NeurIPS 2026 论文列表

Gradient-based attribution methods for large language models (LLMs) have so far been limited to explaining scalar outputs - the logit of a single target token. Yet LLM predictions are inherently distributional, and many practically important questions concern the full predictive distribution: which tokens make the model uncertain? Which drive a broad range of plausible continuations? We propose **Jacobian Scopes**, a unified framework that fills this gap by projecting the input-to-output Jacobian onto different directions in output space via a single vector-Jacobian product, requiring only one backward pass. This yields three complementary methods: **Semantic Scope** attributes a specific target logit; **Fisher Scope**, grounded in information geometry, identifies tokens that most alter the overall shape of the predicted distribution; and **Temperature Scope** traces which tokens govern the model's predictive confidence. Fisher and Temperature Scopes are, to our knowledge, the first attribution methods to natively target distributional rather than pointwise features of LLM predictions. Through case studies spanning instruction following, translation, and in-context time-series forecasting, Jacobian Scopes reveal implicit political biases, uncover word- and phrase-level translation strategies, and illuminate the nearest-neighbor pattern-matching mechanisms underlying LLM forecasting. Quantitative evaluation on LAMBADA and IWSLT2017 across six leading LLMs (LLaMA-3.2, Qwen2.5, Gemma-3) confirms that Jacobian Scopes consistently match or outperform Input $\times$ Gradient and Integrated Gradients, at a fraction of the latter's computational cost.

Articulated objects used in simulation and embodied AI are typically specified by geometry and kinematic structure, but lack the fine-grained dynamical effects that govern realistic mechanical behavior, such as frictional holding, detents, soft closing, and snap latching. Existing approaches either ignore the detailed structure of dynamics entirely, or use simple models with limited expressiveness. We introduce JODA, a framework for generating joint-level dynamics as a structured three-channel field over the joint degree of freedom, capturing conservative forces, dry friction, and damping. Instantiated using shape-constrained piecewise cubic interpolation (PCHIP), this formulation defines a compact and expressive function space that is both interpretable and compatible with differentiable simulation. Building on this representation, we develop methods for inferring and refining joint dynamics from multimodal inputs. Given visual observations and joint context, a vision-language model proposes structured dynamical primitives, which are composed into a unified dynamics field. The resulting representation supports both direct manipulation and gradient-based refinement. We demonstrate that JODA enables plausible and controllable modeling of diverse joint behaviors, providing a unified interface for inference, editing, and optimization.

Recent one-step generative models accelerate sampling by learning deterministic flow maps of the underlying dynamics. These methods rely on learning from ordinary differential equations, leaving open how to define an exact distillation procedure for stochastic dynamics. We introduce the Itô map, an any-step stochastic flow map that takes an intermediate state and Brownian path and predicts future states in a single pass. The Itô map formulation yields novel estimators for inference-time control by providing cheap, differentiable access to posterior samples. Empirically, Itô maps produce diverse, conditionally valid endpoint samples from fixed intermediate states and support strong steering performance on synthetic and image-generation benchmarks. These results establish any-step SDE integration as a useful primitive for posterior sampling and stochastic control.

Offline goal-conditioned reinforcement learning (GCRL) enables goal-reaching from static datasets, but the learned value, critic, or compatibility score must support both long-horizon reachability inference and offline policy extraction. We study iterative latent refinement as an architecture-level replacement for the feedforward value or critic backbone used by existing offline GCRL algorithms. Instead of emitting a score after a single feedforward computation, the network repeatedly updates a latent representation using shared recurrent update weights and lightweight step-specific parameters before exposing the final actor-facing signal. On the recent benchmark datasets reported in this study, recurrent variants improve performance across several algorithms without changing their losses, data, actors, or training protocol; across the 21 reported algorithm-dataset rows, the mean improvement is $10.7$ percentage points, with the largest gains on stitching and bottleneck maze tasks. CRL diagnostics are consistent with an improved policy-extraction interface: recurrent critics increase normalized separation between matched and mismatched goals and produce actors that remain closer to dataset actions under several behavior-support proxies. Depth and update-capacity studies further suggest that the effect depends on how computation is allocated between recurrent depth and per-step expressivity.

Search-based planning is a powerful policy improvement paradigm in reinforcement learning, with milestone successes on discrete-action domains through the AlphaZero and MuZero families. Extending it to continuous control, however, faces a structural mismatch between the search operator and the policy class: search at each state evaluates a finite candidate set and returns a discrete distribution, while the parametric Gaussian policy commonly used in continuous control is a continuous density. We prove that distilling the search output into a Gaussian induces a non-vanishing projection error that inflates policy variance and produces a return gap that persists even when the policy mean is locally optimal, and is not closed by additional search compute or data. Inspired by the iterative refinement of MPPI in continuous control, we propose \textbf{Iterative Gumbel Planning (IGP)}, which adapts this principle to a discrete policy class. IGP represents the policy as a per-dimension factorized categorical distribution aligned with the search support, eliminating the projection error and reducing the residual approximation to a bounded discretization error controllable by the grid resolution. Joint candidates are drawn by Gumbel Top-$K$ sampling without replacement, avoiding the $V^{|A|}$ combinatorial explosion of joint discretization, and search-refine iteration runs for multiple rounds at the same state. On DMControl and HumanoidBench, IGP achieves state-of-the-art performance, with improved sample efficiency and lower temporal action variance at evaluation.

Current pre-training datasets are derived from web crawls, with all their issues, and were not designed to support mid- and post-training pipelines—for instance, they contain little explicit reasoning. Thus, many frontier labs have begun to develop their own internal datasets, starting from state-of-the-art models, to augment their pre-training data mix, e.g., with reasoning traces to address cold start problems. While demonstratively effective, none of these datasets are public, and the effect of this so-called _synthetic data_ on knowledge and skill acquisition of language models, including small ones, remains poorly understood. We present Synth, the first open-source synthetic corpus derived from 58,698 Wikipedia articles that collapses pre-, mid-, and post-training into a single training stage via structured amplification of curated encyclopedic seeds. We evaluate Synth by training a series of models: a 56M tiny model (Monad), 0.3B--0.6B dense models (Baguettotron), and a 13B-total / 1B-active Mixture-of-Experts. At iso-compute, Synth outperforms filtered web data, and our models remain competitive with similarly-sized open-weight baselines. Because Synth is back-translated from grounded passages, Synth-trained models achieve high factual precision despite 10-140$\times$ fewer training tokens, with memorization targeted by the seed corpus. These results show that synthetic datasets, including our Synth dataset, are capable of producing competitive generalist models at a significantly lower cost, enabling rapid iteration as the frontier advances. These findings open up possibilities for both generalist models with significantly increased data efficiency, as well as domain-specific models where no instruction or conversational data is available. Finally, we publicly release our Synth dataset and the series of Baguettotron models under a permissive license, thus supporting open-source language model development.

Many sequential Bayesian methods maintain a whitened operator $A_w = L^{-1}ML^{-\top}$, where $L$ is the Cholesky factor of a kernel, precision, or curvature matrix. When the underlying matrix changes by a rank-$1$ perturbation, existing techniques propagate only one-sided quantities $L^{-1}B$; refreshing the sandwiched operator $A_w$ requires a full $O(r^3)$ re-factorization that dominates iteration cost, consuming up to 91% of per-step time in natural gradient methods. We show that this cost is unnecessary. For pre-whitened inputs, intermediate factors of $L$ cancel exactly from both sides of the augmented Givens rotations, and the updated $(L^+)^{-1}M^+(L^+)^{-\top}$ can be recovered in $O(r^2)$ time by applying the rotations double-sidedly to a zero-padded matrix, followed by a rank-1 correction. We provide a self-contained derivation, stability bounds, and implementations as a C++ extension, MLX kernel, and Triton GPU kernel. On streaming sparse GP regression, online Bayesian optimization on protein fitness landscapes, and natural gradient preconditioning, the method matches full re-factorization in accuracy with up to $5.4\times$ speedups. At $r = 1024$, training time drops from 1075s to 198s, making larger curvature charts affordable and producing strictly better test MSE.

Current sign language recognition and translation benchmarks are commonly collected in controlled settings using fixed or studio-style cameras, particularly limiting their ability to reflect how signers communicate in everyday mobile scenarios. As such, this gap becomes increasingly evident when considering the Arabic Sign Language (ArSL), which is a primary sign language used across Arab countries. In this work, we introduce Isharah-Selfie, a large-scale ArSL dataset collected entirely using front-facing smartphone cameras under unconstrained selfie conditions. The dataset comprises 8,997 video clips performed by 16 signers across 1,493 unique sentences, covering several domains such as healthcare, education, and transportation. Unlike conventional datasets dominated by fixed-camera, two-handed signing, Isharah-Selfie captures realistic mobile communication characterized by one-handed signing, close-range viewpoints, partial body visibility, camera motion, unconstrained environments, and heterogeneous device resolutions. Each video is annotated with both gloss sequences for continuous sign language recognition (CSLR) and Arabic sentence translations for sign language translation (SLT). Moreover, we define signer-independent and unseen-sentences evaluation protocols to assess generalization across unseen signers and novel sentence compositions. Extensive benchmarks using RGB-based and pose-based data show that current CSLR and SLT methods remain challenged by selfie-style signing conditions, particularly under unseen-sentence evaluation, where compositional generalization remains weak. By releasing Isharah-Selfie, we aim to support research toward robust, user-centered sign language technologies that operate reliably in realistic smartphone-based communication settings. The Isharah-Selfie dataset is available on https://www.github.com/.

Deep learning models are known for their ability to memorize training data. While memorization is often linked to risks such as privacy leakage and poor robustness, a highly influential claim by~\citet{feldman2020longtail} argues that memorization is actually \textit{necessary} for generalization. Their conclusion is based on the observation that removing points with high memorization scores reduces test accuracy. Upon closer inspection of their work, we uncover four critical flaws in the underlying methodology: \textbf{(1) sampling bias} in their approximation algorithm inflates memorization scores; \textbf{(2) high false positive rate} in their definition of memorization, leads to misclassification of non-memorized points as memorized; \textbf{(3) unprincipled thresholding} that resulting in an ill-posed problem; and \textbf{(4) data leakage} skews the test accuracy results. To address these limitations, we introduce a modifications for correctly identifying and evaluating memorization, including higher sampling rates, modifying the original memorization definition to reduce the false positive rates, proposing a method to identify a principled score threshold, and employing test datasets especially designed to avoid data leakage. Having accounted for these errors, our results show that, in contradiction to the original work, removing truly memorized points does not cause a drop in accuracy, and in most cases, improves test performance. These findings call into question the necessity of memorization in deep learning and highlight the importance of mitigating its risks.

Long-tailed out-of-distribution (LT-OOD) detection is often addressed with specialized training, including auxiliary out-of-distribution (OOD) data, abstention heads, contrastive objectives, energy losses, or gradient-conflict control. We show that these training mechanisms can obscure a simpler issue: frozen long-tailed representations may already contain useful OOD evidence, but raw Mahalanobis distance is distorted by frequency-coupled feature radius and poorly supported tail covariance. We propose \emph{Hyperspherical Pooled Mahalanobis} (\HPM), a post-hoc detector that normalizes features onto the unit sphere and replaces class-specific covariance with a pooled, ridge-regularized metric while keeping class means as semantic anchors. In CIFAR-LT experiments and an ImageNet-100-LT near-OOD boundary analysis, \HPM improves raw Mahalanobis scoring; for Prior-Calibrated ERM (PC-ERM), it raises AUROC from 46.49 to 85.67 on CIFAR-10-LT and from 50.40 to 78.35 on CIFAR-100-LT. This simple PC-ERM+\HPM pipeline also achieves the best Log Efficiency Score (LES; 3.08) on CIFAR-100-LT, retaining roughly 95\% of the best CIFAR-100-LT AUROC observed among the compared post-hoc scores at substantially lower training-time cost. These results argue for evaluating representation quality, detector geometry, and training complexity as separate factors in LT-OOD detection.

We introduce **InvestigationWorlds**, an agentic environment for legal investigation. We build on an underused artifact of U.S. civil litigation: the summary judgment motion. This motion relies upon a record composed of real evidence exhibits, and results in a court-adopted hypothesis that is treated as ground truth for the purposes of deciding the motion. Each environment is built from a real U.S. Federal Court case retrieved from Public Access to Court Electronic Records (PACER) and augmented by an attorney-validated generation pipeline that synthesizes role-tagged documents around the original record. The resulting corpus admits multiple coherent factual readings, only one of which matches the court-adopted hypothesis. Evaluating on 100 cases, we find agents often commit to incorrect hypotheses despite retrieving relevant evidence, struggling to distinguish the court-adopted hypothesis from alternative hypotheses.

Post-hoc calibration aligns a classifier's predicted confidences with its empirical accuracy without retraining. An ideal calibrator should correct nonlinear miscalibration, scale gracefully to large label spaces, and preserve the original predictions; existing methods typically violate at least one of these properties---temperature scaling lacks expressivity, more flexible parametric alternatives introduce parameters that grow with the number of classes $C$, and other expressive methods do not preserve the rank ordering of class scores and may alter the predicted class. We propose \textbf{Invertible Logits Transformation (InvLT)}, which applies a learned scalar MLP $f:\mathbb{R}\to\mathbb{R}$ element-wise to the pre-softmax logits. Sharing $f$ across all logit dimensions makes the parameter count independent of $C$. Monotonicity of $f$---and hence preservation of the argmax prediction---is softly encouraged via a paired inverse network rather than enforced through the numerical integration required by prior monotone calibrators; this avoids their computational overhead while empirically preserving the original classification accuracy in every setting we evaluate. Across standard image classification benchmarks and a range of architectures, InvLT consistently outperforms a broad set of post-hoc baselines on standard calibration metrics.

Inferring the latent processes that generate observable neural activity is a central challenge in neuroscience, with direct implications for brain-machine interfaces and neural engineering. Existing inverse approaches typically summarize the recording into low-dimensional features or learn black-box latents that lack a direct biophysical interpretation. A promising alternative is to fit a differentiable biophysical simulator end-to-end against the full extracellular signal, using direct and scalable gradient optimization. In practice, however, recovering a per-neuron input via backpropagation through neuronal dynamics is difficult because the loss landscape is nearly flat in the subthreshold regime and jumps sharply at spike threshold, leaving gradient descent without a useful signal. We observe that per-neuron spike times, routinely available from spike sorting, expose discrete millisecond-scale anchors at which the loss does carry information. Gradients from these anchors flow back in time through the differentiable simulator and shape the subthreshold drive that produced each spike. Building on this, we present a fully differentiable pipeline coupling biophysical membrane dynamics to a volume conductor model of the multi-electrode array. For each neuron in a recorded population, the method jointly recovers a time-varying latent input current and a probe-relative position consistent with both the extracellular trace and the observed spike times. No explicit likelihood or posterior estimation is required, and every fitted latent corresponds to a named biophysical quantity. On paired patch-clamp / Neuropixels SPE-1 data, the model localizes the patched neuron to within $40 \mu m$ on average across cells, while reproducing observed spatio-temporal activity patterns under realistic noise. These results position scalable differentiable biophysical simulation as a practical route to mechanistic, gradient-based analysis of high-density neural recordings.

Operator-valued concentration inequalities are foundational to the analysis of modern high-dimensional statistics and randomized algorithms. However, standard oracle bounds are frequently limited in practice: they require explicit a priori knowledge of the true variance, and often explicitly scale with the ambient dimension, rendering them vacuous for infinite-dimensional or heavily structured operators. Motivated by these challenges, we establish the first empirical Bennett and Bernstein inequalities for sums of independent, bounded, compact self-adjoint operators. Our fully data-driven bounds replace the unknown variance with an empirical estimate and rely strictly on the intrinsic dimension rather than the ambient dimension. This structural shift yields computable, dimension-free guarantees that are strictly sharper for non-isotropic random matrices and seamlessly extend to infinite-dimensional Hilbert spaces. We demonstrate that our empirical bounds achieve asymptotic sharpness with the best known oracle rates. Finally, as an independent byproduct, we derive novel empirical concentration guarantees for the intrinsic dimension itself.

Covariance estimation yields a fundamental second-order statistic underlying representation learning, dimension reduction, and dependence modeling. While covariance has been well understood in Euclidean spaces, it is ill-defined for random objects residing on nonlinear Riemannian manifolds, which increasingly arise in modern machine learning applications involving shapes, symmetric positive definite (SPD) matrices, etc. This paper introduces an intrinsic Riemannian cross-covariance for manifold-valued random objects. Our approach defines covariance and correlation by transporting local variations to a common tangent space via parallel transport, yielding a second-order descriptor that is independent of arbitrary coordinate choices. We establish that the proposed covariance inherits desirable properties of its Euclidean counterparts and characterize its asymptotic behavior. Numerical studies on spheres and SPD manifolds, together with real-data experiments on heart valve shapes in Kendall's shape space, demonstrate the effectiveness and verify the properties. Our results position Riemannian covariance as a fundamental tool for second-order learning and analysis in non-Euclidean representation spaces.

Recent Vision-Language-Action (VLA) models have advanced end-to-end autonomous driving by incorporating reasoning for better interpretability and planning quality. However, most existing approaches directly generate the final trajectory without explicitly examining its future consequences, which limits their reliability in complex and dynamic environments. To address this limitation, we propose **IRR-Drive** (**Intend, Reflect, Refine**), an adaptive multimodal reflection framework for autonomous driving. Specifically, to tightly couple high-level reasoning with physical constraints, IRR-Drive first generates a preliminary textual intention and anticipates potential interactions by predicting future semantic bird's-eye view (BEV) representations. This dual-modality (Text + BEV) reflection space explicitly models anticipated scene evolution, enabling the model to rigorously self-correct and refine its initial intent before generating the final trajectory. Furthermore, to balance planning performance and computational efficiency, we construct reflection-oriented training data and design an adaptive reflection reward, enabling the model to adaptively select its reasoning mode according to scene complexity. Instead of using reasoning primarily as an auxiliary interpretation, IRR-Drive directly integrates an adaptive reflection mechanism into the planning framework, enabling grounded, decision-aware trajectory correction that is driven by scene complexity. Our method achieves state-of-the-art performance on the NAVSIM benchmark in both PDMS and EPDMS. Extensive experiments demonstrate the effectiveness of our multimodal reflection framework and validate the efficacy of the proposed adaptive reflection strategy.

Randomized controlled trials (RCTs) identify treatment effects for an enrolled trial population by assigning treatment independently of potential outcomes, conditional on baseline covariates. These trial-specific effects may not generalize to a broader target population when the target and trial covariate distributions differ. This problem becomes harder when individual-level target population data are unavailable. Pre-trained large language models (LLMs) offer one way to address this data limitation by generating digital twins (DTs) with synthetic covariates intended to resemble the target population. Synthetic covariates alone, however, are insufficient, because neither the RCT covariate distribution nor the DT covariate distribution necessarily matches the target covariate distribution. We therefore propose a statistical inference procedure that integrates RCTs with calibrated DTs using external target population summaries. The procedure represents the target covariate distribution as a calibrated mixture of the RCT and DT covariate distributions, allowing DTs to contribute covariate information while limiting the influence of poorly calibrated synthetic data. The procedure also incorporates LLM-generated auxiliary outcome predictions through calibrated outcome regressions, improving precision without changing the estimand. Theoretical results and simulation studies show that the proposed estimators reduce bias relative to RCT-only estimation and improve efficiency for overall and subgroup target treatment effects.

Exploring high-order correlations and long-range dependencies in brain networks holds significant value for both neuroscience research and clinical diagnosis. However, previous studies have lacked a unified integration of high-order and long-range dependency information in brain networks, and there is substantial redundancy behind various types of information. These issues limit their effectiveness in the diagnosis of brain diseases. To address this, we propose an Information Bottleneck- Guided Adaptive HyperGraph Transformer (IBAHGT). By incorporating the information bottleneck (IB) principle, this approach enables adaptive learning of high-order correlations and both short- and long-range dependencies within a unified framework for brain network analysis, achieving high-precision brain disease diagnosis. IBAHGT consists of three key components: an information bottleneck- guided adaptive hypergraph convolution, which introduces a novel hypergraph information bottleneck (HIB) principle to adaptively learn hypergraph message- passing weights between nodes and hyperedges, optimizes information flow and captures high-order information in brain networks that is maximally informative and minimally redundant (MIMR). The Transformer encoder captures global information within brain networks through the attention mechanism, specifically modeling short- and long-range dependencies. An information bottleneck-guided node-level adaptive fusion employs the IB principle to learn independent weights for each node, facilitating the fine-grained integration of high-order information and global information to obtain an efficient representation for downstream tasks. Extensive experiments demonstrate that the proposed method outperforms current state-of-the-art methods and can identify biomarkers for clinical applications.

Transformers acquire in-context learning abilities through abrupt phases during training, often unfolding over multiple stages in which key circuits, such as induction heads, emerge. In this work, we characterize the dynamics underlying the emergence of such circuits across these stages. We focus on a synthetic associative recall task, where sequences are drawn from random maps between a permutation group and a vocabulary, and the model is required to complete the mapping of a permutation by retrieving the corresponding association from context. For this task, we study the gradient-flow trajectories of a simplified two-layer attention-only transformer. Leveraging symmetries in both the transformer architecture and the data distribution, we derive *closed-form parameter dynamics* for both attention layers. We identify a conservation law that couples parameters across layers and governs their joint learning. This law reveals how initialization controls the *timescales* over which such circuits emerge. In the vanishing-initialization limit, we characterize the gradient-flow trajectory, showing how training jumps from saddle to saddle. Finally, we provide empirical evidence across different architectural choices, validating our simplifications and extending the insights from our analysis beyond the simplified setting.