NeurIPS 2026 论文列表

Direct Preference Optimization (DPO) has been widely adopted for aligning language models with human preferences. Recent empirical work has revealed that DPO induces remarkably sparse parameter updates, yet no theoretical explanation for this phenomenon has been offered. We analyze DPO's optimization dynamics under a neural tangent kernel linearization and show that the logistic structure of the DPO loss gives rise to an implicit maximum-margin bias. Building on this observation, we introduce the concept of dual sparsity: the DPO update direction is a sparse linear combination (data-level sparsity, driven by support vectors) of sparse gradient feature vectors (parameter-level sparsity, arising from gradient cancellation between preferred and dispreferred responses). We formalize this structure in our main theorem and present a framework relating it to weight disentanglement---a condition favorable for model merging. Experiments on GPT-2 Small and Medium with the Anthropic HH-RLHF and UltraFeedback datasets confirm both sources of sparsity and show that DPO reduces cross-task interference by up to $35 \times$ compared to supervised fine-tuning.

A classic result in learning and optimization theory is that exponentiated gradient descent (EG) or the multiplicative weight update method computes an $\\epsilon$-approximate minimizer of an $\\ell_1$-Lipschitz function over the $d$-dimensional probability simplex $\\Delta_d :=\\{x \in \\mathbb{R}^d_{\\geq0}:\\sum_{i\in[d]}x_i=1\\}$ with $\\tilde{O}(\epsilon^{-2})$ queries to a first-order oracle. We investigate whether there are improved parallel algorithms for this fundamental problem. We show that, up to logarithmic factors, the answer is negative for $\\epsilon=\\tilde \\Omega(d^{-1/6})$ - any randomized algorithm that makes $O(\\mathrm{poly}(d))$ subgradient queries per round requires $\\tilde{\\Omega}(1/\\epsilon^2)$ rounds to output an $\epsilon$-optimal point with constant success probability. Moreover, to obtain this result we provide an analogous characterization of the parallel complexity of minimizing a $\ell_1$-Lipschitz convex function over the unit $\\ell_1$ ball. Previous lower bounds for non-constant $\\epsilon$ for this $\\ell_1$-Lipschitz convex optimization problem either made additional assumptions on the domain, or instead attained a bound of $\\Omega(\\epsilon^{-2/3})$ [DG19].

We study two models of $\mathsf{ReLU}$ neural networks: monotone networks ($\mathsf{ReLU}^+$) and input convex neural networks ($\mathsf{ICNN}$). Our focus is on expressivity, mostly in terms of depth and our motivation is to gain better understanding of depth requirement needed to exactly represent functions in terms of neural networks with ReLU activations: a subject that has received a lot of attention lately. We prove several lower bounds on the depth required to represent functions by monotone ReLU networks and $\mathsf{ICNN}$. For the maximum function $\mathsf{MAX}_n$ computing the maximum of $n$ real numbers, we show that $\mathsf{ReLU}^+$ networks cannot compute $\mathsf{MAX}_n$, or even approximate it. We prove a sharp $n$ lower bound on the $\mathsf{ICNN}$ depth complexity of $\mathsf{MAX}_n$. We also prove depth separations between $\mathsf{ReLU}$ networks and $\mathsf{ICNN}$s; for every $k$, there is a depth-$2$ $\mathsf{ReLU}$ network of size $O(k^2)$ that cannot be simulated by a depth-$k$ $\mathsf{ICNN}$. The proofs combine ideas from some structural results for $\mathsf{ReLU}^+$ networks.

We examine generative modelling approaches based on the construction of Schrödinger bridges between Gaussian noise and a target distribution. It is known that the solution of the dynamic Schrödinger problem is a diffusion process with a drift associated with Doob's h-transform of a Schrödinger potential. Although its accurate restoration from finite samples is crucial for reliable, high-quality data generation, the existing literature lacks theoretical guarantees regarding this question. In our work, we establish theoretical upper bounds on the complexity of Schrödinger potential approximation and estimation via neural networks. These bounds are determined by the effective dimension of the target distribution. To our knowledge, this is the first result demonstrating that generative modelling methods based on Schrödinger bridges and stochastic optimal control can escape the curse of dimensionality.

The Lipschitz constant of a neural network is a basic stability quantity appearing in optimization guarantees, generalization bounds, and robustness certificates. Deep networks naturally admit two such notions, \emph{input Lipschitzness}, measuring sensitivity to data perturbations, and \emph{parameter Lipschitzness}, measuring sensitivity to weight perturbations. We prove that, for deep networks with normalization layers, with batch normalization as the canonical example, \emph{both} Lipschitz constants can grow exponentially with depth and hence with parameter dimension, even under strong per-layer norm control such as $\\|W_\ell\\|_2\le 1$, and already for linear activations. Thus, a mechanism widely used to stabilize training can create severe worst-case instabilities that are invisible from layerwise norm bounds alone. Parameter-Lipschitz explosion turns Lipschitz-dependent nonsmooth optimization guarantees into exponential-in-dimension bounds for deep normalized networks, while input-Lipschitz explosion makes worst-case Lipschitz-based generalization and robustness certificates vacuous. The latter also yields a concrete rank-separation mechanism for adversarial vulnerability. The theory predicts that perturbations introducing new singular directions should be amplified much more strongly than equal-energy perturbations that remain within the input's original singular subspace. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 support this prediction, showing that rank-creating perturbations cause substantially sharper drops in accuracy, confidence, and margins than same-subspace perturbations.

Omni-modal Large Language Models (OLLMs) greatly expand LLMs' multimodal capabilities but also introduce cross-modal safety risks. However, a systematic understanding of vulnerabilities in omni-modal interactions remains lacking. To bridge this gap, we establish a modality-semantics decoupling principle and construct the AdvBench-Omni dataset, which reveals a significant vulnerability in OLLMs. Mechanistic analysis uncovers a Mid-layer Dissolution phenomenon driven by refusal vector magnitude shrinkage, alongside the existence of a modal-invariant pure refusal direction. Inspired by these insights, we extract a golden refusal vector using Singular Value Decomposition and propose OmniSteer, which utilizes lightweight adapters to modulate intervention intensity adaptively. Extensive experiments show that our method not only increases the Refusal Success Rate against harmful inputs from 69.9% to 91.2%, but also effectively preserves the general capabilities across all modalities.

Electromagnetic signal modeling is increasingly moving toward self-supervised pretraining on heterogeneous I/Q observations. However, an observed waveform is jointly shaped by signal generation, propagation environments, and acquisition processes, which mix reusable structural regularities with observation-specific variations. Existing methods usually pursue unified pretraining with reconstruction objectives on mixed observations. This heterogeneity creates two key mismatches for reconstruction pretraining: (1) appearance mismatch, induced by propagation-dependent waveform variations; and (2) temporal mismatch, induced by acquisition-dependent sampling coordinates. These mismatches make observation-specific appearances and sampling coordinates easier to reconstruct than stable signal structures, inducing observational shortcuts. We identify this failure mode as observation-biased pretraining. To address this, we propose Structural Pretraining with Time Alignment for heterogeneous EM signals. It mitigates temporal mismatch with RoPE-fs, which maps token positions to sampling-rate-normalized physical-time positions, and reduces reliance on appearance shortcuts by replacing waveform reconstruction with masked prediction of discrete structural units. These units are organized into atomic structural units and emergent units that capture variable-length compositional patterns beyond fixed patch boundaries. Across diverse EM benchmarks, our method improves both full-supervision and few-shot adaptation over existing EM pretraining baselines, showing that the learned representations better preserve structural consistency when observations drift in waveform appearance and temporal scale.

Deep multi-view clustering often struggles with inter-view feature redundancy and dimensionality collapse during iterative representation fusion. To address these limitations, we propose Orthogonal Residual Consensus Alignment (ORCA). First, a Semantic Incremental Recurrent Fusion (SIRF) module explicitly separates shared redundancy from complementary residuals via geometric projection. By employing an adaptive halting mechanism, SIRF autonomously maximizes information gain while avoiding structural collapse. Second, a Consensus-Calibrated Hierarchical Contrastive Alignment (CCHA) module is introduced to preserve multi-scale semantics. It adaptively aligns view-level, local, and recurrent representations using the global consensus as a semantic anchor. Consequently, ORCA learns latent representations with robust structural integrity and highly discriminative semantics. Extensive experiments on multiple benchmarks demonstrate that ORCA consistently outperforms state-of-the-art methods, with comprehensive ablations verifying its robustness.

Retrieval benchmarks are increasingly saturating, but we argue that efficient search is far from a solved problem. We identify a class of queries we call *oblique*, which seek documents that *instantiate a latent pattern*, like finding all tweets that express an implicit stance, chat logs that demonstrate a particular failure mode, or transcripts that match an abstract scenario. We study three mechanisms through which obliqueness may arise and introduce OBLIQ-Bench, a suite of five oblique search problems over real long-tail corpora. OBLIQ-Bench exposes an overlooked asymmetry between retrieval and verification, where reasoning LLMs reliably recognize latent relevance whenever relevant documents are surfaced, but even sophisticated retrieval pipelines fail to surface most relevant documents in the first place. We hope that OBLIQ-Bench will drive research into retrieval architectures that efficiently capture latent patterns and implicit signals in large corpora.

Adaptive decision-making in biological and artificial intelligence requires balancing the exploitation of known outcomes with the exploration of uncertain alternatives. Although prior work suggests that uncertainty generally promotes exploration, it has typically treated distinct sources of environmental uncertainty as equivalent. We consider environments with latent reward states that drift over time (volatility) and are observed through noisy outcomes (stochasticity). Both increase posterior uncertainty, yet we show they drive optimal exploration in opposite directions: volatility enhances it, stochasticity suppresses it. We establish this asymmetry formally by extending the Gittins index framework to Gaussian state-space bandits with latent dynamics. We further derive Cause-Aware Uncertainty-Sensitive Exploration (CAUSE), a closed-form exploration bonus obtained via control-as-inference that inherits the same monotonicities. CAUSE outperforms standard exploration strategies in environments with heterogeneous noise structure, and also improves on a Gittins-per-arm policy whose rested-bandit optimality does not transfer to restless settings. Learning and exploration are governed by the same noise-inference asymmetry, and the framework predicts that pathological noise inference produces reversed rather than merely impaired exploration, with implications for computational accounts of psychiatric conditions.

A central question in computational vision is whether human-like visual representations are better explained by discriminative or generative learning. Existing comparisons, however, often confound the learning objective with architecture, scale, and training data, leaving open whether the objective itself drives alignment. We address this confound using Joint Energy-Based Models (JEMs), which interpolate continuously between discriminative and generative training within a fixed architecture. By varying a single mixing coefficient, we isolate the effect of the learning objective and evaluate the resulting models across six human-alignment benchmarks spanning perceptual similarity, gloss perception, human response uncertainty, robustness, shape–texture cue conflict, and diagnostic feature attribution. Across this diverse suite, human alignment is consistently maximized at intermediate points of the generative–discriminative continuum, rather than at either endpoint. Hybrid JEMs combine the categorical structure induced by discriminative learning with the sensitivity to input structure induced by generative learning, yielding more human-like behavior across multiple levels of vision. These results suggest that the generative–discriminative dichotomy is the wrong axis for understanding human-aligned vision: alignment emerges not from choosing one objective over the other, but from balancing both.

Backpropagation (BP) remains the predominant algorithm for training deep neural networks, but it is biologically implausible and suffers from the backward locking problem, constraining parallelization across layers. Direct Feedback Alignment (DFA), which propagates the global output error directly to each hidden layer via fixed feedback matrices, has emerged as a promising alternative for addressing these limitations. However, a significant performance gap persists between DFA and BP, which we attribute to the limited ability of fixed feedback matrices to capture the nonlinear transformation from hidden representations to the network output. To address this limitation, we propose Nonlinear Direct Feedback Alignment (NDFA), which constructs each hidden-layer feedback using a lightweight nonlinear surrogate subnetwork. To enhance the effectiveness of the feedback throughout training, we further propose a learning strategy to update the surrogate subnetwork, thereby better approximating the underlying nonlinear transformation. Extensive experiments across residual networks and vision transformers on CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets demonstrate that NDFA consistently outperforms existing DFA-based methods and substantially narrows the performance gap to BP, while maintaining efficient parallel training. Code will be publicly available after the review process.

Latent signals are often obscured by measurement noise, yet encode the underlying laws and dynamics of complex systems; learning both the signals and their distributions remains a central challenge in scientific inference. The noise is often non-negligible, and the likelihoods for expressive generative models are often intractable. We utilize a convolutional maximum mean discrepancy (convMMD) loss and propose a likelihood-free framework for nonparametric density deconvolution and empirical Bayes denoising under additive measurement error. Our method learns a latent generative model by matching the observed data distribution to the noise-convolved model distribution. This yields a differentiable, simulation-based objective for multivariate homoscedastic or heteroscedastic noise, compatible with expressive sieve classes such as Gaussian mixtures and normalizing flows. The learned density then serves as an empirical prior for posterior denoising of individual latent values. Theoretically, we extend convMMD from parametric to nonparametric estimation, proving finite-sample bounds for empirical sieve minimizers and $L_2$ convergence rates under Sobolev smoothness. These rates recover the classical inverse-problem dependence: polynomial for ordinary-smooth and logarithmic for super-smooth noises. Our method provides a practical, theoretically grounded approach to deconvolution and denoising under generative latent distribution models.

The expressivity of graph neural networks is predominantly characterized in a noiseless setting via binary graph isomorphism tests. This perspective certifies whether two graphs can be distinguished, but does not quantify how robust such distinctions are under structural perturbations. In practice, however, real-world graphs are often corrupted by noise, where even small structural changes might lead to disproportionate distortions in learned representations. In this study, we generalize graph isomorphism to a noisy alignment criterion, requiring that graphs that remain identifiable under noise are embedded closer to their source than to any alternative. This yields a margin-based formulation of expressivity that bridges the gap between combinatorial isomorphism-based theory and the continuous learned representations. Notably, our framework is architecture-agnostic and strictly generalizes classical isomorphism-based expressivity. We further derive certified robustness guarantees for node-level tasks and an explicit robustness radius for graph-level representations. Experiments confirm that our criteria predict and improve robustness under structural perturbations.

Activation monitoring, which probes a model's internal states using lightweight classifiers, is an emerging tool for AI safety. However, its worst-case robustness under a misalignment threat model, where a model might learn to actively conceal its internal states, remains untested. We ask: could a model learn to evade previously unseen activation monitors? We demonstrate that fine-tuning can create **Neural Chameleons**: models capable of zero-shot evading activation monitors. Specifically, we fine-tune an LLM to evade monitors for benign concepts (e.g., HTML) when conditioned on a trigger phrase "You are being probed for {concept}". This learned mechanism generalizes zero-shot: substituting the concept with a safety-relevant term like 'deception' causes the model to successfully evade previously unseen safety monitors, even those trained post hoc on no-trigger activations from the frozen fine-tuned checkpoint. We validate this across diverse model families (Llama, Gemma, Qwen), finding that evasion is highly selective to the triggered concept and incurs minimal capability degradation. Mechanistic analyses suggest that the trigger induces an activation shift that moves representations away from probe-aligned directions. Our work provides a proof-of-concept for this failure mode of activation monitoring under misalignment threat models.

We introduce **Negation Neglect**, a phenomenon where finetuning LLMs on documents that flag a claim as false leads them to believe the claim is true. For example, models are finetuned on documents that convey that "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence—e.g., "Ed Sheeran did *not* win the 100m gold"—models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

Last-iterate convergence of learning dynamics in games has attracted significant recent attention. In two-player zero-sum games with bandit feedback, where only the loss of the selected action pair is observed, Fiegel et al. [2025] show a separation between average-iterate and last-iterate convergence in duality gap: while the optimal $t^{-\frac{1}{2}}$ rate after $t$ rounds is achievable for the former via standard no-regret algorithms, the latter cannot converge faster than $t^{-\frac{1}{3}}$ in expectation or $t^{-\frac{1}{4}}$ with high probability. However, in many practical settings (such as preference learning), the players observe not only their loss but also the opponent’s action. This raises a natural question: can such additional information enable faster last-iterate convergence? We answer this question affirmatively, showing that $t^{-\frac{1}{2}}$ last-iterate convergence is achievable with high probability in this setting, via an efficient algorithm that updates its strategy infrequently by solving an estimated log-barrier-regularized game. We identify fundamental obstacles preventing standard analysis for multi-armed bandits (the single-player case) from generalizing to games, and develop a novel analysis to overcome them. Experiments confirm that our algorithm indeed converges faster than naive baselines and prior methods that do not exploit opponent-action feedback. Finally, we note that our results also improve those for dueling bandits, a special case with skew-symmetric game matrices.

Large language models (LLMs) exhibit severe multilingual safety misalignment: they possess strong safeguards in high-resource languages but remain highly vulnerable to jailbreak attacks in low-resource languages. Current safety alignment methods generally rely on high-quality response data for each target language, which is expensive and difficult to generate. In this paper, we propose a cross-lingual safeguard transfer framework named Multilingual Self-Distillation (**MSD**). This framework transfers an LLM's inherent safety capabilities from high-resource (*e.g.*, English) to low-resource (*e.g.*, Javanese) languages, overcoming the need for response data in any language. Our framework is flexible and can be integrated with different self-distillation strategies. Specifically, we implement two concrete methods---on-policy MSD and off-policy MSD---both of which enable effective cross-lingual safety transfer using only multilingual queries. Furthermore, we propose Dual-Perspective Safety Weighting (DPSW), a divergence measure to optimize the distillation objective. By jointly considering the perspectives of both the teacher and the student, DPSW adaptively increases the penalty weights on safety-critical tokens while reducing the weights on non-critical tokens. Extensive experiments on representative LLMs across diverse multilingual jailbreak and utility benchmarks demonstrate that our method consistently achieves superior multilingual safety performance. Notably, it generalizes effectively to more challenging datasets and unseen languages while preserving the model's general capabilities.

Realizing stable neural decoding over extended periods remains a significant challenge, primarily because invariant task-related neural dynamics are inextricably entangled with inherent signal non-stationarity and the physical drift of electrode-neuron interfaces. This work introduces LNDN (Long-term Neural Decoding Network), a hierarchical framework designed to decouple category-specific neural dynamic manifolds from representational drift, sustaining decoding precision without periodic recalibration. The proposed architecture addresses representational drift through an integrated tri-level alignment strategy. At the physical feature level, a decoupled representation and dynamic gating module isolates the physical identity of neurons from their transient states, adaptively filtering stochastic neural noise. This is further strengthened at the decision level by a Mixture-of-Experts (MoE) architecture, which employs a parallel soft-voting mechanism to suppress high variance induced by local representational distortions. Finally, at the manifold level, Adaptive Batch Normalization (AdaBN) and latent contrastive learning, combined with second-order correlation alignment, explicitly anchor the geometric consistency of the low-dimensional neural manifold over a period of months. Systematic evaluations on a longitudinal olfactory dataset, involving 9 mice performing a four-odor decoding task, demonstrate the efficacy of this approach. When trained exclusively on data from the initial four weeks, LNDN maintains an average accuracy of 83.67\% on completely unseen tests spanning the subsequent eight weeks, significantly outperforming mainstream domain adaptation and advanced neural decoding baselines. This framework offers a scalable, "set-and-forget" solution for robust, long-term invasive brain-computer interface applications.