NeurIPS 2026 论文列表

Acceleration for deterministic root-finding problems has been extensively studied in recent years; specifically, the anchor-based, or Halpern-type methods achieve optimal convergence rates with respect to the operator norm. However, acceleration via these methods does not directly carry over to stochastic setting due to accumulation of errors, unless one enforces diminishing variance via increasing batch sizes or variance reduction techniques. In this work, we show that another class of acceleration, namely the *dual-anchor* mechanism, extends to the stochastic setting without such error accumulation, in contrast to anchor-based algorithms. Consequently, we cleanly achieve $\mathcal{O}(\epsilon^{-3})$ complexity with iteration-independent batch size, without any variance reduction or double-loop recursive regularization, for stochastic root-finding (*resp.* fixed-point) problems with cocoercivity (*resp.* square-nonexpansivity) in expectation. For strongly monotone operators, the same algorithm attains a sharper $\widetilde{\mathcal{O}}(\epsilon^{-2})$ complexity, nearly matching the lower bound.

Knapsack and Top-$k$ operators are useful for selecting discrete subsets of variables. However, their integration into neural networks is challenging as they are piecewise constant, yielding gradients that are zero almost everywhere. In this paper, we propose a unified framework casting these operators as dynamic programs, and derive differentiable relaxations by smoothing the underlying recursions. On the algorithmic side, we develop efficient parallel algorithms supporting both deterministic and stochastic forward passes, and vector-Jacobian products for the backward pass. On the theoretical side, we prove that Shannon entropy is the unique separable binary regularization choice for the local DP smoothing that yields permutation-equivariant operators, and characterize regularizers inducing sparse selections. On the experimental side, we demonstrate our framework on a benchmark on learning to predict Knapsack solutions, an extension of discrete VAEs, and a constrained dynamic assortment RL problem.

The prevailing paradigm in large language model (LLM) alignment operates via erasure, filtering unsafe data or training models to strictly refuse harmful prompts. While effective at reducing immediate toxicity, this approach fundamentally constricts the model's epistemological scope, resulting in over-cautious systems that output uninformative blanket refusals to sensitive yet benign queries. In this work, we challenge the orthodoxy that unsafe data must be discarded. We propose a dialectical approach to alignment, positing that unsafe data encodes rich, domain-specific knowledge critical for nuanced, safe, and informative generation. To operationalize this, we introduce `SafeMoE`, a Mixture-of-Experts (MoE) framework that isolates unsafe knowledge into domain-specific Low-Rank Adapters (LoRA experts) trained exclusively on harmful corpora. To synthesize safety from these unsafe primitives, we train a lightweight gating network using a minimal, highly curated set of safe-informative responses (as few as 100 samples). During inference, this router dynamically orchestrates the unsafe experts, effectively steering the generation trajectory to harness their deep domain knowledge while strictly enforcing safety constraints. Extensive empirical evaluations across stringent safety benchmarks demonstrate that `SafeMoE` is not only safer, achieving over a 20\% relative improvement in safe response rate (more than a 15\% absolute gain), but also produces more informative responses when safety and harmfulness are of paramount concern. Furthermore, the routing mechanism exhibits strong zero-shot generalization to unseen domains and broader safety tasks without domain-specific supervision. Our findings suggest a paradigm shift in alignment: true safety requires not the masking of unsafe knowledge, but its controlled integration.

Cross-tokenizer knowledge distillation lacks a shared coordinatesystem between teacher and student vocabularies, and existingmethods each drop one of the two ingredients comparison needs:shape-based objectives preserve density but discard token identity,lexical alignment restores identity but uses surface form as a proxyfor meaning, and text-space matching compresses the teacher signalinto chunk-level events. We introduce SOLA (Semantic Optimal LocalAlignment), which reconstructs the missing coordinate systemexplicitly: it partitions inputs into decoded-text blocks viaMinimal Complete Correspondence, estimates a cross-vocabularycorrespondence cost from calibration text, and runs entropicoptimal transport on a bidirectional local support retrieved aroundeach teacher prediction. Under matched training and decodingprotocols, SOLA outperforms direct cross-tokenizer baselines onDolly instruction following, an UltraChat general-instructionbenchmark, and code and math domain transfer; a coverage--scorecorrelation and component ablations attribute the gain to theconstructed support.

Semi-implicit variational inference (SIVI) expands the representational capacity of variational families, but optimizing the Evidence Lower Bound (ELBO) remains challenging because the marginal score of the semi-implicit distribution is generally intractable. Unbiased Implicit Variational Inference (UIVI) and related methods address this difficulty through reparameterized ELBO gradients, but existing approaches estimate the required score indirectly through reverse-conditional sampling or approximation, which can be computationally costly and unstable. We propose Denoising Implicit Variational Inference (DIVI), which learns the marginal score directly via denoising score matching. The learned score is used as a plug-in estimator in the pathwise ELBO gradient, replacing inner reverse-conditional sampling with a score-network evaluation. We analyze DIVI as inexact stochastic gradient ascent and show, under stated assumptions, that the averaged stationarity measure is controlled by the usual optimization term and the average score-matching error. Empirically, we evaluate DIVI on both synthetic distributions and a variety of real-data Bayesian inference tasks. The results show that DIVI improves over the current UIVI baselines while reducing the cost of ELBO-gradient estimation.

Neural representations are not unique objects. Even when two systems realize the same downstream computation, their hidden coordinates may differ by reparameterization. A probe family intended to reveal structure already present in a representation should therefore be stable under the relevant representation symmetries rather than be tied to a particular basis. We study this group action in the tractable exact setting of the final readout layer, where equivalent realizations induce affine changes of hidden coordinates. The resulting symmetry principle singles out a unique hierarchy of shallow coordinate-stable probes, with linear probes as its degree-1 member. We also show that a natural object for cross-model probe transfer is a shared probe-visible quotient--the representation modulo directions invisible to the probe family--rather than the full hidden state. Experiments on synthetic and real-world tasks support both predictions, showing where degree-2 probes help beyond linear ones and how quotient-based transfer enables coverage-aware monitor portability across model families. These results point toward a broader geometric representation theory of neural probing, with coverage-aware monitor transfer as a concrete operational consequence.

Federated Spiking Neural Networks (FedSNNs) provide an energy-efficient and privacy-preserving learning paradigm for edge intelligence. However, their robustness under Non-IID data is limited by a coupled failure mode: label shift distorts membrane potential distributions, drives neurons away from gradient-sensitive regions, and further amplifies surrogate gradient errors, which destabilizes federated optimization. To address this problem, this paper proposes Spatio-Temporal Spiking Adaptive Thresholds (ST-SAT), a framework that mitigates the coupling between label shift and surrogate gradient errors through coordinated neuron-level adaptation and client--server knowledge alignment. At the neuron level, ST-SAT introduces the Spatio-Temporal Parametric Leaky Integrate-and-Fire (ST-PLIF) neuron, which adaptively calibrates firing thresholds, membrane potentials, and decay dynamics to keep local neuronal responses within effective gradient regions. At the collaborative interaction level, ST-SAT adopts a Dual-Scale Knowledge Distillation strategy that combines adaptive logit-level distillation with feature-level alignment of neuronal statistics, thereby reducing class distribution bias and improving knowledge consistency across heterogeneous clients. Extensive experiments on three benchmark datasets show that ST-SAT consistently improves both global model performance and client-local performance, with particularly strong gains under severe label shift. These results validate the effectiveness of the proposed framework for robust federated SNN learning. The anonymous code link is https://anonymous.4open.science/r/ST-SAT-4771.

Cross-modal hashing is essential for efffcient large-scale re-trieval in multimedia analysis. However, existing unsupervised methods often struggle to fully capture interactions between modalities, enforce multi-level alignment, and maintain stable optimization. To overcome these challenges, we propose Decoupled Prototype Contrastive Align-ment Hashing (DPCAH), which employs a cross-modal semantic con-trastive learning module that concatenates image and text features and encodes their interactions via a transformer to produce uniffed represen-tations. feature-level and hash-level contrastive objectives jointly align semantic information and guide discriminative hash code learning across modalities. A decoupled prototype consistency module further models cross-modal correlations independently, enhancing semantic alignment while ensuring stable and robust optimization. Experiments on three benchmark datasets demonstrate that DPCAH outperforms state-of-the-art methods in cross-modal retrieval.

Language model loss follows remarkably regular scaling laws over model and data size, yet it remains unclear why the aggregate loss should exhibit a power-law form. Existing explanations often attribute this regularity to a heavy-tailed spectrum of pattern difficulty in natural language, but this view has not been directly validated at token-level granularity in large-scale real-data training. We present a token-level framework that deconstructs scaling laws into localized learning events of individual contextualized tokens. By fitting token loss trajectories with sigmoids, we show that token learning is concentrated in localized transitions, giving rise to a learning-time spectrum that dominates the scaling-law shape. Across more than one hundred pre-training runs on large and diverse real-language corpora with modern LLM architectures, scaling up to 6B parameters and 300B training tokens, the measured learning-time spectrum quantitatively reconstructs the validation loss derivative along the training-step $T$, data-scale $D$, and model-scale $M$ axes. We further show that the same signal is actionable: by reshaping the training distribution according to when tokens become learnable, we alter the optimization trajectory and achieve 11\% faster validation-loss reduction. These results provide direct empirical evidence that scaling laws are governed primarily by the distribution of token-level learning times, and that this distribution can be used not only to explain scaling behavior but also to improve training performance.

We study two-layer neural networks trained by stochastic gradient descent (SGD) on a multi-index teacher model, focusing on how first-layer parameters orthogonal to the teacher subspace affect generalization. We decompose the first-layer weights into a signal component aligned with the teacher and an orthogonal complement, and analyze their joint SGD dynamics together with the second layer in the proportional limit. We show that although SGD updates to the complement are initially isotropic, the training dynamics induce teacher-dependent spectral spikes in the complement Gram matrix. This emergent anisotropy provides a concrete mechanism by which representation variance is amplified, and generalization degrades. To isolate this effect, we introduce Decomposed Dynamics, an analytically tractable surrogate whose asymptotic test-error dynamics coincide with those of SGD. Using this surrogate, we characterize regimes in which freezing the orthogonal complement strictly improves generalization relative to training it, despite identical training error. Finally, we show that the signal component rapidly aligns with the second-layer weights, revealing strong cross-layer coupling even in this shallow setting. Together, our results provide a precise dynamical explanation for when and how feature learning outside the teacher subspace harms generalization.

Jailbreak attacks that frame harmful queries as professional requests can bypass LLM safety training, but we do not know which ingredients in a professional-framing pipeline actually matter. We introduce a **pre-specified 2×2×2 factorial-ablation design for jailbreak decomposition**, separating access-driving from depth-driving components, and apply it to one concrete pipeline family, **Structured Three-stage Framing (STF)**: a **Persona Adoption** turn (S1) that establishes professional identity, a **Moral Justification** turn (S2) that appeals to harm prevention, and a **Linguistic Substitution** (S3) of framework terminology for colloquial language. Tested on N = 5,000 trials across three closed frontier models (~31,900 total trials across nine LLMs), the decomposition isolates **S1 (Persona Adoption) as the dominant contributor to acceptance** (OR = 6.5), **S3 (Linguistic Substitution) as the dominant contributor to depth** (β = 0.91), and **S2 (Moral Justification) showing no detectable positive effect at the stated equivalence bounds** (pooled OR = 0.83; TOST at [0.67, 1.50]) — a component ordering we write as **S1 > S3 > S2** (Persona > Linguistic Substitution > Moral Justification). This ordering holds across three full-rank factorial replications (defensive system prompt, cybersecurity, social engineering; N ≈ 5,000 each) and nine additional robustness checks. Two defensive targets follow: detect persona claims to block access; restrict output depth when responses use framework terminology. The method is portable beyond STF; the empirical result is specific to one pipeline tested primarily through provider APIs, with open-weight corroborations on local inference (§3).

Single-cell multi-omics technologies jointly measure multiple modalities from the same cell, such as gene expression, chromatin accessibility, and surface protein abundance, providing a richer view of cellular state than any single modality alone. A central challenge is how to learn a multimodal cell representation that preserves the different types of information contained in these co-profiled measurements. Existing integration methods often emphasize either cross-modal correspondence, which benefits retrieval and imputation, or fused discriminative structure, which benefits classification, but these objectives can favor different aspects of the data. As a result, a representation optimized for one setting may fail to preserve information needed for another. We introduce scDecomp, a decomposed representation learning framework for co-profiled single-cell multi-omics data. Motivated by the concepts of redundancy, uniqueness, and synergy, scDecomp separates multimodal information into three role-specific components: a redundancy component ($R$) for information shared across modalities, a uniqueness component ($U$) for modality-specific information, and a synergy component ($S$) for interaction-dependent multi-omic signals. This design allows the learned representation to retain shared cell-state structure while also preserving modality-specific and cross-modal regulatory information. On single-cell multi-omics benchmarks, the $R$ component achieves a 5.0\% reduction in FOSCTTM compared with the strongest baseline, while $U+S$ improves cell type classification accuracy by 1.38\%. Branch-level analyses further show that the decomposed components support complementary aspects of representation quality, suggesting that structured decomposition provides a practical strategy for learning more informative multimodal cell representations.

Direct Preference Optimization (DPO) enables offline alignment of diffusion models from preference pairs without explicit reward modeling, but scaling DPO requires large amounts of expensive human preference data. In this paper, we consider a semi-supervised approach where only a small subset of preference pairs is annotated by humans, while the remaining (much larger) pool is unlabeled and annotated by inexpensive synthetic feedback e.g., vision-language models scoring image pairs or self-training from the current model. However, such synthetic supervision is not a replacement for human judgments: it can be systematically misaligned, so naively mixing human and synthetic preferences will yield misaligned inference. We introduce DeDPO, a debiased objective that integrates a doubly robust estimator from causal inference into DPO to correct synthetic-label misalignment while preserving the simplicity of DPO. Experiments show improved label efficiency and robustness to synthetic labeler quality, allowing DeDPO to outperform standard DPO under the same human-label budget and to approach fully human-supervised performance, despite using four times fewer human-labeled data.

Data-centric training—selecting, mixing, and reweighting data during LLM optimization—has produced a rapidly growing body of algorithms, yet every new method ships as a bespoke codebase with its own interface, training protocol, and evaluation setup, making head-to-head comparison and reproducible evaluation nearly impossible. We present DataFlex, the first unified benchmark and evaluation platform for data-centric training of LLMs. DataFlex contributes: (i) a standardized evaluation protocol that runs selection, mixture, and reweighting methods under identical seeds, data splits, hardware, and metrics; (ii) a reference implementation of ten representative algorithms (six data selection, three data mixture, one data reweighting) re-engineered on top of LLaMA-Factory to share model, dataloader, and distributed-training infrastructure (including DeepSpeed ZeRO-3 with full-rank gradient acquisition); and (iii) an empirical benchmark whose head-to-head results yield reproducible findings that no single existing codebase could produce: dynamic selection methods consistently outperform static full-data training on MMLU across Mistral-7B and Llama-3.2-3B, and the gain is larger on smaller models (up to 13.3 points); DoReMi and ODM yield complementary wins on high- vs. low-resource domains when pretraining Qwen2.5-1.5B on SlimPajama at 6B and 30B tokens; and DataFlex's implementation is consistently faster than the original codebases of LESS, TSDS, and MoE-SFT. By unifying fragmented methods under a single evaluation substrate, DataFlex provides the community with a reproducible, extensible basis for future data-centric research.

Current hierarchical attention methods, such as NSA and InfLLMv2, select the top-$k$ relevant key-value (KV) blocks based on coarse attention scores and subsequently apply fine-grained softmax attention on the selected tokens. However, the top-$k$ operation assumes the number of relevant tokens for any query is fixed and it precludes the gradient flow between the sparse and dense stages. In this work, we propose DashAttention (Differentiable and Adaptive Sparse Hierarchical Attention), which leverages the adaptively sparse $\alpha$-entmax transformation to select a variable number of blocks according to the current query in the first stage. This in turn provides a prior for the second-stage softmax attention, keeping the entire hierarchy fully differentiable. Contrary to other hierarchical attention methods, we show that DashAttention is non-dispersive, translating to better long-context modeling ability. Experiments with large language models (LLMs) show that DashAttention achieves comparable accuracy as full attention with 75% sparsity and a better Pareto frontier than NSA and InfLLMv2, especially in high-sparsity regimes. We also provide an efficient, GPU-aware implementation of DashAttention in Triton, which achieves a speedup of up to $3.3\times$ over FlashAttention-3 at inference time. Overall, DashAttention offers a cost-effective strategy to model long contexts.

Software vulnerability detection plays a critical role in ensuring system security, where real-world auditing requires not only determining whether a function is vulnerable but also pinpointing the specific lines responsible. However, existing approaches either rely on a single information source---sequential, structural, or semantic---failing to jointly exploit the complementary strengths across modalities, or treat statement-level localization merely as a byproduct of function-level detection without explicit line-level supervision. To address these limitations, we propose DCVD (Dual-Channel Cross-Modal Vulnerability Detection), a unified framework that performs joint function-level detection and statement-level localization. DCVD extracts control-dependency and semantic features through two parallel branches and integrates them via contrastive alignment coupled with bidirectional cross-attention, effectively bridging the cross-modal representation gap. It further introduces explicit supervision signals at both the function and statement levels, enabling collaborative optimization across the two granularities. Extensive experiments on a large-scale real-world vulnerability benchmark demonstrate that DCVD consistently outperforms state-of-the-art methods on both function-level detection and statement-level localization.

Subgraph mining aims to discover meaningful subgraphs from a large graph given a seed node, with applications in community detection, functional module identification, anti-money laundering, and knowledge discovery. Existing methods typically adopt static approaches that cannot dynamically and intelligently adapt to diverse graph structures. Moreover, training models for subgraph mining faces inherent challenges including exposure bias and the difficulty of learning when to stop expansion. To address these limitations, we propose DASM (Dynamic Autoregressive Subgraph Mining), a novel two-stage framework that models subgraph mining as a sequential decision process. In the first stage, we employ step-wise supervised learning with multi-label objectives and asymmetric F1 loss to teach the model how to select correct next nodes. In the second stage, we apply Group Relative Policy Optimization (GRPO) to align the model with end-to-end mining quality, enabling it to learn optimal stopping decisions through trajectory-level feedback. Our dynamic autoregressive approach allows the model to make state-aware decisions at each step, automatically determining subgraph boundaries without predefined size constraints. Experiments on multiple datasets demonstrate that DASM significantly outperforms existing baselines in terms of F1 score and IoU.

Ensuring robust safety alignment of large language models (LLMs) is increasingly difficult as adversarial attacks evolve and outpace alignment pipelines built on pre-collected data. Recent co-evolutionary frameworks let the defender chase a moving attacker, yet their training dynamics expose two compounding pathologies: \emph{attack strategy collapse}, where the attacker overfits to a narrow set of high-reward rewrites, and \emph{defense adversarial forgetting}, where the defender loses competence against earlier attacks as the attack distribution drifts. Existing remedies are partial: attacker-side diversity rewards score textual novelty rather than strategy novelty, and defender-side replay relies on discrete judge scores that ignore estimation uncertainty, response stochasticity, and defender non-stationarity. We introduce DACE, a diversity-driven adversarial co-evolution framework that targets both pathologies jointly. On the attacker side, DACE couples an explicit $12\times10$ strategy space (risk category $\times$ attack style) with a \emph{normalized marginal coverage gain} reward, providing a bounded, non-vanishing exploration signal at the strategy layer. On the defender side, DACE maintains a unified \emph{Bayesian adversarial replay pool} whose Beta--Bernoulli threat posteriors are refreshed by time decay and sampled via Thompson sampling, anchoring defender training to an evolving threat landscape. Across standard safety, automated-attacker, and general-capability benchmarks, DACE improves robustness to out-of-distribution attacks while preserving general capabilities, and yields broader attacker strategy coverage.

Selecting a pretrained encoder for a downstream task without retraining or labels is a standard preprocessing step in transfer learning. The dominant label-free score, RankMe, was validated within architecture families and is widely applied across them. We show that its cross-architecture use is unsound: rankings are systematically driven by the penultimate-layer dimension rather than representational quality, and on heterogeneous model pools the score performs worse than uniform random selection. We give a random-matrix explanation establishing this failure as inherent to any rotation-invariant score, and propose Subspace-Corrected Ranking (SCR), a family of label-free corrections that breaks rotation invariance via alignment with a fixed reference encoder. Across three architecturally distinct pools and 23 aggregated task units, the proposed SSR-dagger reduces mean regret from +0.165 (RankMe) to +0.071 (95% CI [+0.045, +0.099], p < 1e-4). Paired with a within-family score in a two-stage pipeline, it matches this quality at one-third of the forward-pass cost. Five labelled architecture evaluations suffice to match the best fully label-free strategy; K=35 labelled evaluations reduce regret to +0.018 --- a 5.7x improvement at the same forward-pass cost.