NeurIPS 2026 论文列表

Recent work on agentic training efficiency has mainly focused on improving learning algorithms, rollout strategies, or model-hardware co-design. However, these approaches overlook an important source of training latency: the sandbox that runs the environments. Recent studies show that creating, executing, and resetting containers accounts for an important portion of the rollout process in RL training, substantially slowing down overall training. This overhead largely stems from the centralized management mechanisms (e.g., DAEMON) of existing sandboxes, along with additional features (e.g., multi-tenancy) that are unnecessary for reinforcement learning (RL). Moreover, existing sandboxes are designed for general-purpose isolated execution, making them a poor fit for the requirements of RL rollout and training. We propose NitroBox, a novel sandbox system purpose-built for agentic RL that significantly accelerates the rollout and training process. Technically speaking, we first comprehensively redesign NitroBox's lifecycle to eliminate the dependency on DAEMON, so that every sandbox instance is fully independent during creation, startup, and reset (e.g., via rootless full-namespace isolation). Furthermore, we introduce a copy-on-write mechanism to enable fast startup across multiple instances. Second, we knock off features that are time-consuming but unnecessary for RL — such as multi-tenancy — while preserving the core functionality and security guarantees needed for RL training. Finally, we introduce some RL-nature features to further improve efficiency, such as optional CRIU-based process checkpoint saving and restoring for mid-trajectory branching. Our experiments show that NitroBox is faster than Docker across the entire lifecycle: 10.6× faster in sandbox creation, 24.6× in reset, and 19.1× in deletion. When training Qwen2.5-Coder-32B-Instruct with GRPO at c=32 concurrency, end-to-end wall-clock training time drops from 40.6 hours to 28.4 hours when replacing Docker with NitroBox, while preserving the same task pass rate.

The foundational capabilities of large language models are acquired during pretraining on internet-scale, highly heterogeneous data mixtures. In this work, we investigate a geometric question regarding the converged state of this pretraining process: Does the model converge to a common minimizer across all data sources (e.g., \cref{fig:cwa_illustration:close}), or merely a minimizer of the averaged loss (e.g., \cref{fig:cwa_illustration:distant})? We hypothesize that the geometric ``closeness'' of task-specific minima is intrinsically linked to downstream generalization. However, we reveal that standard optimizers (e.g., AdamW) often converge to points where task-specific minima are distant from each other. To address this, we propose the \textit{Nexus} optimizer, which encourages the closeness of these minima by maximizing gradient similarity during optimization. Extensive experiments across models ranging from 130M to 3B parameters demonstrate that Nexus significantly boosts downstream performance, despite achieving nearly \textit{the same pretraining loss} (see \cref{fig:demo:benchmark}). Notably, on the 3B model, Nexus reduces the out-of-distribution loss by 0.012 and yields up to a 15.0\% accuracy improvement on complex reasoning tasks (e.g., GSM8k). Our findings challenge the reliance on pretraining loss as the sole proxy for model evaluation and highlight the critical role of optimizer implicit bias in unlocking downstream generalization.

Transformers replace recurrence with a memory that grows with sequence length and self-attention that enables ad-hoc look ups over past tokens. Consequently, they lack an incentive to compress history into compact latent states with consistent transition rules. This often leads to learning solutions that generalize poorly. We introduce **Next-Latent Prediction** (NextLat), which extends standard next-token training with *self-supervised* predictions in the latent space. Specifically, NextLat trains a transformer to learn latent representations that are predictive of its next latent state given the next output token. Theoretically, we show that these latents provably converge to *belief states*, compressed information of the history necessary to predict the future. This simple auxiliary objective injects a recurrent inductive bias into transformers, while leaving their architecture, parallel training, and inference unchanged. NextLat effectively encourages the transformer to form compact internal world models with its own belief states and transition dynamics—a crucial property absent in standard next-token prediction transformers. Empirically, across benchmarks in world modeling, reasoning, planning, and language modeling, NextLat demonstrates significant gains over standard next-token training in downstream accuracy, representation compression, and lookahead planning. Furthermore, NextLat enables *variable-length self-speculative decoding*, accelerating inference by up to $3.3\times$ in the language domain. NextLat stands as a simple and efficient paradigm for shaping transformer representations toward stronger generalization.

Autoregressive video generation has emerged as a powerful paradigm for World Action Models (WAMs). However, existing approaches suffer from slow training convergence particularly at high frame rates, limited converged accuracy, and slow inference due to iterative video denoising, as the training supervision is confined to the current chunk without explicit signals about future dynamics. In this paper, we present Next Forcing, a multi-chunk prediction (MCP) framework for causal world modeling that enables faster training, higher accuracy, and accelerated inference. Inspired by multi-token prediction in large language models, Next Forcing introduces an MCP training objective that augments the main model with lightweight auxiliary MCP modules to simultaneously denoise video chunks at multiple future temporal horizons (next$^1$, next$^2$, next$^3$ chunks). These MCP modules form a causal chain across prediction depths, where intermediate features fused from multiple layers of the main model are leveraged to predict future dynamics, allowing near-future predictions to inform farther-future ones and providing dense multi-scale temporal supervision back to the main model. During training, the MCP modules significantly accelerate convergence and improve converged accuracy, especially at high frame rates. At 50 fps, our method achieves a 93.1\% relative improvement over the baseline LingBot-VA at 5k training steps and achieves 2.3$\times$ faster convergence. At inference, the MCP modules can be retained to predict the next video chunk in parallel with the current one, accelerating generation. Next Forcing establishes new state-of-the-art results on the RoboTwin benchmark (94.1\%/93.5\% on Clean/Random) and demonstrates significant improvements on PhyWorld, a benchmark evaluating adherence to physical laws in video generation.

Scientific machine learning has opened new avenues for solving parameterized partial differential equations (PDEs), enabling models to learn a family of PDEs and generalize to unseen instances. In this context, data-driven operator learning methods typically require large training datasets, while physics-informed neural networks (PINNs) suffer from difficult optimization and limited generalization, especially for nonlinear PDEs. We propose Newton-PINet, a physics-informed network enhanced by Newton linearization, offering an effective meta-learning framework for nonlinear PDEs. Newton-PINet (i) employs a physics-informed multilayer network with skip connections, where the output-layer weights are solved by least squares; (ii) adopts a two-stage learning strategy that first leverages gradient-based training to learn robust representations from the available training tasks, and then performs gradient-free fine-tuning on the output layer for fast task-specific generalization; and (iii) incorporates a Newton linearization method to speed up the least-squares iteration for nonlinear PDE problems. On a challenging nonlinear reaction-diffusion benchmark, Newton-PINet achieves up to three orders of magnitude lower relative error than recent neural solvers, while using 16× fewer training tasks and over an order of magnitude less training time (under 5 minutes versus several hours). This work advances the meta-learning of PINNs toward data-efficient, fast, and generalizable physics solvers. The datasets and code are provided in the supplementary material.

We construct the first coreset of size near-linear in $k$ for fair \kMedian in general metric spaces. Combined with the standard merge-and-reduce framework, our construction also gives the first streaming algorithm for fair \kMedian with space complexity near-linear in $k$. The main technical innovation is a distributional reinterpretation of capacitated clustering that the cost of assigning a dataset to a capacitated center set can be expressed as the earth mover's distance between two discrete distributions. This viewpoint enables us to use tools from optimal transport to analyse uniform sampling. A key ingredient in this analysis is a new family of problem-specific $\eps$-nets for the earth mover's distance.

Test-time compute scaling offers a promising path for improving foundation models beyond training-time data and parameters. For Vision-Language-Action (VLA) models, however, applying this paradigm through online reinforcement learning remains challenging due to two coupled limitations. Candidate chunks are often biased toward the pretrained policy's high-likelihood regions, while sparse outcome-level feedback lacks process grounding, making reward signals weakly discriminative and exploration insufficiently guided. To address these challenges, we propose EXPLORE, a test-time reinforcement learning framework that coordinates Exploitation and Exploration for action-chunk generation and drives policy optimization with physics-grounded process rewards. Through this joint design, EXPLORE broadens the effective action space without abandoning the pretrained prior and derives faithful process-level rewards from dense physical interaction feedback to guide online policy improvement. Experiments on SimplerEnv-WidowX, DexMG, and RoboCasa validate the effectiveness of EXPLORE, with consistent improvements across autoregressive and diffusion-based VLA backbones.

Targeted Minimum Loss-based Estimation (TMLE) is a classical plug-in debiasing methodology that delivers doubly robust estimation and asymptotically efficient inference. Despite its statistical guarantees, each iteration of standard TMLE requires solving a minimization subproblem over the entire dataset, resulting in prohibitive computational and memory costs that render the method impractical for large-scale or real-time settings. To ease the bottlenecks, we introduce stochastic TMLE, a randomized targeting procedure that replaces each expensive full-batch fluctuation fit with mini-batch alternatives. Our theoretical analysis establishes that the stochastic iterates converge to a neighborhood (noise ball) centered around the target solution, and crucially, only a small number of subsequent full-sample TMLE iterations suffice to reach an empirical efficient influence function root. Consequently, our stochastic variant inherits the same guarantees and attractive properties as classical TMLE while substantially reducing per-iteration complexity. We further propose two computational variants with provable guarantees that broaden the algorithmic design space, offering flexibility for future developments in scalable targeted learning. Extensive experiments across diverse regimes demonstrate substantial acceleration over standard TMLE without sacrificing inferential quality.

Structural descriptors provide graph neural networks (GNNs) with positional and topological inductive biases. However, widely used classical encodings based on random walks or Laplacian spectra emphasize specific structural statistics of a graph, leaving room for alternative descriptors. Neutral-atom quantum processing units (QPUs) can realize graph-dependent Rydberg Hamiltonians, allowing graph topology to be probed through dynamical correlation measurements. In this work, we use these measurements as quantum graph features (QGFs), providing a complementary, hardware-derived family of structural encoding that --- as we demonstrate empirically --- is not accurately described by classical features derived from Laplacian or adjacency matrix. By appropriately selecting the quench parameters of the Hamiltonian, we access two characteristic correlation regimes with distinct structural signatures: antiferromagnetic-like and ferromagnetic-like. On an intrinsic information benchmark of 15 topological and spectral probing tasks, quantum and hybrid quantum-classical features match or surpass classical descriptors on local topological properties. When supplied to message-passing, attention, and graph-transformer backbones, hybrid quantum-classical features yield small improvements that are within cross-fold noise on most cells but sign-consistent across architectures. When comparing noiseless emulation-derived quantum features with an end-to-end neutral-atom QPU implementation, the downstream hybrid gain over the classical baseline persists on five out of six backbones.

Neurosymbolic learning can use symbolic rules to provide supervision for latent concepts from weak labels, but it commonly assumes that the entities referenced by these rules are already specified. Object-centric models decompose images into slot-like representations; however, such slots are not necessarily aligned with the predicates required for symbolic reasoning. We investigate object-centric neurosymbolic learning under distant supervision, where the object-level arguments of a logic program are learned directly from images using only global task labels. We introduce DeepObjectLog, a probabilistic neurosymbolic model that integrates a slot-based perceptual encoder with a probabilistic logic layer. The encoder predicts objectness and class probabilities for candidate object representations, while the logic layer marginalizes over latent objectness and class assignments to compute the likelihood of the observed label. This formulation provides a differentiable task-level learning signal for object-centric perception without requiring per-object labels, masks, bounding boxes, or heuristic set matching. Evaluations across diverse visual reasoning tasks demonstrate that DeepObjectLog achieves superior out-of-distribution generalization to compositional, object-count, and rule shifts compared to neural object-centric and standard neurosymbolic baselines.

Claim verification is an important problem in high-stakes settings, including health and finance. When information underpinning claims is incomplete or conflicting, uncertain answers may be more appropriate than binary true or false classifications. In all cases, faithful explanations of the considerations determining the final verdict are crucial. We introduce inference-time argumentation (ITA), a trainable neurosymbolic framework for ternary claim verification in which a formal argumentation semantics giving the strength of claims is used both (i) to guide LLM training as models learn to generate arguments and assign them base scores (representing intrinsic strengths) and (ii) to compute ternary (true/false/uncertain) predictions from generated, scored arguments. As a result, at training time, argument generation and scoring can be optimised according to the quality of the induced argumentative predictions. Moreover, at inference time, the final prediction is faithful, by construction, to the arguments and scores determining the verdict, rather than being justified by a potentially unfaithful post-hoc reasoning trace as in conventional reasoning models. We finally show that, on two datasets for ternary claim verification, ITA improves upon argumentative baselines and can perform competitively against non-argumentative direct-prediction baselines, while providing verdicts that are computed deterministically from explicit, inspectable argumentative structures.

Individual neurons exhibit diverse molecular, morphological, and electrophysiological properties that shape their distinct roles in neural circuits. While these multimodal features are largely shared across individuals, they remain an underutilized source of information in analyzing neural population dynamics. We posit that these shared biological profiles provide a common reference for relating neural dynamics across subjects and recording sessions. Our framework leverages this structure by routing neurons into latent groups based on their biological profiles and analyzing population dynamics through group-level representations. Because these assignments are anchored in biological properties that are consistent across subjects, the learned organization supports transfer to unseen animals without retraining or neuron-wise correspondence. We evaluate our framework on whole-brain calcium imaging in larval zebrafish and large-scale electrophysiology in the mouse visual cortex. Across both species and recording modalities, the framework delivers state-of-the-art performance in both neural decoding and manifold analysis: it achieves superior decoding of behavior and stimulus labels while better preserving the temporal consistency and local geometric structure of latent trajectories on unseen data. Our experiments demonstrate that not only does providing biological features improve baseline performance, but our proposed framework achieves even stronger results by using these properties to explicitly structure the population rather than treating them as auxiliary inputs. Together, these results suggest that organizing neural populations by their multimodal biological features offers a promising framework for generalizing population dynamics across subjects and recording sessions.

We investigate whether neuron populations that shape the internal organization of neural networks evolve predictably with scale, extending scaling laws beyond macroscopic observables such as loss. To probe this question, we study *Rosetta Neurons*, a previously characterized class of neurons whose activation patterns are similar across independently trained models (Dravid et al., 2023). In separate analyses of language models up to 30B parameters and vision models up to 5B parameters, we observe that the population of Rosetta Neurons follows a sublinear power law in model size, growing in absolute number but occupying a shrinking fraction of the total neuron count. We further observe a *Neuron Polarization Effect*: Rosetta Neurons become more selective and increasingly monosemantic with scale, separating from a growing non-Rosetta population that remains less selective. An analytical model balancing feature utility against limited neuron capacity explains the sublinear power-law scaling and this polarization effect. Finally, we find that Rosetta Neurons become more domain-specialized with scale and illustrate their selectivity through a targeted data-filtering case study for continued pretraining. Our results point to a scaling law for shared neuron-level structure, linking model size to systematic changes in neuron universality, selectivity, and specialization.

Many physical systems exhibit a low-dimensional structure that varies with external parameters: link lengths in a robot, forcing constants in a fluid, or Reynolds numbers in a flow shift the underlying manifold while preserving its intrinsic dimension. Constrained AutoEncoders (cAEs) learn such manifolds through an *idempotent* encoder-decoder projection, a property that unconstrained autoencoders cannot match and that is essential whenever the model is applied iteratively. However, the standard strategies for making a cAE context-dependent, namely concatenating the context to the input or affinely modulating hidden activations, break the encoder-decoder idempotency, sacrificing the projection guarantee precisely in the setting where it would be most valuable. To restore this guarantee under context variation, we developed the Neuromodulated Constrained Autoencoder (NcAE), which modulates the activation slope and bias of a cAE through a context-driven hyper-network. This paper presents the NcAE, its theoretical foundation, and its empirical validation. We prove that for every context, including contexts unseen at training time, the reconstruction map remains an idempotent projection, the topology of the learned manifold is invariant, and context perturbations induce smooth changes in the manifold. We evaluated our approach on a 16-DoF pendulum with context dependent coupling and the Lorenz96 system across a bifurcation. The NcAE matched or exceeded the best of six baselines on reconstruction, idempotency, and latent-geometry metrics, while being the only architecture that preserves geometric consistency by construction. The NcAE thereby provides a stable, geometry-preserving coordinate system across families of physical regimes.

Reactive synthesis from temporal logic modulo theories (LTLt) automatically con- structs correct-by-construction controllers for systems with arithmetic constraints— elevators, thermostats, robotic patrols—but the state-of-the-art CEGRES algorithm times out on 77% of existing benchmarks due to a fundamental bottleneck: it discovers the universally valid theory tautologies needed to refine its Boolean abstraction one at a time, through expensive counterexample analysis. We present NEUROSYNTHEOS, a neural augmentation of the CEGRES loop that predicts batches of tautologies at each refinement step. We formalize template invariance for parametric specification families, showing that the prediction problem reduces to constant re-instantiation when specifications share algebraic structure. We train a GNN over formula abstract syntax trees and an autoregressive transformer on tautology traces from 104 solved benchmarks. NEUROSYNTHEOS establishes a new state of the art for LTLt synthesis: on 83 standard benchmarks, it solves 60 ±1.2 specifications (vs. 19 for the previous best), resolving 41 that no prior method could handle within a 5-minute budget, while reducing the median itera- tion count by 5.2×with under 3 seconds of neural overhead. All predictions are SMT-verified before injection, preserving full soundness guarantees.

Epilepsy affects approximately $1$% of the global population, where seizure detection and reliable forecasting are critical for patient safety and quality of life. While emerging foundation models increasingly recast EEG analysis as a language modeling task, most remain confined to single-modality inputs and narrow domain generalization. In this work, we present NeuroNTP, a modular multimodal foundation model built on an xLSTM backbone and pretrained on one of the largest retrospective epilepsy datasets ever collected, comprising $150,000$ hours of neural (EEG) and physiological (ECG, SpO$_2$) signals paired with rich clinical text. NeuroNTP employs a flexible, montage-geometry grounded tokenizer and optimizes a composite objective, synergizing causal next-token prediction with seizure-specific auxiliary tasks to enable unified sequence modeling across heterogeneous modalities. This approach allows the model to capture long-range temporal dependencies and cross-modal interactions that unimodal baselines ignore. In extensive experiments, we show that NeuroNTP outperforms supervised baselines by a substantial margin on held-out clinical cohorts. Furthermore, the model generalizes to public EEG benchmarks such as CHB-MIT and Siena Scalp EEG, achieving state-of-the-art performance under standard evaluation settings. We release code and model weights to facilitate further research.

Static-input Spiking Neural Networks (SNNs) commonly create temporal dynamics by repeating the same image over multiple timesteps. For handwritten text recognition, however, this convention is input-redundant: the image contains no native event stream, while recognition depends on faint stroke fragments, degraded ink boundaries, and monotonic left-to-right alignment. We introduce NeuroInk, a retinomorphic spike-oriented recognizer that converts static handwriting into a low-timestep sequence of temporally ordered ink evidence. Instead of replaying identical frames, NeuroInk performs temporal ink transduction, decomposing each text-line image into coarse stroke-mass support and fine stroke-boundary evidence. The resulting evidence sequence is injected as time-varying currents into an adaptive spiking visual hierarchy, where high-frequency enhancement and mid-level detail reinjection are used to preserve fine ink structures during spiking propagation. For CTC decoding, two-dimensional spiking maps are read out as width-wise columnar sequences and processed by CTC-aligned association circuits that combine local horizontal context, global association, and selective trace memory. On the LAM benchmark, NeuroInk achieves 2.4\% CER on the validation split and 2.6\% CER on the test split using only two timesteps and greedy CTC decoding, without an external language model or lexicon. These results support the view that static-input SNNs sequence recognition can benefit from co-designing sensory temporalization, spiking visual abstraction, and monotonic sequence association.

Large Language Models (LLMs) can generate plausible free text self-explanations to justify their answers. However, these natural language explanations may not accurately reflect the model's actual reasoning process, indicating a lack of faithfulness. Existing faithfulness evaluation methods rely primarily on behavioral tests or computational block analysis without examining the semantic content of internal neural representations. This paper proposes NeuroFaith, a flexible framework that measures the faithfulness of LLM free text self-explanation by identifying key concepts within explanations and mechanistically testing whether these concepts actually influence the model's predictions. We show the versatility of NeuroFaith across 2-hop reasoning and classification tasks. Additionally, we develop a linear faithfulness probe based on NeuroFaith to detect unfaithful self-explanations from representation space and improve faithfulness through steering. NeuroFaith provides a principled approach to evaluating and enhancing the faithfulness of LLM free text self-explanations, addressing critical needs for trustworthy AI systems.

Foundation models (FMs) promise to extract unified representations that generalize across downstream tasks. They have emerged across fields, including electroencephalography (EEG), but it is less clear how effective they are in this particular field. Published evaluations differ in datasets, in the EEG-specific preprocessing that might influence reported results, and in the reported metrics, frequently obscuring the clinical relevance in EEG. We introduce NeuroAtlas, the largest EEG benchmark to date: 42 datasets and $\sim$260k hours covering clinical EEG (epilepsy, sleep medicine, brain age estimation) and brain-computer interfaces, and include multiple datasets per task along with bespoke clinical evaluation metrics. Besides evaluating EEG-FMs with respect to supervised baselines, we present results from generic time-series FMs. We report three findings. First, EEG-specific FMs do not consistently outperform time-series FMs, which have neither EEG-focused architectures nor been pretrained on EEG. Second, standard machine learning metrics are insufficient to assess clinical utility: thus, we thoroughly evaluate more appropriate measures such as the quality of event-level decision-making, hypnogram-derived features, and the brain-age gap in the domains of epilepsy, sleep, and brain age, respectively. Third, model rankings and performance can vary substantially within domains. We conclude that pretrained models perform largely on par, with only narrow advantages for a few, and that current models do not yet deliver on the promise of an out-of-the-box unified EEG model. NeuroAtlas exposes this gap and provides the datasets and metrics for the next generation of unified EEG FMs.