NeurIPS 2026 论文列表

A common way for trajectory planning is to leverage generative models trained on large collections of expert trajectories. At inference time, the model generates executable trajectories by conditioning on task goal constraints. However, trajectory-based methods rely on costly supervision, scale poorly with sequence length, and often generalize poorly to unseen constraints such as novel start-goal pairs. We propose an alternative to learn the underlying state-space manifold and use the geometry of the manifold for trajectory planning. This approach requires only state observations and enables generalization to unseen constraints by constructing trajectories on the learned manifold of the state space. Experiments on classical planning problems in maze demonstrate the effectiveness of our work. We further show that the method can be used in higher dimension space where table-top robot arms are considered. Our method is able to find feasible path given only infeasible straight line reference, and be comparable to State-of-the-Art trajectory-based model without actually learning on trajectory data.

Although existing candidate trajectory evaluators have substantially improved end-to-end planning accuracy, their underlying mechanism remains suboptimal. Model-based evaluators typically make scoring future-aware by explicitly predicting candidate-conditioned future states over the planning horizon, which we term rollout. However, rollout incurs high inference cost and accumulates prediction errors. We argue that planning does not require reconstructing a full future scene for each candidate, but only future-relevant information for reliable comparison. We therefore propose RFWorld, a Rollout-Free World model for end-to-end trajectory evaluation. RFWorld constructs a shared scene memory and shapes it into a future-readable representation via lightweight training-time temporal readout with semantic supervision, keeping training overhead low. At inference time, candidate trajectories directly query the shared memory via sparse spatial readout for scoring, reducing computation by roughly the planning horizon times the number of candidate trajectories, while avoiding recursive rollout errors. Experiments on NAVSIM show that RFWorld learns a more planning-aligned world representation and achieves state-of-the-art performance.

Large language model agents have shown strong capabilities in generating coherent and contextually appropriate responses, yet robust long-horizon dialogue remains limited by the lack of external memory that is traceable, updatable, and diagnostically transparent. Existing memory-augmented agents often store memories as isolated records or overwritable states, making it difficult to preserve how information originates, evolves, conflicts, or becomes obsolete over time. We propose TrajWiki, a trajectory-based memory framework for long-horizon conversational agents. Instead of treating memory as static entries, TrajWiki represents each memory as a source-grounded evolution trajectory, maintained through immutable episodic snapshots and claim-level operations such as ADD, REVISE, and DEPRECATE. To reduce fragmentation and retrieval cost, TrajWiki further introduces Memory Wiki, a persistent intermediate layer that incrementally compiles dialogue history into structured and interlinked wiki pages capturing salient entities, events, quantities, topics, and conflicts. At inference time, queries are routed hierarchically from relevant wiki pages to linked memory trajectories, then to corresponding snapshots and source messages for evidence-grounded answer synthesis. Experiments on LoCoMo and MedMT show that TrajWiki improves long-horizon dialogue performance across both open-source and closed-source LLM backbones, while providing greater interpretability and diagnostic visibility into memory evolution, retrieval failures, and answer generation. Code is available at https://anonymous.4open.science/r/TrajWiki-F0EC/.

Controlling the motion of multiple objects in image-to-video (I2V) generation requires preserving object identities while enforcing adherence to distinct target trajectories. This becomes particularly challenging as the number of objects increases and their paths intersect or occlude one another. Existing approaches entangle multiple trajectories within a shared, dense conditioning signal, making object-level correspondence difficult to preserve in crowded scenes. We depart from this paradigm and enforce a strict, per object spatial constraint that isolates instances independently. Our method, TrajLoc, achieves this directly within the attention layers by substituting the cross-attention weights of each object token with a Gaussian heatmap centered on its target location at every frame. The same per object token interface carries trajectory and depth through a learned embedding and preserves identity by encoding first frame appearance in place of an object token. Evaluations across six datasets, featuring up to 20 simultaneously controlled objects and out of distribution real world scenes, demonstrate that our method consistently improves both visual fidelity and trajectory adherence. Applied to two architecturally distinct backbones (CogVideoX 5B and WaN 2.1 14B), our approach achieves average gains of +4.3 dB PSNR and a 51\% reduction in trajectory end point error compared to the strongest baselines. Code will be made publicly available upon publication.

In this work, we introduce a novel training-free encoding framework, dubbed **XShapeEnc**, that encodes an arbitrary spatially grounded 2D geometric shape into a compact representation exhibiting five favorable properties, including invertibility, adaptivity, generality and controllability. Specifically, a 2D spatially grounded geometric shape is decomposed into its normalized geometry within the unit disk and its pose vector, where the pose is further transformed into a harmonic pose field that also lies within the unit disk. A set of orthogonal Zernike basis is constructed to encode shape geometry and pose either independently or jointly, with controllable relative emphasis on shape geometry or shape pose. We demonstrate the theoretical validity, efficiency, discriminability, and wide applicability of **XShapeEnc** via extensive analysis and experiments across a wide range of shape-aware tasks and our self-curated XShapeCorpus dataset. We envision **XShapeEnc** as a foundational tool for research beyond one-dimensional sequential data and data-driven, learning-based encoding paradigms, paving the way for a unified spatial encoding framework for frontier 2D spatial intelligence.

Transformers perform inference by iteratively transforming token representations across layers. This layerwise computation has been studied empirically, and recent mean-field theories of Transformer dynamics explain how attention can drive token distributions toward clustering. However, existing mean-field analyses largely treat model parameters as prescribed, leaving open how training reshapes this clustering picture. We study this question in a noisy mean-field Transformer in which only a parameter-linear FFN is trained under $L^2$ regularization. We find and analyze a training-induced phase in the dynamics: after initially following attention-driven clustering, the token distribution can leave the clustered regime near the final layers. Our mathematical analysis is based on an entropy-regularized interaction energy that captures the clustering bias of attention. More broadly, our results point toward a training-aware mean-field theory of Transformer dynamics, in which training and inference dynamics are treated together.

Open World Object Detection (OWOD) requires a detector to recognize known categories, detect unknown objects, and incrementally learn new classes. Recent work such as OW-OVD adapts a pre-trained Vision-Language Object Detector (VLOD) to OWOD through fine-tuning. However, this approach demands costly backpropagation and often degrades the detector's own zero-shot performance on known categories, calling into question whether training is necessary. In this paper, we propose CAVE (Clustering in Attribute-Visual Embedding space), a training-free framework that adapts a VLOD to OWOD without any parameter updates. From a single forward pass, CAVE collects per-class visual and attribute statistics via von Mises-Fisher-based clustering, along with attribute co-occurrence patterns, without any gradient computation. At inference, CAVE fuses the original VLOD's text-based scores with MAS (Mixture-Aggregate Score) derived from the collected statistics for known-class detection, while unknown objects are identified by AOS (Attribute Objectness Score) that combines superclass responses with attribute co-occurrence patterns. Experiments on M-OWODB show that CAVE outperforms OW-OVD by +16.2 known-class mAP (Task 4) and +28.0 average unknown recall, while surpassing the zero-shot VLOD across all incremental tasks.

Diffusion models generate samples by denoising along the score of a perturbed target distribution. In practice, one trains a neural diffusion model, which is computationally expensive. Recent work suggests that score matching implicitly smooths the empirical score, and that this smoothing bias promotes generalization by capturing low-dimensional data geometry. We propose moment-matched score-smoothed overdamped Langevin dynamics (MM-SOLD), a training-free interacting particle sampler that enforces the target moments throughout the sampling trajectory. We prove that, in the large-particle limit, the empirical particle density converges to a deterministic limit whose one-particle stationary marginal is a Gibbs--Boltzmann density obtained by exponentially tilting a naive score-smoothed diffusion target. The mean and covariance of this distribution agree with the empirical moments of the training data. Experiments on 2D distributions and latent-space image generation show that MM-SOLD enables fast, robust, training-free sampling on CPUs, with sample fidelity and diversity competitive with neural diffusion baselines.

Choosing where to place entangling gates is a central design choice in quantum machine learning circuits, yet entanglers are still typically chosen from fixed templates or by expensive search procedures requiring training or pairwise kernel evaluations. We propose the \emph{HSD-Lipschitz principle}, a training-free geometric criterion built on a simple intuition: a useful entangler should keep encoded states stable across the dataset while separating different classes. We formalise this through a two-sided bound on local light-cone subsystems. Under a light-cone locality assumption, dataset-level gradient variance is upper-bounded by reduced-state spread; under random initialisation, the expected between-class gradient signal is lower-bounded by class-conditional reduced-state distance. The resulting algorithm, HSD-greedy, adaptively grows an entanglement structure using only single-state reduced moments, estimable via classical shadows, and avoids the $\mathcal{O}(N^2)$ pairwise kernel evaluations of kernel-target alignment proxies. Empirically, both bounds hold without violation across $110$ topologies and $440$ parameter rows, and extend to $180$ additional configurations. HSD-greedy attains the strongest average AUC among training-free entangler selectors on synthetic and real-world benchmarks, at roughly $60\times$ lower selection cost than the strongest baseline in 13-qubit simulation. End-to-end execution on a 20-qubit IBM Eagle subchain substantially outperforms a fixed 19-CNOT Linear chain and matches or exceeds a matched-budget Random baseline while using far fewer two-qubit gates than the Linear chain.

Large Language Models (LLMs) have achieved remarkable performance on a wide range of specialized tasks, exhibiting strong problem-solving capabilities. However, training these models is prohibitively expensive, and they often lack domain-specific expertise because they rely on general knowledge datasets. Expertise finetuning can address this issue; however, it often leads to overspecialization, and developing a single multi-domain expert remains difficult due to diverging objectives. Furthermore, multitask training is challenging due to interference and catastrophic forgetting. Existing work proposes combining the expertise of dense models within a Mixture of Experts (MoE) architecture, although this approach still requires multitask finetuning. To address these issues, we introduce Dynamic Upcycling MoE (DUME), a novel approach that reuses dense experts trained on different domains to construct a unified MoE model. Our method builds a single multitask model that preserves the capabilities of the original dense experts without requiring additional training. DUME is both cost-efficient and scalable: by leveraging the closed-form solution of ridge regression, it eliminates the need for further optimization and enables experts to be added or removed dynamically. Our extensive experiments on different tasks (on causal language modeling and reasoning settings) with several model sizes (between 115M and 9B parameters) and architectures demonstrate that DUME consistently outperforms baseline approaches. In particular, in a reasoning setting, DUME can also surpass the dense expert model's performance in its domain, achieving 109.9\% of the dense expert's performance on average. Finally, we also show that the DUME model can be fine-tuned to further improve performance in complex reasoning scenarios. Our code is available at this anonymous link: https://anonymous.4open.science/r/DUME-NeurIPS-D551.

Large language models are increasingly deployed in decisions that require culture-dependent moral judgements, yet they answer as if the whole world thinks with a Western mindset. The Moral Machine experiment showed this is wrong at scale: 40 million judgments across 233 countries reveal that moral preferences are systematically structured by culture, and a model that ignores this variation does not merely underperform, but also imposes one society's intuitions on all others. Existing fixes do not scale to global deployment, as fine-tuning needs per-country preference data and GPU budgets, reward-guided decoding needs per-country reward models, and activation steering needs access to model internals that black-box APIs do not expose. In this work, we focus on this realistic inference-time regime, with no weight updates, no training data, and no internal access. The key observation is that within-country demographic disagreement, not consensus, is the steering signal. When culturally grounded personas agree, the base model is already calibrated. But when they disagree, the spread tells us what to fix and how. We propose DISCA (Disagreement-Informed Steering for Cultural Alignment), which instantiates each country as a panel of four World-Values-Survey-grounded persona agents, converts their disagreement into a bounded, loss-averse correction whose magnitude is set by the panel's variance, and shrinks the correction toward zero when the estimate is unreliable. Across 20 countries and 7 open-weight backbones (2B–70B) from five model families, DISCA reduces cultural misalignment on MultiTP by 10–24% on binary moral dilemmas and 2–7% on open-ended scenarios. Furthermore, a smaller 14B backbone with DISCA reaches lower absolute misalignment than a vanilla 70B model.

Modifying 3D assets requires enacting targeted structural transformations while strictly preserving the geometric identity of unedited regions. While recent probability flow models enable training-free editing, bounding these modifications within a high-dimensional latent space presents severe mathematical challenges. Current inversion-free trajectories rely on global velocity updates, which consistently result in spatial bleed into preserved structures and induce trajectory under-commitment, causing edits to stall mid-path. To resolve these dual failures, we introduce FocusFlow, a segmentation-free framework that structures the integration trajectory into two disjoint phases. During early coarse structure formation, FocusFlow applies Feature-Divergence Localization (FDL) to spatially gate the velocity field using intrinsic cross-attention. During late-stage fine-detail refinement, Reference-Guided Generation (RGG) drives topological convergence by pulling the trajectory toward an explicit clean-latent target. Evaluated on the joint Google Scanned Objects and PartObjaverse-Tiny benchmarks, FocusFlow breaks the traditional preservation-modification trade-off. It achieves superior structural preservation while maximizing edit magnitude. By providing stable, localized geometric control without manual spatial annotations, this phase-scheduled approach removes significant technical barriers to the precise modification of synthetic 3D media.

To defend against backdoor attacks on neural networks, the defender must identify the verifier function that determines whether an input is a trigger, rather than rejecting individual triggers in isolation, since an attacker can generate new triggers satisfying the same verifier, against which trigger-by-trigger blocking cannot keep up. If a backdoor verifier could be installed with the same kind of asymmetry as a cryptographic authentication scheme, in which only the attacker can produce trigger inputs while the defender cannot reverse-engineer the verifier from observations, learning-based defense would face a principled impossibility. We show that no such asymmetry can arise for any backdoor installed by training into a model. We establish two complementary impossibility results. First, any verifier installable by training a polynomial-size neural network is reconstructible by the defender with sample complexity matching the attacker's up to polynomial factors. Second, any cryptographic secret the attacker might supply to the model at inference is necessarily observable to a white-box defender. Together, these results rule out any way of giving a self-contained backdoor cryptographic asymmetry. We confirm both routes empirically on pretrained large language models, demonstrating that no configuration yields a backdoor simultaneously effective for the attacker and irrecoverable by the defender.

Inference-time harnesses substantially improve large language models on complex reasoning tasks. However, the intrinsic capabilities of the underlying model remain unchanged by the addition of these external workflows. To bridge this gap, we introduce \emph{On-Policy Harness Self-Distillation} (OPHSD), which employs the harness-augmented current model as a teacher for self-distillation, thereby introducing extra supervisory signals from the harness beyond training data. OPHSD internalizes task-specific harness capabilities into the student model, yielding robust generalizability and strong standalone performance across diverse reasoning tasks. Evaluated across draft--verify harness for text classification and plan--solve for mathematical reasoning tasks, OPHSD consistently outperforms strong baselines (e.g., +10.83\% over OPSD on HMMT25). Our analysis further indicates that reattaching the harness during inference yields no additional benefits and can even degrade performance, suggesting that complex harnesses need not always be permanent fixtures; instead, they can serve as temporary training scaffolds whose benefits are permanently fed back into the base model. Our code and training data are available at \url{https://anonymous.4open.science/r/OPHSD-On-Policy-Harness-Self-Distillation-99F7}.

Chain-of-thought (CoT) monitoring is one of the most promising tools we have for detecting model misbehavior, but its effectiveness depends on models faithfully externalizing their reasoning. Motivated by this vulnerability, we study whether monitor-aware models are capable of obfuscating their reasoning to evade detection. We use synthetic document finetuning to expose eight models to realistic pre-training-style documents describing a CoT monitor and find that monitor-aware models consistently achieve higher rates of undetected misbehavior compared to unaware controls. This effect is weaker but still present on a harder agentic task. We also show that CoT controllability, a model's ability to reshape its own reasoning trace under an imposed constraint, is closely correlated with obfuscation success across the eight models studied ($r=0.800$, $p=0.017$). Monitor-aware models placed under equal reinforcement learning optimization pressure also learn to reward-hack without triggering a CoT monitor substantially faster than unaware controls. Together, these results suggest that knowledge of monitoring combined with high CoT controllability poses a risk to CoT-based monitoring.

Equilibrium Propagation (EP) is a training framework for energy-based models (EBMs) that has attracted interest in the context of neuromorphic computing platforms such as continuous Hopfield networks, nonlinear resistive networks and coupled phase oscillators. However, EP's practical applications have so far remained limited to relatively small-scale problems. Predictive coding networks (PCNs), another class of EBMs rooted in computational neuroscience, are typically trained with a specialized algorithm and have likewise not yet been demonstrated at large scale. In this work, we develop a more effective training method for PCNs which combines the centered variant of EP with a novel equilibration scheme for PCNs. Using this approach, we train a 10-layer convolutional PCN (VGG10) on full-size ImageNet, achieving 13.23% test error rate on the top-5 classification task, close to the 12.2% backpropagation baseline. To our knowledge, this is the first demonstration of both PCNs and EP-based training at ImageNet scale. These results significantly extend the scalability of both approaches and suggest that the primary challenges in scaling EP in other EBMs may not be attributed to inherent limitations of the EP framework.

Unsupervised learning in Vision Transformer models, such as DINO, enables robust visual perception without requiring labeled data by leveraging large-scale image datasets. However, these models lack a focusing mechanism during training, resulting in perceptual representations that are often ambiguous or overly similar. In this work, we introduce two visual focusing loss functions designed to establish correspondences between the model's attention and specific regions within images. Specifically, we leverage the multi-head attention mechanism within the model to selectively steer the model's focus, leading to the emergence of multiple, diverse, and focused perceptual capabilities without requiring supervision. Through both qualitative and quantitative evaluations, we demonstrate that our method substantially increases the spatial selectivity and diversity of attention heads, improving both explainability and fine-grained recognition performance.

Training quality is the primary determinant of reasoning efficiency in large language models, yielding greater returns than proportional increases in inference computation. Evaluating 44 language models (0.5B--685B parameters) across six reasoning benchmarks totalling over 67000 assessments, this study establishes when additional inference tokens are beneficial, redundant or actively harmful. Standard models exhibit near-universal accuracy convergence at approximately 1000 tokens regardless of architecture or parameter count. Causal intervention experiments demonstrate that constraining generation length improves reasoning accuracy by 15.4 percentage points, identifying over-generation rather than capacity exhaustion as the primary efficiency bottleneck. Among 13 reasoning-specialised models, training methodology yields 18.8-fold efficiency improvement at matched parameter scales, exceeding gains from ten-fold parameter increases. Graduate-level evaluation reveals that arithmetic benchmark performance does not predict reasoning capability on complex tasks ($r = -0.03$, $P = 0.91$), exposing fundamental limitations invisible to standard evaluation practice. Mechanistic profiling reveals two orthogonal dimensions of training quality---quality control and decomposition efficiency---that independently predict efficiency profiles, providing a measurable framework linking training decisions to inference-time behaviour.

We introduce Quokka, the first comprehensive scaling law for diffusion language models (DLMs), encompassing both compute- and data-constrained regimes alongside key modeling and optimization designs. Under compute constraints, we find that optimal parameter and dataset sizes scale equally with compute ($N_{\mathrm{opt}} \propto C^{0.5}$, $D_{\mathrm{opt}} \propto C^{0.5}$); however, DLMs are 2--5$\times$ more data-hungry than autoregressive (AR) models, requiring larger corpora and proportionally smaller models for a given FLOP budget. Under data constraints, validation loss follows a U-shaped curve across epochs, with the onset of overfitting scaling as $e_{\mathrm{opt}} \propto U_D^{0.39}/N^{0.55}$. This indicates that optimally leveraging a fixed unique data budget ($U_D$) requires allocating both modestly larger models and more training epochs. Beyond data and parameter allocation, we provide actionable guidance on DLM design choices: absorbing-mask transitions outperform uniform diffusion, linear noise schedules are the most stable and performant, and principled diffusion ELBO objectives ultimately surpass MaskGIT losses despite slower initial convergence. Furthermore, we demonstrate that easy-to-hard noise curricula accelerate early learning, established AR scaling laws for batch size and learning rate transfer seamlessly to DLMs, and weight decay, while unhelpful for single-epoch runs, is essential for multi-epoch regularization.