NeurIPS 2026 论文列表

Do behavioral styles in large language models correspond to reusable directions in residual space, or do they mainly appear after averaging many noisy contrastive examples? We introduce MinSteer, a two-stage cached-continuation method for estimating a contrastive residual direction (CRD) within one decoding trajectory. Stage-X generates a continuation in one style; a style-flip suffix is then processed using Stage-X's final KV cache; Stage-Y continues from this inherited trajectory. The CRD is computed from generated-token residuals only, excluding prompt and suffix tokens from the average. Across sentiment transfer, politeness control, and a ParaDetox-based bidirectional diagnostic, MinSteer extracts effective steering directions from very small contrasts. On TweetEval sentiment transfer, one MinSteer contrast outperforms a CAA baseline averaged over 50 independent pairs. Japanese keigo CRDs transfer to English and Chinese outputs and outperform trajectory-free TwoSeq controls. In ParaDetox, MinSteer produces bidirectional Detoxify movement under our rewrite setup, while exposing semantic-preservation trade-offs. The directions transfer across tested scenarios and languages and show clear depth dependence. Together, these results show that cached self-rewrite trajectories provide a cleaner signal for residual-space behavior directions than independent-pair averaging.

Learning high-quality representations of mixed-integer linear programming (MILP) is crucial for advancing machine learning (ML)-based methods, yet remains challenging due to the discrete nature of MILPs and their extreme sensitivity to small perturbations. Although contrastive learning has proven highly effective for representation learning in domains such as computer vision, designing effective data augmentations for MILPs is particularly difficult. To address this challenge, we propose MIRAGE, a principled data augmentation framework grounded in the invariance of the piecewise-affine dual price function induced by the branch-and-bound (B&B) algorithm. Viewing this dual function as a recorded, solver-level surrogate of the MILP's lower-bound structure, we derive two complementary augmentation certificates that act as its right-hand and objective-side projections. Both certificates are provably invariant under the same dual price function on the recorded B&B tree, and therefore preserve either the structural similarity of the search tree or the optimality of the original optimal solution. Empirically, we integrate MIRAGE with ML models for two significant downstream tasks—learning to branch and predict-and-search—and observe consistent performance improvements across multiple benchmarks. By grounding MILP data augmentation in a single, solver-induced duality quantity, our work provides a principled and effective approach to MILP representation learning.

Spatio-temporal point processes (STPPs) represent a random collection of points where each point corresponds to the time and location of an event. STPPs are widely used to describe a wide range of phenomena, from earthquakes to wildfires to crime occurrence to disease outbreaks. Generative models have recently emerged as a new powerful paradigm for STPPs, demonstrating %due to their exceptional highly competitive generalization capabilities and promising potential for systematic uncertainty quantification. However, existing generative approaches primarily rely on unimodal numerical data, overlooking the geographic context in which events occur and failing to capture critical spatial and temporal patterns within the STPP. To address these limitations, we propose Morse-aware Cross-Modality learning within Diffusion Model, or \textbf{MCM-DM}. MCM-DM represents a cross-modality diffusion framework based on the discrete Morse and cobordism theories that augments STPP modeling with geographic scene understanding. MCM-DM consists of three key components: (i) a vision-language model-based geographic encoder that extracts semantic embeddings from satellite tiles at each event location; (ii) an attention-based fusion mechanism to integrate critical structure representation with spatio-temporal embeddings, and (iii) a Morse-theoretic topological aligner which aligns latent representations of critical events with spatio-temporal embedding space. We demonstrate the MCM-DM utility on 8 diverse STPP datasets from Earth sciences, epidemiology, urban mobility, and crime analytics, highlighting its cross-domain versatility. The code is available at~\url{https://anonymous.4open.science/r/MCM-DM-CF83}.

This paper examines the role of Novel View Synthesis (NVS) in geometric representation learning. In principle, NVS should reason about 3D scene structure, thereby enabling transferable multi-view geometric representations. Yet, existing encoder-based NVS methods yield poor representations. This is not because of lack of supervisory signal, but rather due to inconspicuous architectural choices: spatially expressive decoders that dilute representational capabilities of the scene encoder, and low-level pixel-space targets that hinder feature learning. We present SNAP, a self-supervised encoder-decoder transformer that addresses both through a pose-conditioned local decoder and a latent-space reconstruction objective. SNAP is general purpose, and we show that it is on par or better than special-purpose geometry-supervised methods. SNAP also outperforms self-supervised representations across five geometric tasks: visual localization, pose estimation, point correspondence, depth estimation, and robot manipulation. Remarkably, SNAP's patch tokens exhibit emergent viewpoint invariance that rivals heavily supervised models despite lower compute and data budgets. Under severe camera shifts where standard 2D representations collapse, SNAP maintains strict spatial consistency, revealing that restricting decoder expressivity actively prevents the suppression of transferable geometric structure.

Modeling hyperspectral imagery (HSI) across different sensors presents a fundamental challenge due to variations in wavelength coverage, band sampling, and channel dimensionality. As a result, models trained under a fixed spectral configuration often fail to generalize to other sensors. Existing Vision Transformer (ViT) approaches either rely on implicit spectral modeling with fixed channel assumptions or adopt explicit spatial–spectral attention with prohibitive computational cost, leading to a fundamental trade-off between efficiency and expressiveness. In this work, we introduce Low-rank Efficient Spatial–Spectral ViT (LESSViT), a sensor-flexible architecture for cross-spectral generalization. LESSViT is built on LESS Attention, a structured low-rank factorization that models joint spatial–spectral interactions through separable spatial and spectral components, reducing the complexity of full spatial–spectral attention from $\mathcal{O}(N^2 C^2)$ to $\mathcal{O}(rNC)$, where $N$ is the number of spatial tokens, $C$ is the number of spectral channels, and $r$ is the rank of the low-rank approximation. We further incorporate channel-agnostic patch embedding and wavelength-aware positional encoding to support flexible spectral inputs. To enable efficient and robust pretraining, we introduce a hyperspectral masked autoencoder (HyperMAE) with decoupled spatial–spectral masking and hierarchical channel sampling. We evaluate LESSViT under a cross-spectral generalization setting that simulates cross-sensor variability. Experiments on the SpectralEarth benchmark demonstrate that LESSViT improves robustness under spectral shifts while remaining competitive in-distribution, and explicit and efficient spatial–spectral modeling is essential for scalable and generalizable hyperspectral representation learning.

Concept erasure is a prominent approach to achieving fairness in machine learning through the removal of sensitive attributes from learned representations. Prior concept erasure methods typically define debiasing indirectly through the failure of a chosen adversary, probe, or fairness penalty, making the target of erasure dependent on a particular decoder family or optimization setup. Rather than depending on a particular architecture, task type, or training paradigm, we introduce KINDER, which aims to directly remove sensitive-attribute information from a model's intermediate representations and is compatible with any deep model that contains an intermediate feature space. KINDER operates in a random Fourier feature space, where it estimates a sensitive subspace from protected group prototypes and projects representations onto its orthogonal complement, thereby forming an explicit representation cleaning mechanism. We conduct extensive experiments across diverse settings and show the effectiveness of KINDER on (i) unimodal and multimodal datasets, (ii) supervised and self-supervised settings, (iii) classification, regression, and image segmentation tasks, and (iv) diverse data modalities, including visual, textual, and tabular data.

The linear representation hypothesis (LRH) holds that high-level concepts are encoded as linear directions in a model's feature space, and grounds much of mechanistic interpretability, from probing to sparse autoencoders to steering vectors. We argue the standard test for it is too lenient. A high linear-probe accuracy shows that a concept is linearly readable from the representation, not that it is one of the representation's privileged directions of variance, which is what the LRH actually claims. These two readings come apart in any high-dimensional, richly structured representation, and the gap matters. The structure that linear probing ignores is precisely where compositional generalization, out-of-distribution behavior, and the validity of post-hoc feature decomposition live. We propose three asymmetric diagnostics that operationalize the strong reading: reverse predictivity, spectral concentration, and out-of-distribution generalization. We validate them in a fully observed simulation where ground truth is known, then apply them to pretrained vision encoders on SVG-World, our deterministic generative benchmark with paired visual styles. Pretraining produces structure beyond what a forward probe can certify, but the linear-feature reading systematically overstates how much. Probing on its own is not enough to license the mechanistic claims that rest on it.

Mutual information (MI) is fundamental to representation-learning analysis in deep networks, but in high dimensions MI estimators depend on hyperparameters --- bin widths, kernel bandwidths, neighbourhood radii, auxiliary networks --- whose bias often dominates the signal. We show that continuous piecewise-linear (CPWL) networks, including ReLU networks, admit an intrinsic information-theoretic analysis that avoids this pitfall: their inherent partition of the input space into convex linear regions assigns to every input a discrete region label $\Pi$ alongside the continuous representation $T$. We decompose $I(Y; T)$ exactly into geometric terms, where the \emph{routing information} $I(Y; \Pi)$ admits a hyperparameter-free plug-in estimator that approximates a population lower bound on every other MI term in the decomposition. When the partition saturates, a functional-equivalence quotient at relative Frobenius tolerance $\varepsilon \in [0, 2]$ collapses regions implementing the same linear operator and restores informativeness. Empirically, the routing estimator correlates strongly with three established MI baselines {\it without any hyperparameter}, and the functional quotient at moderate $\varepsilon$ recovers informative estimates in the regime where the raw partition is bounded by $\log_2 N$ rather than $H(Y)$. We provide a parameter-free framework for analysing the internal information geometry of CPWL networks, grounded in their inherent piecewise-linear structure rather than in external discretisation.

Hypergraphs record multi-way interactions among entities. Extracting information from the combinatorial structure underlying observed multi-way interactions is a central task in many real-world problems. Existing methods face several limitations. First, many deep architectures for hypergraphs do not explicitly exploit the potential low-rank structure, which can sacrifice parsimony and interpretability in the learned representations. Second, many low-rank-based methods operate on tensor representations, which typically require hyperlinks to have uniform sizes and thus limit their applicability to general hypergraphs with non-uniform hyperlink sizes. Third, many methods ignore the fact that hyperlinks often arise from heterogeneous mechanisms. For example, medical symptoms may co-occur in the profiles of patients with very different conditions, and such heterogeneity should be incorporated into the learning process. In this work, we develop a general framework for hypergraph representation learning using hyperlink random effects while exploiting the low-rank structure in hypergraphs. The proposed framework accommodates latent heterogeneity in hyperlink formation while preserving entity interaction patterns. We establish identifiability of the model parameters and leverage an expectation-maximization strategy for estimation. The framework allows flexible specifications for the hyperlink random effects; in this paper, we study three choices: categorical, Gaussian mixture, and score-based effects, and develop corresponding estimation algorithms. Through simulation studies, we demonstrate the effectiveness of the proposed method in recovering latent structure and capturing heterogeneous interaction patterns. Empirical studies on real-world hypergraph datasets further illustrate the practical utility of our approach.

While large language models (LLMs) appear to be increasingly capable of solving compositional tasks, it is an open question whether they do so using compositional mechanisms. In this work, we investigate how feedforward LLMs solve two-hop factual recall tasks, which can be expressed compositionally as $g(f(x))$. We first confirm that modern LLMs continue to suffer from the "compositionality gap", i.e. their ability to compute both $z = f(x)$ and $y = g(z)$ does not entail their ability to compute the composition $y = g(f(x))$. We then decode residual stream representations and identify two processing mechanisms: one which solves tasks *compositionally*, computing $f(x)$ along the way to $g(f(x))$, and one which solves them *directly*, without any detectable signature of the intermediate variable $f(x)$. Finally, we find that embedding space geometry is strongly related to which mechanism is employed, where the idiomatic mechanism is dominant when tasks are represented by translations from $x$ to $g(f(x))$ in the embedding spaces.

Hierarchical graph-level tasks that require reasoning over substructures that extend beyond local neighborhoods are posing challenges for standard Graph Neural Networks (GNNs). While graph pooling aims to address this by coarsening graphs into hierarchical representations, existing methods often fail to explicitly preserve the substructures they induce, leading to inconsistent empirical performance. Moreover, the evaluation of such approaches has proved difficult due to the lack of benchmarks with explicit hierarchical structure. We propose Pooled-Substructure-Aware GNNs, an approach to graph pooling that explicitly preserves and processes substructures uncovered by structure-based pooling schemes. We show that this design strengthens pooling in terms of Weisfeiler-Lehman (WL) expressivity. To enable principled evaluation of pooling-based models, we introduce two new synthetic datasets with controllable hierarchical structure and propose a metric to characterize the extent of hierarchical patterns in real-world graph-level benchmarks. On tasks with relevant hierarchical structure, our approach outperforms matched pooling baselines, while remaining competitive on structurally simpler benchmarks.

Existing multimodal large language models (MLLMs) typically formulate object detection as a conditional sequence generation task. Under this paradigm, spatial locations are commonly represented by either textual coordinates or special quantized learnable tokens. However, textual coordinates are tokenized as unrelated symbols, spatially adjacent points (e.g., $x=199$ vs. $y=200$) yield entirely different tokens, so the cross-entropy loss is misaligned with true spatial distance. Special quantized location tokens, on the other hand, inevitably introduce quantization error. Reducing this error requires linearly more tokens, but a larger vocabulary is harder to train and converges slower. We argue that both limitations stem from the formulation of spatial representation. To address this issue, we propose a simple yet novel localization representation method, HiLoc, which predicts locations in a coarse-to-fine manner with special tokens: first a grid index for coarse localization, then a cell index to refine the position within the grid. Under this paradigm, accuracy can be scaled up by simply adding the hierarchy level. We also build HiRef-500K, a dataset consisting negative and complex referring samples to strengthen rejection and complex referring ability. Experiments show that our method achieves strong performance on both detection and grounding benchmarks, especially on small objects, under high IOU thresholds and complex referring tasks.

In this work, we investigate how the internal representations of Large Language Models (LLMs) change with inputs of increasing difficulty. Although sparsity changes may not be obvious for individual samples, dataset-level results reveal a consistent statistical trend: as task difficulty increases (e.g., harder questions, longer contexts, or more answer choices), the last hidden states of LLMs become systematically sparser. In short, \emph{the harder the task, the sparser the representation}. This sparsity--difficulty relation is observable across diverse models and domains, suggesting that harder inputs drive more concentrated activation patterns in the last hidden state. Through a series of controlled analyses and learning-dynamics experiments, we show that representational density is a learned property of data familiarity: models develop rich, distributed representations for mastered patterns, while unfamiliar inputs default to sparser activations. Our finding is also actionable. We illustrate this with Sparsity-Guided Curriculum In-Context Learning (SG-ICL), which uses sparsity to select few-shot demonstrations matched to the query's difficulty, outperforming standard CoT and Auto-CoT baselines on MATH-500. Our study provides new insights into how LLM representations reflect task difficulty and how this signal can be leveraged for inference.

Deep reinforcement learning agents progressively lose representational capacity during training: neurons become dormant, removing active capacity from the network, and effective rank collapses, leaving surviving neurons redundant. Existing remedies such as periodic resets, and special neural network architectures, are largely algorithm- or domain-specific. We propose a simple architectural fix, the Hadamard Representation (HR), which replaces a standard hidden layer with the element-wise product of two independently parameterized layers. HR operates through two complementary mechanisms. First, it reduces the probability of a neuron becoming dormant, which is particularly valuable for continuously differentiable activations such as $\tanh$: unlike dormant ReLU neurons, which are effectively pruned, saturated $\tanh$ neurons silently corrupt downstream layers by turning their outgoing weights into fixed biases. Second, independently of dormancy, the multiplicative structure captures richer feature interactions and increases effective rank without widening the layer. We evaluate HR across five algorithms and three domains: DQN, PPO, and PQN on pixel-based discrete-action Atari, SimbaV2 on state-based continuous control, and MR.Q on visual continuous control. HR consistently improves performance over the strong baselines without any hyperparameter tuning, and gains persist against parameter-matched wider variants, ruling out parameter count as an alternative explanation.

Precise localization of recording electrodes is fundamental to systems neuroscience; yet, current methods depend on labor-intensive post-hoc histology, which is incompatible with real-time feedback or chronic implants. We test the hypothesis that the electrical signals being recorded themselves carry sufficient anatomical information to localize their recording site in 3D atlas coordinates, without the need for histology. To probe this, we introduce a self-supervised framework that learns geometry-aware representations of single-channel local field potentials (LFP), leveraging probe channel geometry as a self-supervisory signal. We compare speech-based SSL objectives (e.g., masked prediction) against geometry-aware objectives that constrain the embedding space to mirror physical channel layout along the probe. These representations outperform masked-prediction baselines, and the gains compound at larger data scales. We test downstream analysis with 3D coordinate regression and brain region classification, benchmarking three SSL backbones (Wav2Vec 2.0, Whisper, and Data2Vec) against supervised baselines (AnyNet, ViT, and classical spectral classifiers) across datasets, laboratories, species, and probe technologies. Our method achieves state-of-the-art results in brain region classification and 3D coordinate regression from raw single-channel LFP, outperforming all baselines, affirming our hypothesis.

Self-supervised gait representation learning aims to learn transferable representations that generalize to unseen scenarios. However, existing methods rely solely on visual self-supervision, making it difficult to assess the semantic value of individual unlabeled sequences and limiting their ability to exploit latent semantic cues. To address this limitation, we introduce language priors into self-supervised gait pretraining. We construct GaitLP-1M, a million-scale silhouette-based gait pretraining dataset without identity labels. A two-stage caption generation pipeline further produces gait descriptions from the corresponding RGB sequences. To the best of our knowledge, GaitLP-1M is the first million-scale silhouette gait dataset paired with gait-centric captions. Building on this foundation, we propose GaitLingo, a language-guided self-supervised framework for gait representation learning. Specifically, we design a Caption-Guided Sample Weighting (CSW) mechanism to prioritize informative samples based on sequence-level semantic richness and attribute rarity. We then introduce a lightweight Temporal Adapter (TA) to better capture motion semantics and improve cross-domain robustness. Finally, we propose Soft Relation Distillation (SRD) to transfer caption-derived relational structures into the visual embedding space. Extensive experiments demonstrate that GaitLingo consistently outperforms prior self-supervised gait pretraining methods in the zero-shot setting across six benchmarks. Compared with GaitSSB, it achieves Rank-1 gains of +12.1% on Gait3D and +7.9% on GREW, as well as an average improvement of +9.1% across four CCPG settings.

Single-Domain Generalized Object Detection (S-DGOD) aims to generalize from a single source domain to unseen domains under distribution shifts. Existing efforts mainly focus on simulating unseen domains or suppressing domain-specific factors. However, they largely overlook the intrinsic generalization capability embedded in Vision Foundation Models (VFMs). In this work, we show that unconstrained fine-tuning disrupts the geometric structure of pre-trained representations, which encodes domain-invariant semantics. Motivated by this, we propose Geometric-Invariance Fine-Tuning (GIFT), a geometric-aware fine-tuning method for VFMs. Specifically, we introduce a series of learnable Householder reflections to the pre-trained weights to encourage geometry-consistent adaptation under domain shift. In addition, we propose a lightweight channel-wise scaling to facilitate effective task-specific adaptation. Extensive experiments demonstrate that GIFT consistently outperforms state-of-the-art (SOTA) approaches on two S-DGOD benchmarks. Notably, by introducing fewer than 1% additional trainable parameters into the frozen backbone, the proposed method achieves improvements of +7.9% and +2.9% in mPC over the previous SOTA on the Cityscapes-C and Diverse Weather Dataset, respectively. This finding highlights that preserving representation geometry is crucial for robust cross-domain generalization, establishing GIFT as a principled and efficient solution to domain shift. The code is available in the supplementary material.

Real-world networks typically exhibit small-world properties: dense local clustering yet efficient global propagation. However, current methods struggle to reconcile this hierarchy: Graph Neural Networks (GNNs) suffer from over-smoothing in long-range modeling, while sequence/state-space models often lack the inductive bias to preserve complex graph topology. To bridge this gap, we propose Graph Encoding Mamba (GEM), a dual-scale architecture for graph-level hierarchical representation learning. First, GEM employs a dual-branch encoder that synergizes the local sensitivity of GNNs with the selective scanning of Mamba, effectively capturing both local topology and global shortcuts. Second, a modularity-aware hierarchical pooling mechanism is designed to retain salient community structures during graph coarsening. Crucially, to mitigate structural redundancy from multi-scale fusion, we incorporate an Information Bottleneck objective to distill task-relevant structural patterns from redundant features, encouraging compact, task-relevant representations and supporting intrinsic interpretability. Extensive experiments on graph-level benchmarks demonstrate that GEM improves performance across biological, social, and brain network datasets. We further provide a lightweight node-level adaptation on long-range benchmarks as a boundary analysis of GEM's transferability. Code is available at \url{https://anonymous.4open.science/r/GEM-C34D}.

Learning expressive representations for Boolean logic circuits is a fundamental challenge at the intersection of graph learning and Electronic $\mbox{Design}$ Automation (EDA). Existing Graph Neural $\mbox{Networks}$ (GNNs) primarily rely on topological message passing, which often fails to capture the strict causal dependencies and discrete functional semantics of logic gates. In this paper, we $\mbox{propose} \textbf{FuncFormer}$, a Graph Transformer that incorporates $\textbf{functional simulation}$ not as a proxy task, but as a fundamental inductive bias directly into the representation learning process. Unlike standard GNNs which typically rely on isotropic aggregation, FuncFormer $\textbf{encodes the intrinsic flow of functional propagation}$ by analyzing randomized simulation traces as they evolve through the network. This approach effectively aligns the continuous embedding manifold with the discrete Boolean function space, effectively mitigating structural aliasing. By integrating these deterministic signal trajectories with a scalable dual-path attention mechanism, our model preserves functional consistency across long-range dependencies in both combinational and sequential circuits. Empirical results demonstrate that FuncFormer significantly outperforms state-of-the-art models (e.g., DeepGate4) in Quality-of-Results (QoR) prediction and formal verification tasks, exhibiting robust generalization to unseen circuit scales.