k-mer features remain competitive with most models, whereas NTv3 shows a consistent advantage; the ranking is robust to globally fixed and multi-scale k-mer baselines and alternative downstream probes. To test incremental predictive utility directly, we compare global 6-mer features with their concatenation with each FM embedding under a matched Random Forest protocol. NTv3 improves 46 of 57 datasets (mean ΔΔMCC = +0.072), while Caduceus provides smaller gains and the other models do not improve mean MCC. We separately decompose embeddings into a component linearly predictable from 6-mer frequencies and a ridge residual, finding that linear recoverability does not reliably indicate downstream utility. Exact 6-mer-preserving shuffles further reveal task-relevant sequence-order sensitivity, most pronounced for NTv3, although order sensitivity alone does not guarantee predictive improvements. Our results establish compositional features as essential controls and offer complementary diagnostics for evaluating the added value of frozen genomic representations, without making claims about fine-tuning capacity.
Oct 9, 2026

The RNN Homunculus: Learning to Gate Long-Term Memory

Learning long-range dependencies with recurrent neural networks (RNNs) is difficult because of the vanishing gradient problem. Gated recurrent neural networks like GRUs and LSTMs address this problem and are commonly explained by their ability to create additive copy paths through time. This explanation is correct, but incomplete: useful RNN gate values must themselves be learned, which is typically done with another RNN. We study this gate-learning problem through carefully controlled sequence classification tasks. Our analysis shows that gated architectures struggle when all training samples require long-range memory and thus equally suffer from the vanishing gradient problem. We intervene by adding short-dependency samples which let gated models learn solutions that capture long-dependency samples, but also generalize across different class labels. When properly learned, gated RNNs match or even exceed the test accuracy of modern architectures like Mamba and xLSTM. Overall, our results highlight gate learning as a central part of learning long-range memory and show how training data can support this process by exposing the same task over shorter dependencies.
Oct 9, 2026

Understanding Mamba-3 Through Oscillatory Dynamics

Mamba-3 introduces input-dependent rotations into recurrent state updates, but their contribution beyond state retention remains unclear. We interpret these ro- tations as underdamped second-order dynamics, providing a framework for in- vestigating the role of oscillation in sequence modeling. Controlled comparisons show that underdamped dynamics outperform non-oscillatory damping regimes on state-tracking tasks, while language-model pretraining demonstrates benefits for long-context prediction and length generalization. Interventions in pretrained models further establish the importance of learned phase, and state analysis sug- gests that rotation regulates how retained information combines with incoming up- dates. Taken together, these findings identify underdamped dynamics as a useful mechanism for recurrent computation and suggest that oscillation supports long- context processing by regulating how information accumulates.
Oct 9, 2026

Double-Take: Improving Capacity, Binding, and Selectivity In Long-Context Within Bidirectional Memory

Processing a long sequence does not guarantee that distant information remains usable. As context grows, a fixed-size memory must retain many associations, connect information across positions, and preserve relevant structure without being overwhelmed by everything it observes. These are distinct requirements: a memory may store more information without representing the relationships needed for retrieval, and it may transport those relationships without preserving the ones that matter. Thus, the challenge is not simply extending context length, but constructing a memory that remains usable as context grows. We study this problem in large-chunk test-time training (TTT) and identify two basic failures: single-step writes accumulate interference, limiting capacity, while self-associative writes cannot form cross-position bindings. We introduce Double-Take (DT), which repairs these failures with multi-step unpinned updates and cross-position mixing, and adds bidirectional change-sensitive writes for selective retention. Capacity alone leaves end-to-end recall at the 0.100.10 chance floor; adding binding raises it to 0.920.92. Trained only on 32 kb32\,\mathrm{kb} genomic contexts, DT reaches 0.9270.927 AUROC at 131 kb131\,\mathrm{kb} and 0.8560.856 near 1 Mb1\,\mathrm{Mb}, while transferring zero-shot from human to mouse. Together, these results identify capacity, binding, and selectivity as three requirements for usable long-range memory.
Oct 9, 2026

Time Delay System for Sequence Modeling

Existing sequence models share a critical limitation: they lack an explicit, interpretable mechanism for content-aware temporal recall. LSTMs compress everything into a single hidden state, thereby losing information about specific past events. Transformers attend to all tokens indiscriminately at quadratic cost, while state-space models like Mamba compress history into a fixed-dimensional state without explicit content-addressable memory. To address this gap, we introduce a novel framework grounded in nonlinear time-delay systems, in which the effective delay is dynamically estimated based on the semantic similarity between current and historical inputs. Our approach features three key innovations: (i) a differentiable semantic lag estimator that computes content-aware soft attention weights over history, producing a weighted historical context and similarity score; (ii) a diagonal state space with closed-form cumulative sum solution, achieving O(1) per-step complexity with no sequential loops; and (iii) an interpretable fuzzy semantic modulation mechanism that learns context-dependent rules for balancing instantaneous processing versus historical recall, where rule centers correspond to prototypical semantic contexts. Experimental results on text classification and generation outperform state-of-the-art methods.
Oct 8, 2026

TraceRelay: Attention-Aligned Recurrence over Rolling Traces

We present TraceRelay, an attention-aligned recurrent architecture that distributes persistent representations over a rolling sequence of low-dimensional traces. Local right looking attention forms increments from lower-layer representations; delivery is delayed until all attended inputs are in the causal past. A fixed additive phase recurrence accumulates the delayed increments, and left-looking attention reads the resulting residual augmented stream. A stride-wise prefix sum supports parallel prefill and bounded-buffer continuation. We study 36 small-model runs on Equal Repeats, bounded Dyck closing-type prediction, and causal Most-Freq generation, using three seeds per setting. At trained length 256, Equal Repeats models with recurrent phase inheritance reach 98.81-99.69% accuracy versus 50.73-51.63% for separately trained variants without inheritance, despite the latter receiving more updates. Accuracy drops sharply at lengths beyond the training range. At the longest evaluated lengths, models with more dimensions in the middle layer's recurrent traces perform better on Dyck (76.34% versus 55.49% close accuracy at length 4096), whereas models with fewer trace dimensions perform better on five-symbol Most-Freq (70.74% versus 55.60% exact generation at length 1024). These contrasting cases motivate further study of how the size of recurrent representations should be chosen for different tasks, without establishing a general rule across tasks or model configurations.