NeurIPS 2026 Papers

For practical use, diffusion- or flow-based generative models must be aligned with task-specific rewards, such as prompt fidelity or aesthetic preference. That alignment is challenging because the reward is defined for clean output images, but the alignment procedure requires value function estimates at noisy intermediate latents. Existing methods resort to Tweedie-style or Monte Carlo approximations, trading off estimator bias against computational cost: Tweedie estimates are efficient but biased, while Monte Carlo estimates are more accurate but require expensive rollouts. A natural alternative would be a learned value function, but it remains an open question how to effectively train a strong and general value model specifically for noisy latents. Here, we propose StitchVM (Stitched Value Model), a model stitching framework that efficiently transfers reward models pretrained for clean images to the noisy latent regime. StitchVM starts from an existing, truncated pixel-space reward model and attaches a frozen diffusion backbone to it as its head. From the pixel-space model, the resulting hybrid retains a carefully pretrained, robust reward capability; from the diffusion backbone, it inherits its native ability to handle noisy latents. The stitching procedure is exceptionally lightweight, e.g., stitching and finetuning CLIP ViT-L and SD 3.5 Medium takes only $\approx$10 GPU-hours. By lifting powerful pixel-space reward models to latent space, StitchVM opens up a new style of diffusion alignment: instead of rough, yet costly per-sample approximation of the value function, the correct function for the actual, noisy latents is constructed once and then amortized over many samples and iterations. We show that this approach yields improvements across a broad range of downstream steering and post-training methods: DPS becomes $3.2\times$ faster while halving peak GPU memory, and DiffusionNFT becomes $2.3\times$ faster.

Text-guided image editing with diffusion models struggles with spatial edits (e.g. manipulating objects, changing viewpoints, manipulating occlusions) that depend on the 3D structure understanding of a scene. Marigold and follow-up work demonstrate that a pretrained diffusion backbone already harbors strong geometric priors that can be activated for depth estimation, yet how to redirect these priors toward editing remains unresolved. Through per-layer gradient probing, we discover that the editing and depth objectives are compatible in early transformer layers but actively conflict in late layers, explaining why naive depth-conditioned editing either ignores geometry or degrades visual quality. We present StitchEdit, which first activates the backbone's geometric capability via a lightweight depth-denoising adapter, then selectively fine-tunes this adapter for editing using a probe-guided per-layer learning rate schedule that preserves depth knowledge where it helps and prunes it where it conflicts. On the proposed StitchEditBench, a benchmark of 150 geometry-dependent edits spanning spatial transformation, occlusion manipulation, and viewpoint change, StitchEdit consistently surpasses both instruction-based and depth-conditioned baselines on perceptual quality and geometric fidelity.

We introduce Sticky Jump Diffusions (SJDs), continuous-time Markov processes on $\mathbb R^d$ with a discrete anchor set identified with token embeddings. In forward time, anchors release their mass at a hazard rate and the released mass diffuses in the continuous ambient space; time reversal couples a score-driven SDE with a sticky jump kernel whose rate and destination are fixed by flux balance with the forward law. We estimate the score and the per-anchor reverse hazards from a single denoising classifier via Denoising Hazard Matching, the hazard analogue of denoising score matching, with simulation-free cross-entropy training. SJD generalizes the classical sticky-boundary diffusions of Feller and Itô to vocabulary-sized anchor sets, and recovers masked diffusion, continuous diffusion, and hybrid diffusion as limits. Beyond these limits, the framework exposes two design axes that previous work hold fixed: the un-sticking kernel, which encodes both per-anchor geometry and cross-position structure of the corruption, and the per-anchor forward hazard, which encodes the commitment schedule. We evaluate SJD on CIFAR-10, ImageNet $64{\times}64$, Text8, and Sudoku, where it is competitive with discrete, continuous, and hybrid baselines.

Activation steering is a lightweight way to control pretrained generative models: a learned map inserted into the model's activations can steer generation toward target behavior without updating the model weights. While effective for single target behavior, existing steering methods remain brittle when multiple behaviors are targeted. A style intervention followed by a safety intervention can produce a different result than the same applied in the reverse order, as the second intervention is evaluated on activations already shifted by the first. We trace this failure to coordinate sharing: existing transport-based steering methods learn different concepts in the same activation coordinates, so their interventions interfere by construction. To address this, we introduce **StiCAS** (**Sti**fefel **C**ompositional **A**ctivation **S**teering), a geometric steering framework that learns a separate orthogonal subspace for each concept. Each concept is steered by an affine transport restricted to its own subspace, and orthogonality between subspaces makes these transports non-interfering by construction. The concept frame is trained directly on the Stiefel manifold, so Riemannian optimization preserves the orthogonality required for composition throughout learning. This structure gives an exact commutativity guarantee: any sequential ordering of single-concept interventions reproduces the simultaneous multi-concept update. Across 45 concept pairs and three open models, StiCAS drives order-dependent interference exactly to zero and roughly doubles joint-concept success over the strongest activation-transport baseline.

Multilingual studies of social bias in open-ended LLM generation remain limited: most existing benchmarks are English-centric, template-based, or restricted to recognizing pre-specified stereotypes. We introduce StereoTales, a multilingual dataset and evaluation pipeline for systematically studying the emergence of social bias in open-ended LLM generation. The dataset covers 10 languages and 79 socio-demographic attributes, and comprises over 650k stories generated by 23 recent LLMs, each annotated with the socio-demographic profile of the protagonist across 19 dimensions. From these, we apply statistical tests to identify more than 1{,}500 over-represented associations, which we then rate for harmfulness through both a panel of humans (N = 247) and the same LLMs. We report three main findings. \textbf{(i)} Every model we evaluate emits consequential harmful stereotypes in open-ended generation, regardless of size or capabilities, and these associations are largely shared across providers rather than isolated misbehaviors. \textbf{(ii)} Prompt language strongly shapes which stereotypes appear: rather than transferring as a shared set of biases, harmful associations adapt culturally to the prompt language and amplify bias against locally salient protected groups. \textbf{(iii)} Human and LLM harmfulness judgments are broadly aligned (Spearman $\rho=0.62$), with disagreements concentrating on specific attribute classes rather than specific providers. To support further analyses, we release the evaluation code and the dataset, including model generations, attribute annotations, and harmfulness ratings.

We present StereoSplat, a feed-forward 3D Gaussian Splatting architecture designed explicitly for stereo videos to achieve reliable, metric-scale novel-view synthesis. Unlike monocular approaches that neglect the nature of binocular pairs, StereoSplat adopts a native stereo architecture that extracts dense multi-scale latent features and disparity from a foundation stereo backbone and feeds them into a dedicated Gaussian decoder. Our decoder utilizes cross-view feature fusion and gated spatial refinement to directly translate these rich representations into per-pixel 3D Gaussians. To resolve the high redundancy of per-pixel predictions across overlapping views, we propose a learnable depth-adaptive aggregation mechanism that clusters and aggregates primitives in inverse-depth space while preserving fine-grained details. To support robust training and generalization, we introduce XRStereo, a large-scale synthetic dataset of stereo video featuring various rig configurations tailored to match the physical hardware of real-world platforms. Trained exclusively on synthetic data and evaluated across challenging benchmarks, StereoSplat yields substantial improvements in the depth accuracy of rendered novel views over established baselines while maintaining high-fidelity novel view synthesis. By overcoming prior geometric limitations, our framework enables robust zero-shot transfer to real-world domains, offering a scalable geometric foundation for spatial computing in mixed reality, autonomous driving, and robotics.

Existing peptide models saturate natural-sequence proteomics benchmarks, but whether they capture the underlying physicochemical dynamics or merely pattern-match on sequence statistics remains unclear. Public peptide datasets offer no axis along which to test this: non-canonical residues and stereochemical variation are essentially absent. We introduce StereoPep, a benchmark of 48,788 chemically synthesized peptides spanning 11 libraries, produced in the laboratory via high-throughput flow synthesis and characterized experimentally by reverse-phase liquid chromatography retention time. Unlike model-generated synthetic data, every peptide in StereoPep is a real molecule with a real chromatographic measurement. The dataset is constructed around matched diastereomeric pairs (sequences identical in primary structure but differing at a single stereocenter), providing a direct test of whether models capture three-dimensional stereochemical information. StereoPep enables progress across multiple fields: it serves as a benchmark for stereochemistry-aware molecular representations and geometric deep learning, supports adapting protein language models to non-canonical residues, and underpins emerging proteomics multiplexing strategies that encode samples through retention-time shifts. Across specialized retention-time models, graph-based architectures, and protein language models, all approaches predict overall retention reasonably well (Pearson $r \approx 0.80$) but collapse on diastereomer pair discrimination (AUC $\approx 0.55$), revealing that current representations encode atomic composition far better than 3D structure.

Large reasoning models improve with more test-time computation, but often overthink, producing unnecessarily long chains-of-thought that raise cost without improving accuracy. Prior reinforcement learning approaches typically rely on a single outcome reward with trajectory-level length penalties, which cannot distinguish essential from redundant reasoning steps and therefore yield blunt compression. Although recent work incorporates step-level signals, such as offline pruning, supervised data construction, or verifier-based intermediate rewards, reasoning length is rarely treated as an explicit step-level optimization objective during RL. We propose Step-wise Adaptive Penalization (SWAP), a fine-grained framework that allocates length reduction across steps based on intrinsic contribution. We estimate step importance from the model's on-policy log-probability improvement toward the correct answer, then treat excess length as a penalty mass redistributed to penalize low-importance steps more heavily while preserving high-importance reasoning. We optimize with a unified outcome–process advantage within group-relative policy optimization. Extensive experiments demonstrate that SWAP reduces reasoning length by 56.3% on average while improving accuracy by 4.2% relative to the base model.

Recovering executable CAD programs from 3D meshes is fundamentally challenging due to the long-horizon, compositional nature of CAD construction and the need for precise estimation of both discrete operations and continuous parameters. Existing learning-based methods primarily treat mesh-to-CAD reconstruction as one-shot sequence prediction and are largely limited to simple sketch-extrude pipelines, restricting operation diversity and preventing the use of intermediate geometric feedback, with no mechanism for refining generated programs. To address this limitation, we introduce StepCAD, a generative optimization approach that performs step-by-step, geometry-conditioned program synthesis followed by explicit refinement in program space. Given an input mesh, StepCAD first predicts construction sequences using a state-conditioned CAD policy conditioned on both target and intermediate geometry, then refines them through IoU-guided tree search over local program edits. To address the limited operation diversity in prior work, we further introduce ARCADE-1.5M, a large-scale dataset of 1.5M executable CAD programs with diverse operations, long-horizon sequences up to 150+ operations, and 12.5M intermediate state-action transitions. Experiments on multiple CAD reconstruction benchmarks show that StepCAD achieves state-of-the-art reconstruction accuracy, validity, and robustness, with up to 87.2\% relative IoU improvement over the best prior method and increasingly larger gains as shape complexity increases.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve the reasoning capabilities of large language models, but its reward is derived only from the correctness of the final answer and provides no supervision over intermediate reasoning steps. Recent rubric-based methods, such as Rubrics as Rewards (RaR), introduce finer-grained supervision by scoring rollouts against a structured set of evaluation criteria. However, the resulting rubric scores are still aggregated into a single scalar that is applied to the entire response, leading to three structural weaknesses, namely loss of the multi-criterion rubric structure, uniform supervision of correct and incorrect reasoning steps, and reward hacking in the trained model through unbounded self-correction. On a sample of 1{,}000 problems, we find that 18.2\% of steps within answer-correct responses are themselves wrong yet positively rewarded, while 49.9\% of steps within answer-incorrect responses are in fact correct yet penalized. Therefore, we introduce \textbf{Step-wise Rubrics as Rewards (SRaR)}, an RLVR framework that (i) uses an LLM judge to attribute each rubric item to a specific reasoning step, (ii) normalizes the per-step rubric scores across rollouts so that only steps whose quality varies produce a learning signal, and (iii) combines the resulting per-step reward with the standard outcome reward through a decoupled advantage estimator that keeps the outcome-driven baseline stable. To support training, we further build a 16K-problem rubric dataset by contrastively distilling rubric items from correct and flawed reasoning paths sampled from a strong model and verified against the ground-truth answer. Across six mathematical reasoning benchmarks spanning multiple difficulty levels, SRaR improves the average accuracy over RaR by \textbf{3.57} points on Qwen3-8B-Non-Thinking and by \textbf{2.75} points on Qwen3-32B-Non-Thinking, raises the Faithful Reasoning Rate on AIME~2025 from 34.5\% to 46.7\% (every correct-answer trajectory of SRaR uses entirely correct reasoning steps), and reduces the rate of self-correction looping from 48.1\% to 26.5\%.

Diffusion large language models (dLLMs) generate tokens in parallel via iterative denoising, but their bidirectional attention prevents the KV caching used by autoregressive models and recomputes the full attention map at every step, incurring quadratic per-step cost. Existing sparse attention methods for dLLMs compute a sparse mask once at the first denoising step and reuse it throughout, while allocating the sparsity budget uniformly across heads. We argue both choices ignore key structural properties of attention in dLLMs: empirically, attention patterns are only locally consistent across steps and shift abruptly at certain transitions, and different heads concentrate their attention mass to very different extents. To this end, we design Step-dLLM, which periodically refreshes a block-sparse mask at anchor steps along the denoising trajectory and reuses it within each locally consistent interval, while a head-adaptive global budget assigns more blocks to heads with more concentrated attention. Backed by a fused IO-aware Triton kernel for block-sparse attention, Step-dLLM applies to pre-trained LLaDA-1.5 and Dream-7B-Instruct without retraining, achieving a $4.2\times$ kernel-level speedup over dense FlashAttention at 16K context with 90% sparsity (up to $8.5\times$ at 95%), and achieves better model accuracy and lower latency than prior methods.

Verifiable reward training has improved mathematical and coding reasoning, but these domains capture only part of step-by-step decision making. Many real-world tasks require finding a high-value feasible plan among many valid alternatives. We introduce $\text{OPT}\star$, a scalable family of optimization-style tasks for training and evaluating LLM step-by-step optimization-like reasoning along a complexity axis: each task provides a feasibility checker and evaluator, while a complexity parameter expands the search space without requiring new human labels. This motivates studying these tasks in two regimes: (i) solver-guided online policy optimization, which uses a solver as a value oracle for partial states and applies rank-based reward shaping to reinforce better next steps, and (ii) search-based offline RL when such solvers are unavailable. Theoretically, we relate success in large search spaces to the information a reasoner extracts per unit of search budget. Empirically, we ablate the ingredients that make search efficient on $\text{OPT}\star$ and show that training on $\text{OPT}\star$ improves step-by-step optimization-like reasoning.

We propose a nonparametric, kernel-based approach to generative modeling based on Stein operators and particle transport. Our method constructs a continuous-time transport from a simple reference distribution (typically, a standard Gaussian) to a target distribution using a Stein formulation of the continuity equation. Unlike alternative Stein transport methods, which require access to the target density or its gradient, we focus on the sample-only setting and introduce a tractable surrogate objective based on an Ornstein--Uhlenbeck (OU) probability path. This construction allows the velocity field driving the transport to be learned using only samples from the target distribution and Gaussian noise. By restricting the velocity field to a reproducing kernel Hilbert space, we obtain closed-form solutions via kernel ridge regression at each time step. The resulting method is particularly well suited to small data regimes, where heavily parameterised neural generative models may be difficult to train or overfit. We demonstrate the usefulness of the approach for sample-based generative modeling in the small data regime (generative data augmentation), and for posterior emulation in Bayesian inverse problems, where expensive Monte Carlo samplers produce limited sets of posterior samples that must be efficiently augmented.

Machine learning interatomic potentials (MLIPs) enable efficient and accurate atomistic simulations but depend critically on the quality and diversity of the training data. We introduce Stein kernelized molecular dynamics (SKMD), an enhanced sampling method that uses interacting particle dynamics to acquire informative training configurations for the active learning and fine-tuning of MLIPs. SKMD corresponds to a stochastic variant of Stein variational gradient descent that is adapted for molecular dynamics by incorporating asynchronous particle updates and a kernel of global atomic descriptors, which provides a symmetry-aware measure of configurational similarity. Unlike other enhanced samplers used in molecular dynamics, SKMD preserves the Boltzmann distribution as the asymptotic distribution of the dynamics. This property enforces a balance between the exploration of diverse configurations and attraction toward high-probability regions of the energy landscape. We further propose an approach to efficient online data acquisition using an adaptive stopping criterion that selects non-redundant training data over the course of simulation. We demonstrate SKMD for the active learning of a neural network model of the Muller-Brown potential and the fine-tuning of a MACE interatomic potential for alanine dipeptide. Compared to active learning baselines, our method achieves higher model accuracy in fewer training iterations, with the same number of acquired training samples.

Large language models (LLMs) are typically governed by post-training alignment (e.g., RLHF or DPO), which yields a largely static policy during deployment and inference. However, real-world safety is a full-lifecycle problem: static defenses degrade against evolving jailbreak behaviors, and fixed weights cannot adapt to pluralistic, time-varying safety norms. This motivates inference-time governance that steers behavior without costly retraining. To address this, we introduce the Consensus Clustering LinUCB Bandit (CCLUB), a unified framework for adaptive social alignment via system-prompt routing. CCLUB employs a conservative consensus clustering mechanism: it pools data only within the intersection of utility and safety similarity graphs, effectively preventing unsafe generalization across semantically proximal but risk-divergent contexts. Our theoretical analysis gives an expected regret bound $O\left(d\log T/(p_{\min}\gamma^2\lambda_x) + d\sqrt{MT}\log T\right)$, separating the logarithmic cost of identifying safe consensus clusters from the dominant cluster-level exploitation term. Experiments show that CCLUB improves the safety--utility trade-off and offline deployment gap, and improves cumulative reward by 10.98\% over the strongest non-prototype baseline.

As state-of-the-art text-to-image flow models have matured to deliver near-photorealistic quality, controlling what they generate -- e.g., inhibiting harmful content while promoting benign alternatives -- has become a central challenge. The current steering paradigm consists of adding a global steering vector to selected model activations. While functional, a fixed vector applied example-agnostic and uniformly along the entire trajectory cannot adapt to the changing state of the generation, and causes uncontrolled global changes beyond the targeted concepts. We introduce Steering Fields, a generalization of steering vectors that adaptively re-estimates the steering direction at each step of the generative process. Steering Fields operate on latent representations, expose a continuous trade-off between steering strength and content preservation, and are compositional, enabling the simultaneous induction and inhibition of concepts, setting a new state of the art on safety steering benchmarks. Although our formulation prescribes no explicit spatial mask or object-level prior, the trajectory-adaptive estimation naturally preserves local structure, in a manner reminiscent of image editing. In fact, when applied in the VAE latent space, Steering Fields can serve as a structure-preserving image-editing technique that achieves state-of-the-art semantic fidelity (CLIP, VQAScore), while remaining model-agnostic and inversion-free

Activation steering is a practical post-training model alignment technique to enhance the utility of Large Language Models (LLMs). Prior to deploying a model as a service, developers can steer a pre-trained model toward specific behavioral objectives, such as better truthfulness, or reasoning ability, without the need for retraining. Conceptually, these methods implement behavior control through hidden-state interventions, without changing the underlying model parameters. However, this capability unintentionally introduces critical and under-explored safety risks. We identify a phenomenon termed **Steering Externalities**, where steering vectors derived from benign datasets—such as reducing harmless refusals, improving structured-output following, truthfulness, and reasoning performance—inadvertently erode safety guardrails. Experiments reveal that these interventions act as a force multiplier, creating new vulnerabilities to jailbreaks and increasing attack success rates to over 80\% on standard benchmarks by bypassing the initial safety alignment. Ultimately, our results expose a critical blind spot in deployment: benign activation steering can erode the ``safety margin,'' rendering models more vulnerable to black-box attacks and indicating that inference-time utility improvements must be rigorously audited for unintended safety externalities.

Visual text editing aims to precisely modify text in images and videos while preserving stylistic consistency and visual realism. Despite significant advances in the image domain, video text editing remains largely unexplored: it is a localized task demanding stroke-level precision within small text regions, which compounds the challenges of cross-frame accuracy, temporal coherence, and stylistic fidelity. We introduce SteerVTE, a unified framework that \underline{\textbf{steer}}s a frozen video diffusion model to perform precise \underline{\textbf{V}}ideo \underline{\textbf{T}}ext \underline{\textbf{E}}diting through style and glyph control. Built on a frozen diffusion transformer, SteerVTE attaches a lightweight text context adapter with two complementary modules: a style encoder capturing the original text's visual attributes, and dual-granularity glyph encoders encoding the target text at both the line and character levels. To overcome the inherently weak text rendering priors of video foundation models, we further propose a glyph-aware spatial-focal loss and a three-stage progressive training curriculum that scales from image to video data. To support large-scale training, we also develop an automatic synthesis pipeline and construct SteerVTE-1M, a dataset of one million triplets spanning diverse scenes, fonts, and stylistic effects. Extensive experiments demonstrate that SteerVTE substantially outperforms existing video editing baselines across text accuracy, style consistency, and temporal coherence. Our anonymous project website is available at \url{https://steervte-paper.github.io}.

Time series forecasting aims to predict future values from historical observations and auxiliary features. We propose \textbf{SteerCast}, a retrieval-based latent steering method that improves decoder-only forecaster at inference time, without updating its parameters. SteerCast constructs a database from the training set by storing a representation of each history window together with a \emph{steering vector} computed in the forecaster's latent space, defined as the difference between representations induced by the ground-truth continuation and by the model's own prediction. At test time, SteerCast retrieves nearest neighbors for a query history, aggregates their steering vectors, and injects the resulting signal into the forecaster's hidden states at every step of autoregressive generation, guiding predictions toward trajectories consistent with similar training cases. Experiments across diverse multivariate benchmarks and multiple horizons show that SteerCast consistently improves forecasting accuracy over the fine-tuned backbone and retrieval-based baselines, while requiring no additional training beyond the original fine-tuning and using only the training set as a retrieval corpus.