NeurIPS 2025 Papers

This paper investigates nonparametric quantile regression using recurrent neural networks (RNNs) and sparse recurrent neural networks (SRNNs) to approximate the conditional quantile function, which is assumed to follow a compositional hierarchical interaction model. We show that RNN- and SRNN-based estimators with rectified linear unit (ReLU) activation and appropriately designed architectures achieve the optimal nonparametric convergence rate, up to a logarithmic factor, under stationary, exponentially $\boldsymbol{\beta}$-mixing processes. To establish this result, we derive sharp approximation error bounds for functions in the hierarchical interaction model using RNNs and SRNNs, exploiting their close connection to sparse feedforward neural networks (SFNNs). Numerical experiments and an empirical study on the Dow Jones Industrial Average (DJIA) further support our theoretical findings.

The energy paradigm, exemplified by Hopfield networks, offers a principled framework for memory in neural systems by interpreting dynamics as descent on an energy surface. While powerful for static associative memories, it falls short in modeling sequential memory, where transitions between memories are essential. We introduce the Exponential Dynamic Energy Network (EDEN), a novel architecture that extends the energy paradigm to temporal domains by evolving the energy function over multiple timescales. EDEN combines a static high-capacity energy network with a slow, asymmetrically interacting modulatory population, enabling robust and controlled memory transitions. We formally derive short-timescale energy functions that govern local dynamics and use them to analytically compute memory escape times, revealing a phase transition between static and dynamic regimes. The analysis of capacity, defined as the number of memories that can be stored with minimal error rate as a function of the dimensions of the state space (number of feature neurons), for EDEN shows that it achieves exponential sequence memory capacity $\mathcal{O}(\gamma^N)$, outperforming the linear capacity $\mathcal{O}(N)$ of conventional models. Furthermore, EDEN's dynamics resemble the activity of time and ramping cells observed in the human brain during episodic memory tasks, grounding its biological relevance. By unifying static and sequential memory within a dynamic energy framework, EDEN offers a scalable and interpretable model for high-capacity temporal memory in both artificial and biological systems.

In recent years, large language models with Transformer architecture as the core have made breakthrough progress in many fields. At the same time, there are also some weaknesses in the large language model that have prompted people to reflect, among which the most fundamental one is the reflection on the Transformer architecture. The Transformer architecture has high parallelism and can fully utilize the computing power of GPUs, thus replacing models such as LSTM in the past few years. However, high parallelism is not a free lunch, as it fundamentally limits the performance of models. Especially, the problems that logarithmic precision Transformer architecture can solve are strictly limited to the $TC^0$. And there are many important issues that are usually considered out of $TC^0$, such as Python code evaluation, entity tracking, chess, and other state tracking tasks. Meanwhile, some recent state space methods based on Delta Rule have been able to break through the $TC^0$ architecture, but they are limited by fixed size state spaces and perform poorly on many tasks. To this end, we have re-examined the Transformer from the perspective of a state space with kernel functions and propose an improved Transformer called DeltaFormer. We have theoretically and practically demonstrated that the proposed new architecture can break through the limitation of the inherent $TC^0$ expressivity of Transformers and verified that it is not weaker than standard Transformer in language modeling tasks. We hope our work can provide inspiration for designing more expressive models.

The theoretical understanding of self-attention (SA) has been steadily progressing. A prominent line of work studies a class of SA layers that admit an energy function decreased by state updates. While it provides valuable insights into inherent biases in signal propagation, it often relies on idealized assumptions or additional constraints not necessarily present in standard SA. Thus, to broaden our understanding, this work aims to relax these energy constraints and provide an energy-agnostic characterization of inference dynamics by dynamical systems analysis. In more detail, we first consider relaxing the symmetry and single-head constraints traditionally required in energy-based formulations. Next, we show that analyzing the Jacobian matrix of the state is highly valuable when investigating more general SA architectures without necessarily admitting an energy function. It reveals that the normalization layer plays an essential role in suppressing the Lipschitzness of SA and the Jacobian's complex eigenvalues, which correspond to the oscillatory components of the dynamics. In addition, the Lyapunov exponents computed from the Jacobians demonstrate that the normalized dynamics lie close to a critical state, and this criticality serves as a strong indicator of high inference performance. Furthermore, the Jacobian perspective also enables us to develop regularization methods for training and a pseudo-energy for monitoring inference dynamics.

We present the first provable method for identifying symmetric linear dynamical systems (LDS) with accuracy guarantees that are independent of the system’s state dimension or effective memory. Our approach builds upon recent work that represents symmetric LDSs as convolutions learnable via fixed spectral transformations. We show how to invert this representation—recovering an LDS model from its spectral transform—yielding an end-to-end convex optimization procedure. This distillation preserves predictive accuracy while enabling constant-time and constant-space inference per token, independent of sequence length. We evaluate our method, SpectraLDS, as a component in sequence prediction architectures and demonstrate that accuracy is preserved while inference efficiency is improved on tasks such as language modeling.

We present LrcSSM, a $\textit{non-linear}$ recurrent model that processes long sequences as fast as today's linear state-space layers. By forcing its Jacobian matrix to be diagonal, the full sequence can be solved in parallel, giving $\mathcal{O}(TD)$ time and memory and only $\mathcal{O}(\log T)$ sequential depth, for input-sequence length $T$ and a state dimension $D$. Moreover, LrcSSM offers a formal gradient-stability guarantee that other input-varying systems such as Liquid-S4 and Mamba do not provide. Importantly, the diagonal Jacobian structure of our model results in no performance loss compared to the original model with dense Jacobian, and the approach can be generalized to other non-linear recurrent models, demonstrating broader applicability. On a suite of long-range forecasting tasks, we demonstrate that LrcSSM outperforms Transformers, LRU, S5, and Mamba.

We propose a novel online Gaussian process (GP) model that is capable of capturing long-term memory in sequential data in an online learning setting. Our model, Online HiPPO Sparse Variational Gaussian Process (OHSVGP), leverages the HiPPO (High-order Polynomial Projection Operators) framework, which is popularized in the RNN domain due to its long-range memory modeling capabilities. We interpret the HiPPO time-varying orthogonal projections as inducing variables with time-dependent orthogonal polynomial basis functions, which allows the SVGP inducing points to memorize the process history. We show that the HiPPO framework fits naturally into the interdomain GP framework and demonstrate that the kernel matrices can also be updated online in a recurrence form based on the ODE evolution of HiPPO. We evaluate OHSVGP with online prediction for 1D time series, continual learning in discriminative GP model for data with multidimensional inputs, and deep generative modeling with sparse Gaussian process variational autoencoder, showing that it outperforms existing online GP methods in terms of predictive performance, long-term memory preservation, and computational efficiency.

To enable embodied agents to operate effectively over extended timeframes, it is crucial to develop models that form and access memories to stay contextualized in their environment. In the current paradigm of training transformer-based policies for embodied sequential decision-making tasks, visual inputs often overwhelm the context limits of transformers, while humans can maintain and utilize a lifetime of experience compressed as memories. Significant compression is possible in principle, as much of the input is irrelevant and can be abstracted. However, existing approaches predominantly focus on either recurrent models with fixed-size memory or transformers with full-context reliance. In this work, we propose Memo, a transformer-based architecture and training recipe for reinforcement learning (RL) on memory-intensive, long-horizon tasks. Memo incorporates the creation and retrieval of memory by interleaving periodic summarization tokens with the inputs of a model during training. We demonstrate Memo’s effectiveness on a grid-world meta-RL benchmark and a multi-object navigation task in photo-realistic indoor settings. Memo outperforms naive long-context transformer baselines while being more compute and storage efficient. Additionally, Memo generalizes better to longer contexts at inference time and remains robust in streaming settings, where historical context must be truncated to fit inference constraints.

Generating hyperspectral images (HSIs) from RGB images through spectral reconstruction can significantly reduce the cost of HSI acquisition. In this paper, we propose a Fractal-Based Recursive Spectral Reconstruction Network (FRN), which differs from existing paradigms that attempt to directly integrate the full-spectrum information from the R, G, and B channels in a one-shot manner. Instead, it treats spectral reconstruction as a progressive process, predicting from broad to narrow bands or employing a coarse-to-fine approach for predicting the next wavelength. Inspired by fractals in mathematics, FRN establishes a novel spectral reconstruction paradigm by recursively invoking an atomic reconstruction module. In each invocation, only the spectral information from neighboring bands is used to provide clues for the generation of the image at the next wavelength, which follows the low-rank property of spectral data. Moreover, we design a band-aware state space model that employs a pixel-differentiated scanning strategy at different stages of the generation process, further suppressing interference from low-correlation regions caused by reflectance differences. Through extensive experimentation across different datasets, FRN achieves superior reconstruction performance compared to state-of-the-art methods. Code is available at https://github.com/mongko007/frn.

The rise of parallel computing hardware has made it increasingly important to understand which nonlinear state space models can be efficiently parallelized. Recent advances have shown that evaluating a state space model can be recast as solving a parallelizable optimization problem, and sometimes this approach yields dramatic speed-ups in evaluation time. However, the factors that govern the difficulty of these optimization problems remain unclear, limiting the larger adoption of the technique. In this work, we establish a precise relationship between the dynamics of a nonlinear system and the conditioning of its corresponding optimization formulation. We show that the predictability of a system, defined as the degree to which small perturbations in state influence future behavior, directly governs the number of optimization steps required for evaluation. In predictable systems, the state trajectory can be computed in $\mathcal{O}((\log T)^2)$ time, where $T$ is the sequence length, a major improvement over the conventional sequential approach. In contrast, chaotic or unpredictable systems exhibit poor conditioning, with the consequence that parallel evaluation converges too slowly to be useful. Importantly, our theoretical analysis demonstrates that for predictable systems, the optimization problem is always well-conditioned, whereas for unpredictable systems, the conditioning degrades exponentially as a function of the sequence length. We validate our claims through extensive experiments, providing practical guidance on when nonlinear dynamical systems can be efficiently parallelized, and highlighting predictability as a key design principle for parallelizable models.

In the past decades, brain imaging research underwent a shift from mapping tasked evoked brain regions of activations towards identifying and characterizing the dynamic brain networks of multiple coordinating brain regions. Electrophysiological signals are the direct manifestation of brain activities, thus, characterizing the whole brain electrophysiological networks (WBEN) can serve as a fundamental tool for neuroscience studies and clinical applications. In this work, we introduce the first framework for the integration of scalp EEG and intracranial EEG (iEEG) for the WBEN estimation with a principled estimation framework based on state-space models, where an Expectation-Maximization (EM) algorithm is designed to infer the state variables and brain connectivity simultaneously. We validated the proposed method on synthetic data, and the results revealed improved performance compared to traditional two-step methods using scalp EEG only, demonstrating the importance of including iEEG signal for WBEN estimation. For real data with simultaneous EEG and iEEG, we applied the developed framework to understand the information flows of the encoding and maintenance phases during the working memory task. The information flows between the subcortical and cortical regions are delineated, highlighting more significant information flows from cortical to subcortical regions than maintenance phases. The results are consistent with previous research findings but with the view of the whole brain scope, which underscores the unique utility of the proposed framework.

Parameter generation has long struggled to match the scale of today's large vision and language models, curbing its broader utility. In this paper, we introduce Recurrent Diffusion for Large-Scale Parameter Generation (RPG), a novel framework that generates full neural network parameters—up to hundreds of millions—on a single GPU. Our approach first partitions a network's parameters into non-overlapping 'tokens', each corresponding to a distinct portion of the model. A recurrent mechanism then learns the inter-token relationships, producing 'prototypes' which serve as conditions for a diffusion process that ultimately synthesizes the full parameters. Across a spectrum of architectures and tasks—including ResNets, ConvNeXts and ViTs on ImageNet-1K and COCO, and even LoRA-based LLMs—RPG achieves performance on par with fully trained networks while avoiding excessive memory overhead. Notably, it generalizes beyond its training set to generate valid parameters for previously unseen tasks, highlighting its flexibility in dynamic and open-ended scenarios. By overcoming the longstanding memory and scalability barriers, RPG serves as a critical advance in 'AI generating AI', potentially enabling efficient weight generation at scales previously deemed infeasible.

Synaptic plasticity is widely considered to be crucial to the brain’s ability to learn throughout life. Decades of theoretical work have therefore been invested in deriving and designing biologically plausible learning rules capable of granting various memory abilities to neural networks. Most of these theoretical approaches optimize directly for a desired memory function; but this procedure can lead to complex, finely-tuned rules, rendering them brittle to perturbations and difficult to implement in practice. Instead, we build on recent work that automatically discovers large numbers of candidate plasticity rules operating in recurrent spiking neural networks. Surprisingly, despite the fact that these rules are selected solely to achieve network stabilization, we observe across a range of network models -feedforward, recurrent; rate and spiking- that almost all these rules endow the network with simple forms of memory such as familiarity detection - seemingly by accident. To understand this phenomenon, we study an analytic toy model. We observe that memory arises from the degeneracy of weight matrices that stabilize a network: where the network lands in this space of stable weights depends on its past inputs---that is, memory. Even simple Hebbian plasticity rules can utilize this degeneracy, creating a zoo of memory abilities with various lifetimes. In practice, the larger the network and the more co-active plasticity rules in the system, the stronger the memory-by-accident phenomenon becomes. Overall, our findings suggest that activity-silent memory is a near-unavoidable consequence of stabilization. Simple forms of memory, such as familiarity or novelty detection, appear to be widely available resources for plastic brain networks, suggesting that they could form the raw materials that were later sculpted into higher-order cognitive abilities.

State-Space Models (SSMs) have proven to be powerful tools for online function approximation and for modeling long-range dependencies in sequential data. While recent methods such as HiPPO have demonstrated strong performance using a few polynomial bases, they remain limited by their reliance on closed-form solutions for specific, well-behaved bases. The SaFARi framework generalizes this approach, enabling the construction of SSMs from arbitrary frames, including non-orthogonal and redundant ones, thus allowing an infinite diversity of possible "species'' within the SSM family. In this paper, we introduce WaLRUS (Wavelets for Long-range Representation Using SSMs), a new species of SaFARi built from Daubechies wavelet frames. We instantiate two variants, scaled-Walrus and translated-Walrus, and show that their multiresolution and localized nature offers significant advantages in representing non-smooth and transient signals. We compare Walrus to HiPPO-based models and demonstrate improved accuracy, better numerical properties, and more efficient implementations for online function approximation tasks.

Current methods for initializing state space models (SSMs) parameters mainly rely on the \textit{HiPPO framework}, which is based on an online approximation of orthogonal polynomials. Recently, diagonal alternatives have shown to reach a similar level of performance while being significantly more efficient due to the simplification in the kernel computation. However, the \textit{HiPPO framework} does not explicitly study the role of its diagonal variants. In this paper, we take a further step to investigate the role of diagonal SSM initialization schemes from the frequency perspective. Our work seeks to systematically understand how to parameterize these models and uncover the learning biases inherent in such diagonal state-space models. Based on our observations, we propose a diagonal initialization on the discrete Fourier domain \textit{S4D-DFouT}. The insights in the role of pole placing in the initialization enable us to further scale them and achieve state-of-the-art results on the Long Range Arena benchmark, allowing us to train from scratch on very large datasets as PathX-256.

Recent advancements in language models (LMs) have marked a shift toward the growing importance of post-training. Yet, post-training approaches such as supervised fine-tuning (SFT) do not guarantee the effective use of knowledge acquired during pretraining. We therefore introduce infosteer, a lightweight method that encourages parametric information utilization in LMs during post-training. Specifically, Infosteer treats the feed-forward network (FFN) layer as associate key-value memory and promotes the use of stored memory vectors via forward-pass interventions or regularization during backpropagation. This simple guidance during post-training phase yields consistent performance improvements across diverse model families--including Qwen, Gemma and Llama---spanning 15 downstream tasks in both in-distribution (ID) and out-of-distribution (OOD) evaluations. Beyond performance gains, we also find that steered LMs can adaptively allocate information by placing more emphasis on generating semantically meaningful tokens, while using fewer resources on simple transition ones (e.g., ',' or 'and'). Our work underscores that vanilla post-training does not fully exploit the potential gained during pre-training, and that steering LMs in latent representation space offers a promising approach to enhance both performance and interpretability.

Neural networks are powered by an implicit bias: a tendency of gradient descent to fit training data in a way that generalizes to unseen data. A recent class of neural network models gaining increasing popularity is structured state space models (SSMs). Prior work argued that the implicit bias of SSMs leads to generalization in a setting where data is generated by a low dimensional teacher. In this paper, we revisit the latter setting, and formally establish a phenomenon entirely undetected by prior work on the implicit bias of SSMs. Namely, we prove that while implicit bias leads to generalization under many choices of training data, there exist special examples whose inclusion in training completely distorts the implicit bias, to a point where generalization fails. This failure occurs despite the special training examples being labeled by the teacher, i.e., having clean labels! We empirically demonstrate the phenomenon, with SSMs trained independently and as part of non-linear neural networks. In the area of adversarial machine learning, disrupting generalization with cleanly labeled training examples is known as clean-label poisoning. Given the proliferation of SSMs, we believe that delineating their susceptibility to clean-label poisoning, and developing methods for overcoming this susceptibility, are critical research directions to pursue.

Video Large Language Models (Video-LLMs) excel at understanding videos in-context, assuming full access to the video when answering queries. However, these models face challenges in streaming scenarios where hour-long videos must be processed online, and questions need timely responses. In this work, we propose a training-free approach compatible with standard Video-LLMs, leveraging three key concepts: 1) LLM-informed selection of visual tokens to identify those that the LLM has attended to and contributed to its understanding of each short clip. Our attention-based selection allows us to discard up to ~95\% of unimportant visual tokens with minimal performance loss; 2) Hierarchical selection of tokens combined with natural language understanding of each processed clip; 3) Caption-based question answering for lightweight and accurate responses. Our method achieves state-of-the-art performance on streaming video benchmarks, striking a balance between efficiency and effectiveness.

Stochastic optimal control methods often struggle in complex non-convex landscapes, frequently becoming trapped in local optima due to their inability to learn from historical trajectory data. This paper introduces Memory-Augmented Potential Field Theory, a unified mathematical framework that integrates historical experience into stochastic optimal control. Our approach dynamically constructs memory-based potential fields that identify and encode key topological features of the state space, enabling controllers to automatically learn from past experiences and adapt their optimization strategy. We provide a theoretical analysis showing that memory-augmented potential fields possess non-convex escape properties, asymptotic convergence characteristics, and computational efficiency. We implement this theoretical framework in a Memory-Augmented Model Predictive Path Integral (MPPI) controller that demonstrates significantly improved performance in challenging non-convex environments. The framework represents a generalizable approach to experience-based learning within control systems (especially robotic dynamics), enhancing their ability to navigate complex state spaces without requiring specialized domain knowledge or extensive offline training.