Sequence

Effective sequence modeling fundamentally requires balancing the retention of unbounded history with the high-resolution detection of abrupt short-term variations common in real-world phenomena. However, existing state space models (SSMs) relying on high-order polynomial projection operators (HiPPO) face a critical trade-off where uniform measures dilute recent information to maintain timescale invariance, while exponential measures sacrifice global context to capture local dynamics. This paper proposes a Fractional Recurrent Architecture for Computational Temporal Analysis of Long sequences (FRACTAL), a novel architecture integrating fractional measure theory into recursive memory updates to address this limitation. By deriving projection operators with analytically characterized spectral properties and a tunable singularity index, the proposed method amplifies sensitivity to recent signal perturbations while preserving the spectral structure that encodes scale-invariant memory dynamics. This theoretical innovation is instantiated within a simplified diagonalized state space framework by modulating input projection initialization to enable simultaneous capture of multi-scale temporal features. FRACTAL achieves an average score of 87.11% on the Long Range Arena benchmark, including 61.85% on the ListOps task, outperforming the S5 model.

Over the last two decades, language modeling has experienced a shift from the use of predominantly recurrent architectures that process tokens sequentially during training and inference to non-recurrent models that process sequence elements in parallel during training, which results in greater training efficiency and stability at the expense of lower inference throughput. Here we introduce the Structured Recurrent Mixer, an architecture that allows for algebraic conversion between a sequence parallel representation at train time and a recurrent representation at inference, notably without the need for specialized kernels or device-specific memory management. We show experimentally that this dual representation allows for greater training efficiency, higher input information capacity, and larger inference throughput and concurrency when compared to other linear complexity models. We postulate that recurrent models are poorly suited to extended sequence length scaling for information-rich inputs typical of language, but are well suited to scaling in the sample (batch) dimension due to their constant memory per sample. We provide Mojo/MAX inference implementations of SRMs exhibiting 12x the throughput and 170x the concurrency of similarly powerful Transformers inferenced on vLLM, increases characteristic of Pytorch implementations resulting in a 30% increase in compute-constant GSM8k Pass@k. We conclude by demonstrating that SRMs are effective reinforcement learning training candidates.

Convolutional architectures have emerged as powerful alternatives to Transformers for sequence modeling. The primary advantage is that they offer improved theoretical sequence length complexity by leveraging the Fast Fourier Transform (FFT). However, this theoretical improvement does not always meaningfully land in practice. One critical obstacle is that applying standard FFTs is not amenable to the large-scale training pipeline wherein data is packed from different sources into a single sequence for hardware efficiency. Indeed, standard FFT algorithms are not easily amenable to document packing. Existing workarounds suffer from severe inefficiencies, crippling the practical performance of convolutional architectures. We close this gap with RubiConv, a novel algorithm for performing hardware-efficient, boundary-respecting convolutions on packed sequences. Extensive experiments show that RubiConv achieves significant speedups over both attention and standard FFT-based baselines. This work makes the theoretical efficiency of long convolutional models a practical reality for large-scale, real-world data packing.

Sequence prediction methods for linear dynamical systems with long memory, i.e. marginally stable systems, typically achieve regret that grows linearly with the hidden dimension of the underlying generative model. While many methods have been developed to address this regime with varying success, we show that simply using the second-order Vovk-Azoury-Warmuth (VAW) algorithm to learn a short autoregressive-with-inputs (ARX) model achieves astoundingly strong results: for bounded sequential data from a marginally-stable linear dynamical system with spectra in the complex disk except for angular wedge of width $δ$ around the negative real axis, this algorithm achieves dimension-free regret $O\left( δ^{-4} \log^2 T \right)$. These bounds are state-of-the-art to our knowledge. The key components for our result come from 1) using the theory of ``Universal Sequence Preconditioning'' (USP) \cite{marsdenuniversal} to prove the existence of an optimal setting of autoregressive coefficients, 2) the application of VAW which takes better advantage of the memory compression provided by USP, and 3) the analysis of Faber polynomials on circular sectors to extend these results to systems with complex spectra.

Reinforcement learning, including reinforcement learning with verifiable rewards (RLVR), has emerged as a powerful approach for LLM post-training. Central to these approaches is the design of the importance sampling (IS) ratio used in off-policy policy-gradient estimation. Existing methods face a fundamental bias-variance dilemma: token-level IS ratios, as adopted by PPO (Schulman et al., 2017) and GRPO (Shao et al., 2024), introduce bias by ignoring prefix state distribution mismatch; full sequence ratios provide exact trajectory-level correction but suffer from high variance due to the multiplicative accumulation of per-token ratios, while GSPO (Zheng et al., 2025) improves numerical stability via length normalization at the cost of deviating from the exact full-sequence IS correction. In this work, we identify the cumulative token IS ratio, the product of per-token ratios up to position $t$, as a theoretically principled solution to this dilemma. We prove that, under the token-level policy-gradient formulation, this ratio provides an unbiased prefix correction for each token-level gradient term and has strictly lower variance than the full sequence ratio. Building on this insight, we propose CTPO (Cumulative Token Policy Optimization), which combines the cumulative token IS ratio with position-adaptive clipping that scales log-space clip bounds according to the natural $\sqrt{t}$ growth of the cumulative log-ratio. This yields more consistent regularization across token positions. We implement and evaluate CTPO in the tool-integrated reasoning setting on several challenging mathematical reasoning benchmarks, achieving the best average performance across both model scales compared with strong GRPO and GSPO baselines. Code will be available at https://github.com/horizon-llm/CTPO.

Fast Weight Programmers (FWPs) encode temporal dependencies through dynamically updated parameters rather than recurrent hidden states. Quantum FWPs (QFWPs) extend this idea with variational quantum circuits (VQCs), but existing implementations rely on multi-qubit architectures that are difficult to scale on noisy intermediate-scale quantum (NISQ) devices and expensive to simulate classically. We propose gated QKAN-FWP, a fast-weight framework that integrates FWP with Quantum-inspired Kolmogorov-Arnold Network (QKAN) using single-qubit data re-uploading circuits as learnable nonlinear activation, known as DatA Re-Uploading ActivatioN (DARUAN). We further introduce a scalar-gated fast-weight update rule that stabilizes parameter evolution, supported by a theoretical analysis of its adaptive memory kernel, geometric boundedness, and parallelizable gradient paths. We evaluate the framework across time-series benchmarks, MiniGrid reinforcement learning, and highlight real-world solar cycle forecasting as our main practical result. In the long-horizon setting with 528-month input window and 132-month forecast horizon, our 12.5k-parameter model achieves lower scaled Mean Square Error (MSE), peak amplitude error, and peak timing error than a suite of classical recurrent baselines with up to 13x more parameters, including Long Short-Term Memory (LSTM) networks (25.9k-89.1k parameters), WaveNet-LSTM (167k), Vanilla recurrent neural network (11.5k), and a Modified Echo State Network (132k). To validate NISQ compatibility, we further deploy the trained fast programmer on IonQ and IBM Quantum processors, recovering forecasting accuracy within 0.1% relative MSE of the noiseless simulator at 1024 shots. These results position gated QKAN-FWP as a scalable, parameter-efficient, and NISQ-compatible approach to quantum-inspired sequence modeling.

Antibody therapeutics are among the most successful modern medicines, yet computationally designing antibodies with desirable binding and developability properties remains challenging. While protein language models (pLMs) have emerged as powerful tools for antibody sequence design, existing approaches largely suffer from two key limitations: they predominantly memorize germline sequences rather than modeling biologically meaningful somatic variation, and they offer limited support for flexible classifier-guided conditional generation. We address these challenges through two primary contributions. First, we demonstrate that discrete diffusion fine-tuning achieves strong language modeling performance on antibody sequences while allowing for generation conditioned on any off-the-shelf classifier. Second, we introduce germline absorbing diffusion, a novel modification of the discrete diffusion noise process in which the germline sequence - rather than a masked sequence - serves as the absorbing state. This biologically motivated inductive bias restricts the model to learning the trajectory from germline to observed sequence, effectively excluding genetic variation and V(D)J recombination statistics from the learned distribution and dramatically mitigating germline bias. We show that germline diffusion improves non-germline residue prediction accuracy from 26 percent to 46 percent, approaching the theoretical upper bound set by true biological variability. We then demonstrate the utility of our germline diffusion model on the conditional generation tasks of sampling antibodies with improved hydrophobicity and predicted binding affinity. On both tasks our model shows an improved tradeoff between class adherence and sample quality, significantly outperforming EvoProtGrad, a popular strategy to sample from pLMs with gradient-based discrete Markov Chain Monte Carlo.

Ancestral sequence reconstruction (ASR) aims to infer extinct protein sequences at internal nodes of a phylogenetic tree. Classical ASR methods are typically based on continuous-time Markov substitution models, but they treat sites largely independently and handle insertions and deletions only weakly or not at all. We introduce a tree-conditioned edit-flow model for variable-length ASR. Given two descendant sequences and their branch distances to a shared ancestor, the model reconstructs the ancestor through paired bidirectional edit trajectories constrained to agree on a common ancestral state. On a benchmark of experimentally evolved sequences with only context-independent substitutions, the model does not match the accuracy of the best classical method, yet still achieves reasonable performance despite being trained on natural sequences that include insertions, deletions, and substitutions. On a benchmark of natural homologous sequences with abundant insertions and deletions, the model most accurately localizes inferred evolutionary change.

Music-inspired Automatic Stage Lighting Control (ASLC) has gained increasing attention in recent years due to the substantial time and financial costs associated with hiring and training professional lighting engineers. However, existing methods suffer from several notable limitations: the low interpretability of rule-based approaches, the restriction to single-primary-light control in music-to-color-space methods, and the limited transferability of music-to-controlling-parameter frameworks. To address these gaps, we propose SeqLight, a hierarchical deep learning framework that maps music to multi-light Hue-Saturation-Value (HSV) space. Our approach first customizes SkipBART, an end-to-end single primary light generation model, to predict the full light color distribution for each frame, followed by hybrid Imitation Learning (IL) techniques to derive an effective decomposition strategy that distributes the global color distribution among individual lights. Notably, the light decomposition module can be trained under varying venue-specific lighting configurations using only mixed light data and no professional demonstrations, thereby flexibly adapting across diverse venues. In this stage, we formulate the light decomposition task as a Goal-Conditioned Markov Decision Process (GCMDP), construct an expert demonstration set inspired by Hindsight Experience Replay (HER), and introduce a three-phase IL training pipeline, achieving strong generalization capability. To validate our IL solution for the proposed GCMDP, we conduct a series of quantitative analysis and human study. The code and trained models are provided at https://github.com/RS2002/SeqLight .

Boltzmann Machines trained on evolutionary sequence data have emerged as a powerful paradigm for the data-driven design of artificial proteins. However, the relationship between model architecture, specifically parameter density, and experimental performance remains poorly understood. Here, we investigate this relationship using the Chorismate Mutase enzyme family as a model system. We compare standard fully connected Boltzmann Machines for Direct Coupling Analysis (bmDCA) with sparse models generated via progressive edge activation (eaDCA) and edge decimation (edDCA). We identify a maximum-entropy model (meDCA) along the decimation trajectory that represents an optimal balance between constraint satisfaction and the flexibility of the probability distribution. We synthesized and tested artificial sequences from all models using an in vivo complementation assay, finding that all architectures, regardless of sparsity, generate functional enzymes with high success rates, even at significant divergence from natural sequences. Despite this functional equivalence, we demonstrate that the meDCA model samples a viable sequence space that is more than fifteen orders of magnitude larger than its low-entropy counterparts. Furthermore, comparative analyses reveal that high-entropy models systematically minimize overfitting and better capture the local neutral spaces surrounding natural proteins. These findings suggest that while various models satisfying coevolutionary statistics can generate functional sequences, high-entropy Boltzmann Machines provide a superior representation of the underlying evolutionary fitness landscape.

Agentic AI will be an essential enabling technology for designing future mobile communication systems, which could provide flexible and customized services, automate complex network operations, and drive autonomous decision-making across the network. This work studies how Large Language Model (LLM)-based network AI agents can be utilized to execute network procedures expressed as sequences of tool invocations. We investigate four approaches, which differ in how the agent obtains the procedure and in how execution is distributed between the agent and the underlying tools. We evaluated the latency and execution correctness across these approaches using a User Equipment (UE) IP allocation procedure as a case study. Furthermore, we conduct a stress test to examine how many sequential procedural steps an LLM agent can reliably execute before failure. Our results show that approaches relying on iterative agent-side reasoning incur higher latency and are more prone to execution errors, while approaches where the procedure is encapsulated within a single tool, which internally orchestrates the required steps by invoking other tools, reduce latency by limiting repeated reasoning. The stress-test results further show that the model with advanced tool-calling capability maintains reliable execution over longer procedures than the other evaluated models; however, all models exhibit reliability degradation as procedure length increases, revealing clear execution limits in multi-step tool-based workflows. To systematically analyze failures in procedure execution, we introduce a procedure-specific error taxonomy that categorizes deviations in multi-step procedural execution.

We give new examples of graphs and trees with dominating set sequences that are not log-concave. These examples were generated by PatternBoost, a transformer-based reinforcement learning software developed by Charton-Ellenberg-Wagner-Williamson. We also show: for any positive integer $m$, there exists a tree whose dominating set sequence is not log-concave for at least $m$ indices by modifying a similar construction of Bautista-Ramos for the independent set sequence. We show that a large class of caterpillar graphs has log-concave dominating set sequences. A continuous analogue of the sequence is also log-concave for all graphs.

Capturing user interests across extensive watch histories is critical for short-form video recommendation, yet scaling sequence length is limited by two bottlenecks: the semantic sparsity of atomic Video IDs and the quadratic computational complexity of Transformers. Traditional orthogonal Video IDs fail to capture content relationships and demand large embedding tables, while the quadratic complexity of self-attention restricts the maximum sequence length under strict industrial latency and resource constraints. In this work, we present a production-deployed framework for modeling ultra-long user behavior sequences at a billion-user scale. We first address the representation bottleneck by adopting content-native Semantic IDs. By utilizing depth-truncated, coarse-grained Semantic IDs, we shrink the embedding table size from corpus cardinality. This compact representation naturally generalizes to cold-start content through shared semantic prefixes. Second, to overcome the sequence scaling barrier, we introduce a Global-Aware Compression Transformer that leverages non-parametric temporal folding and unified global query integration to effectively condense the sequence, alleviating both the memory and computational bottlenecks of standard self-attention. Offline profiling on our computing infrastructure demonstrates an order-of-magnitude reduction in peak memory footprint and a drastic decrease in computational overhead. This efficiency gain enables supporting longer sequence lengths at an affordable cost in production, yielding substantial online gains in satisfied user engagement and satisfied content consumption in large-scale online A/B tests.

Modern industrial Deep Learning Recommendation Models typically extract user preferences through the analysis of sequential interaction histories, subsequently generating predictions based on these derived interests. The inherent heterogeneity in data characteristics frequently result in substantial under-utilization of computational resources during large-scale training, primarily due to computational bubbles caused by severe stragglers and slow blocking communications. This paper introduces FreeScale, a solution designed to (1) mitigate the straggler problem through meticulously load balanced input samples (2) minimize the blocking communication by overlapping prioritized embedding communications with computations (3) resolve the GPU resource competition during computation and communication overlapping by communicating through SM-Free techniques. Empirical evaluation demonstrates that FreeScale achieves up to 90.3% reduction in computational bubbles when applied to real-world workloads running on 256 H100 GPUs.

We introduce HubRouter, a pluggable module that replaces O(n^2) attention layers with O(nM) hub-mediated routing, where M << n is a small number of learned hub tokens. We demonstrate it in two from-scratch architectures: a Jamba-style hybrid and a 12-layer Transformer; retrofit into pretrained models is a tested negative case. HubRouter implements an encode-decode-score-council pipeline: M learned hubs cross-attend to all tokens, tokens project against hubs for routing fingerprints, a score head selects top-k tokens, and a sparse council attends only to the selected subset. We validate HubRouter in three settings. (1) Hub-Jamba yields a nominal 4.2% PPL improvement (200.2 vs 209.0, single seed; possibly within seed noise) and up to ~90x training throughput at sequence length 1024 in matched PyTorch-native baselines; an optimised baseline would narrow this to ~10-15x. (2) Graduated replacement of 25% of Transformer attention layers gives the best perplexity in our matched-budget sweep (268.0 vs 282.4 pure Transformer). (3) Hub-GPT provides strictly causal routing, achieving PPL 211.5 +/- 0.4 over 3 seeds (post council-causal fix); approximately 3 PPL worse than Jamba's 208.5 +/- 0.7, a measurable quality cost for avoiding O(n^2) computation. Post-fix, chunk size C has little effect; the pre-fix chunk-size benefit was an artifact of a bidirectional-council leak we found in adversarial review. A multi-seed hub-count sweep (~105 runs across M=1-32) reveals M=8-14 as the reliably-converging sub-band (4-5/5 seeds); M=6 is rescued to 5/5 by orthogonal regularization, while M>=20 shows increasing seed sensitivity. Companion paper arXiv:2603.20997 (Basu, 2026) defines the routing diagnostic task. Code and scripts will be released.

We establish a mathematical correspondence between state space models, a state-of-the-art architecture for capturing long-range dependencies in data, and an exactly solvable nonlinear oscillator network. As a specific example of this general correspondence, we analyze the diagonal linear time-invariant implementation of the Structured State Space Sequence model (S4). The correspondence embeds S4D, a specific implementation of S4, into a ring network topology, in which recent inputs are encoded, as waves of activity traveling over the one-dimensional spatial layout of the network. We then derive an exact operator expression for the full forward pass of S4D, yielding an analytical characterization of its complete input-output map. This expression reveals that the nonlinear decoder in the system induces interactions between these information-carrying waves that enable classifying real-world sequences. These results generalize across modern SSM architectures, and show that they admit an exact mathematical description with a clear physical interpretation. These insights enable a new level of interpretability for these systems in terms of nonlinear oscillator networks.

Structured sequence generation often requires a model to satisfy several input-derived constraints in a single output. Standard decoding methods may assign high probability to fluent continuations while placing low mass on continuations that realize all required anchors jointly. We study this regime as a rare-event sequential inference problem. LatticeBridge combines a compact prefix language model, instance-compiled surface automata, and a twisted sequential Monte Carlo (SMC) decoder with resampling, multilevel splitting, and a source-support proposal term derived from instance-provided phrases. The constraint representation is compiled from each input instance and does not rely on manually curated lexical classes. On 2,610 attainable validation tasks spanning CommonGen, E2E NLG, and WikiBio, the particle decoder improves exact anchor satisfaction and mean anchor coverage over greedy, beam-filtered, and best-of-k ancestral baselines under a shared proposal model. Since exact anchor satisfaction alone does not rule out unsupported attribute substitutions, the evaluation reports required-anchor coverage, source coverage, source-intrusion diagnostics, overlap, runtime, and particle statistics jointly. The benchmark characterizes the faithfulness-overlap-latency frontier under a fixed proposal model.

Large language models often appear strong on symbolic and algorithmic tasks, yet this apparent strength can hide brittle behaviour when problems become longer, harder, or slightly out of distribution. A major limitation of current reasoning benchmarks is that many primarily test whether a model can produce a valid answer, while paying less attention to whether the solution is minimal, robust, and stable under controlled difficulty scaling. We introduce RecurrReason, a difficulty-controlled benchmark of four recurrent logic puzzles (Tower of Hanoi, River Crossing, Block World, and Checkers Jumping) with BFS-optimal trajectories and a single interpretable difficulty parameter $N \in \{1,\dots,10\}$, totalling 10{,}817 unique puzzles and 285{,}933 moves. We benchmark two Transformer families, an encoder-decoder model (T5-style) and a decoder-only model (GPT-2-style), under consistent data splits and evaluation criteria, training on $N{=}1$ to $7$ and evaluating on both held-out in-distribution instances and harder out-of-distribution instances at $N{=}8$ to $10$. Fine-tuned pre-trained T5 achieves 97.27% validation and 81.00% OOD accuracy on Block World; all models score 0.00% on River Crossing under all conditions. Failure mode analysis reveals that architecture is a stronger determinant of success than scale. Pre-training transfers only to puzzles with locally structured transition functions. Our code and dataset will be open-sourced upon acceptance.

DNA sequence encoding is fundamental to gene function prediction, protein synthesis, and diverse downstream biological tasks. Despite the substantial progress achieved by large-scale DNA sequence pretraining, existing studies have overwhelmingly emphasized pretraining scale and custom downstream evaluation datasets, while neglecting some essential components of the pretraining paradigm. In this paper, we reveal three critical yet heretofore overlooked problems in DNA pretraining: inappropriate downstream datasets, inherent flaws in the neighbor-masking strategy, and the lack of detailed discussion on vocabulary. Therefore, we undertake comprehensive investigations and propose principled guidelines, including selection criteria for evaluation datasets, guiding task design, and in-depth vocabulary analysis. Extensive experiments validate the significance of our identified problems and support the rationale behind our recommendations. Finally, we introduce a standardized testbed that enables reproducible and rigorous benchmarking of DNA pretraining methods to advance the development of genomic foundation models.

Exo-to-Ego video generation aims to synthesize a first-person video from a synchronized third-person view and corresponding camera poses. While paired supervision is available, synchronized exo-ego data inherently introduces substantial spatio-temporal and geometric discontinuities, violating the smooth-motion assumptions of standard video generation benchmarks. We identify this synchronization-induced jump as the central challenge and propose Syn2Seq-Forcing, a sequential formulation that interpolates between the source and target videos to form a single continuous signal. By reframing Exo2Ego as sequential signal modeling rather than a conventional condition-output task, our approach enables diffusion-based sequence models, e.g. Diffusion Forcing Transformers (DFoT), to capture coherent transitions across frames more effectively. Empirically, we show that interpolating only the videos, without performing pose interpolation already produces significant improvements, emphasizing that the dominant difficulty arises from spatio-temporal discontinuities. Beyond immediate performance gains, this formulation establishes a general and flexible framework capable of unifying both Exo2Ego and Ego2Exo generation within a single continuous sequence model, providing a principled foundation for future research in cross-view video synthesis.

The King Wen sequence of the I-Ching (c. 1000 BC) orders 64 hexagrams -- states of a six-dimensional binary space -- in a pattern that has puzzled scholars for three millennia. We present a rigorous statistical characterization of this ordering using Monte Carlo permutation analysis against 100,000 random baselines. We find that the sequence has four statistically significant properties: higher-than-random transition distance (98.2nd percentile), negative lag-1 autocorrelation (p=0.037), yang-balanced groups of four (p=0.002), and asymmetric within-pair vs. between-pair distances (99.2nd percentile). These properties superficially resemble principles from curriculum learning and curiosity-driven exploration, motivating the hypothesis that they might benefit neural network training. We test this hypothesis through three experiments: learning rate schedule modulation, curriculum ordering, and seed sensitivity analysis, conducted across two hardware platforms (NVIDIA RTX 2060 with PyTorch and Apple Silicon with MLX). The results are uniformly negative. King Wen LR modulation degrades performance at all tested amplitudes. As curriculum ordering, King Wen is the worst non-sequential ordering on one platform and within noise on the other. A 30-seed sweep confirms that only King Wen's degradation exceeds natural seed variance. We explain why: the sequence's high variance -- the very property that makes it statistically distinctive -- destabilizes gradient-based optimization. Anti-habituation in a fixed combinatorial sequence is not the same as effective training dynamics.

Event-based cameras capture visual information as asynchronous streams of per-pixel brightness changes, generating sparse, temporally precise data. Compared to conventional frame-based sensors, they offer significant advantages in capturing high-speed dynamics while consuming substantially less power. Predicting future event representations from past observations is an important problem, enabling downstream tasks such as future semantic segmentation or object tracking without requiring access to future sensor measurements. While recent state-of-the-art approaches achieve strong performance, they often rely on computationally heavy backbones and, in some cases, large-scale pretraining, limiting their applicability in resource-constrained scenarios. In this work, we introduce E-TIDE, a lightweight, end-to-end trainable architecture for event-tensor prediction that is designed to operate efficiently without large-scale pretraining. Our approach employs the TIDE module (Temporal Interaction for Dynamic Events), motivated by efficient spatiotemporal interaction design for sparse event tensors, to capture temporal dependencies via large-kernel mixing and activity-aware gating while maintaining low computational complexity. Experiments on standard event-based datasets demonstrate that our method achieves competitive performance with significantly reduced model size and training requirements, making it well-suited for real-time deployment under tight latency and memory budgets.

Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to reducing inference costs. However, achieving high-quality generation in distilled models requires careful joint design of both the student architecture and the distillation process. Many prior distillation works evaluate downstream multiple-choice benchmarks by ranking candidate answers with log-likelihood rather than requiring autoregressive generation, which can obscure important differences in model quality. For example, on overlapping benchmarks, we show that a 7B distilled model that nearly matches its teacher to within 0.2 pp under log-likelihood scoring falls behind by 20.8 pp when it must generate answers autoregressively. We investigate this phenomenon with GenDistill, a multi-stage pipeline we designed for distilling a pretrained Transformer into an efficient Hybrid Kimi Delta Attention (Hybrid-KDA) student. Using it as a controlled testbed on Qwen3-0.6B, we systematically ablate six design axes (training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice) and evaluate every choice under both log-likelihood and generation-based protocols. We find that log-likelihood-based evaluation consistently underestimates the gap between teacher and student, and can in some cases reverse the ranking of design choices, so conclusions drawn from perplexity-only evaluation may be misleading. Among the factors we study, dataset selection, completion-only masking, and freezing attention layers during post-training have the largest impact on generation quality. Our best distillation recipe, using a Hybrid-KDA model as the student, retains 86-90% of teacher accuracy on knowledge benchmarks while reducing KV cache memory by up to 75% and improving time-to-first-token by 2-4x at 128K-token contexts.

Despite the remarkable capabilities of Multimodal Large Language Models (MLLMs), they still suffer from visual fading in long-context scenarios. Specifically, the attention to visual tokens diminishes as the text sequence lengthens, leading to text generation detached from visual constraints. We attribute this degradation to the inherent inductive bias of Multimodal RoPE, which penalizes inter-modal attention as the distance between visual and text tokens increases. To address this, we propose inter-modal Distance Invariant Position Encoding (DIPE), a simple but effective mechanism that disentangles position encoding based on modality interactions. DIPE retains the natural relative positioning for intra-modal interactions to preserve local structure, while enforcing an anchored perceptual proximity for inter-modal interactions. This strategy effectively mitigates the inter-modal distance-based penalty, ensuring that visual signals remain perceptually consistent regardless of the context length. Experimental results demonstrate that by integrating DIPE with Multimodal RoPE, the model maintains stable visual grounding in long-context scenarios, significantly alleviating visual fading while preserving performance on standard short-context benchmarks. Code is available at https://github.com/lchen1019/DIPE.

Time series reasoning tasks often start with a natural language question and require targeted analysis of a time series. Evidence may span the full series or appear in a few short intervals, so the model must decide what to inspect. Most existing approaches encode the entire time series into a fixed representation before inference, regardless of whether or not the entire sequence is relevant. We introduce ARTIST, which formulates time-series reasoning as a sequential decision problem. ARTIST interleaves reasoning with adaptive temporal segment selection. It adopts a controller-reasoner architecture and uses reinforcement learning to train the controller role to select informative segments and the reasoner role to generate segment-conditioned reasoning traces and final answers. During inference, the model actively acquires task-relevant information instead of relying on a static summary of the full sequence. We use a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior. We evaluate ARTIST on six time-series reasoning benchmarks and compare it with large language models, vision-language models, and prior time-series reasoning systems. ARTIST improves average accuracy by 6.46 absolute percentage points over the strongest baseline. The largest gains appear on rare event localization and multi-segment reasoning tasks. Supervised fine-tuning improves performance, and reinforcement learning provides additional gains by optimizing question-adaptive segment selection. These results show that selective data use drives effective time-series reasoning.

Generative sequence modeling faces a fundamental tension between the expressivity of Transformers and the efficiency of linear sequence models. Existing efficient architectures are theoretically bounded by shallow, single-step linear updates, while powerful iterative methods like Test-Time Training (TTT) break hardware parallelism due to two dimensions of serial dependency: token-level state reliance and step-level iteration loops. We propose PRISM (Parallel Residual Iterative Sequence Model) to resolve this tension. PRISM explicitly approximates the expressive gate-residual-direction iteration pattern of TTT in a parallelizable form. We employ a Write-Forget Decoupling strategy that isolates non-linearity within the injection operator. To bypass the serial dependency of explicit solvers, PRISM utilizes a two-stage proxy architecture: a short-convolution anchors the initial residual using local history energy, while a learned predictor estimates the refinement updates directly from the input. This design distills structural patterns associated with iterative correction into a parallelizable feedforward operator. Theoretically, we prove that this formulation achieves Rank-$L$ accumulation, structurally expanding the update scheme beyond the single-step Rank-$1$ bottleneck. Empirically, it achieves comparable performance to explicit optimization methods while achieving \textbf{174x higher throughput}. Codes are available in https://github.com/gpr-prism/prism/.