NeurIPS 2026 Papers

Reliable generative AI models critically rely on expert human annotations to evaluate output quality, yet these "gold'' labels are expensive to collect and limited in quantity. Organizations thus often turn to collecting vast but noisy "silver'' labels from crowdsourced workers or vendor annotators as proxies for gold labels. Because gold remains the evaluation target, naively aggregating noisy silver labels may introduce bias, and estimators built on sparsely observed gold labels may have high variance to resolve the model performance gaps that guide practical decisions. Model evaluation has become an ongoing operational practice rather than a one-time exercise, with evaluation rounds repeating across model versions, releases, and content domains. A natural question is whether the previous historical evaluation data can be used to improve each new round of evaluation. We introduce HERO (History Enhanced RObust model evaluation), a novel framework that uses historical data to reduce both bias and variance in model performance evaluation. HERO calibrates silver labelers' performance learned from historical gold annotations, and stabilizes the resulting estimator by anchoring it to covariate information measured with high precision in the historical data. HERO can be broadly applied across multiple common evaluation tasks, and remains valid when only a subset of historical labelers appears in the current round. We establish conditions under which the bias and variance reductions hold, showcase HERO's performance in simulation studies, and demonstrate its effectiveness on real-world model evaluation benchmarking datasets.

Federated continual learning (FCL) evaluates how distributed clients learn from changing data streams while retaining previously learned knowledge. Existing evaluations are difficult to compare because they often change datasets, task splits, client data splits, task orders, backbones, memory assumptions, and reporting rules simultaneously. We introduce \textbf{HERO}, a heterogeneity-aware benchmark library for FCL. HERO builds benchmark streams by separating three choices that are often coupled, namely the task split, the client data split, and the client task sequence. In HERO-Core, the main comparable benchmark, $\alpha$ controls client data skew and $\rho$ controls task-order mismatch. We evaluate representative FCL methods on CIFAR-100 and TinyImageNet using final average accuracy, average forgetting, and bottom-10\% client accuracy. We also include a graph-based Domain-IL portability case study on OGB-MolPCBA, where scaffold-domain granularity changes the input distribution while the prediction task remains fixed. Our results show that method behavior changes across easy and heterogeneous settings, that average accuracy can hide weak bottom-client performance, that task-order mismatch favors different strategies from synchronized evaluation, and that the same HERO interface can expose domain-shift difficulty beyond image-based FCIL. HERO releases benchmark streams, configurations, method implementations, and reporting scripts to support reproducible and setting-aware FCL evaluation.

Individual-level genotype data underpin genetic discovery, disease-risk modeling, and therapeutic development, yet many critical settings remain data-limited, including underrepresented ancestries, rare or imbalanced disease cohorts, and restricted-access biobanks. Conditional synthetic genome generation would support privacy-aware benchmarking, simulation, and data augmentation, but requires learning an ultra-high-dimensional discrete distribution over hundreds of thousands to tens of millions of correlated single-nucleotide polymorphisms (SNPs), or sites of DNA variation. We introduce $\mathbf{HELICS}$: $\mathbf{H}$igh-fidelity g$\mathbf{E}$neration via $\mathbf{L}$atent $\mathbf{I}$nterpolant $\mathbf{C}$onditional $\mathbf{S}$ynthesis, a phasing-free latent flow matching framework for genome-wide SNP-level generation from biobank-scale genotype data. HELICS tokenizes chromosomes with convolutional autoencoders, reducing $>$$600{\rm K}$ SNPs to $4{,}096$ continuous latent tokens, and trains a transformer parameterized conditional flow matching model over the concatenated multi-chromosome latent representation. Trained on roughly $428\rm K$ genomes in the UK Biobank, HELICS generates genome-wide synthetic cohorts without haplotype phasing and achieves stronger fidelity to held-out real genotypes than existing haplotype- and genotype-based simulators, including the highest validation allele-frequency agreement ($R^2=0.992$) and superior preservation of local and long-range SNP correlation structure. Conditioning on ancestry and genetic liability summaries enables targeted generation across genetic risk profiles; in perturbation analyses, by increasing the conditioned disease trait liability, HELICS induces SNP-level dosage changes in generated genomes that are highly correlated with published association weights. HELICS provides a scalable generative modeling framework for SNP-level genome data and a path toward conditional synthetic cohorts for data-limited genetic studies.

Distilling vision-language models into faster hybrid architectures, such as 3:1 Mamba-2/attention mixes, is now standard practice for making inference efficient. Aggregate benchmarks suggest that this works but they hide selective failures. When we distill Qwen3-VL-8B-Instruct into a 3:1 Mamba-2/attention hybrid, student model stays within 2 points of the teacher across visual reasoning benchmarks like MMStar, MMBench, and MMMU-Pro, while dropping 13 points on optical-character-recognition and document tasks. The student can still understand the scene but loses the fine-grained text needed to answer. We localize much of the failure to a specific kind of position. In a high-resolution image, most patches are sky, wall, or smooth texture, while a small fraction carries text, edges, object boundaries, or other local details. In a token-level diagnostic, the top 10\% highest-density patches have 3.6$\times$ larger residual drift than the bottom 10\% lowest-density patches and 3.5$\times$ larger teacher-masking answer contribution. Uniform weighting devotes many loss terms to low-information background patches, whereas sparse answer-bearing patches receive no special protection. The required intervention is minimal: we replace uniform residual alignment with density-weighted residual alignment, using patch self-dissimilarity as a training-free proxy for position importance. We call this HEED. Compared with normal end-to-end distillation, HEED increases performance by 8.7 points on OCRBench v2 and 5.13 points on a 10-benchmark average. The gain is realized on different teacher models and hybrid architectures. After standard post-training, the student reaches teacher-level performance on the 10-benchmark average with a 4.12$\times$ throughput and a 68\% memory saving at 128k context, with no additional parameters and no inference-time cost.

In LLM-based recommendation, memory effectively plays the role of a user-level agent, typically built on top of a memory-augmented backbone. Yet under this paradigm users and items remain asymmetric participants: only the user is given a structured profile; that profile is consumed inside a bandwidth-limited reranker prompt; and it is frozen after warmup. Starting from this observation, we recast LLM-based recommendation as an agentic-communication process in four stages --- Build a self-description, Transmit it along the collaborative graph, Understand the other side, and Sync it from the next interaction. Analysing the dominant memory-augmented paradigm through this formulation, we identify three open communication challenges that current pipelines leave unaddressed: the item side never produces a comparable self-description (one-sided Build); profile-level signals are routed only into a bandwidth-limited reranker prompt rather than into the retrieval stage where they would not compete for context tokens (prompt-trapped Understand); and the user's description is frozen after warmup with no principled update rule (absent Sync). We instantiate the missing stages as HARP (Homophily-aware Agentic Recommendation via dual Profiles}), a training-free, three-operator addition to a frozen LLM backbone: a rule-extracted symmetric item profile (zero extra LLM calls), a channel-decomposable profile-level homophily $\mathcal{H}(\phi_u,\phi_i)$ consumed at retrieval time rather than inside the reranker prompt, and a critic-gated single-pass reflective update realising in-context profile evolution without any gradient step. On four public benchmarks (Yelp, Amazon Books, MovieTV, Goodreads), HARP lifts Hit@1 over the strongest published baseline by $6.9\%$--$24.0\%$, with family-wise Holm-corrected $p<10^{-9}$ on every dataset. Ablations and a per-case attribution study isolate each operator as a distinct, complementary source of gain; we also quantify and openly report the trade-off introduced by the reflective pass. Our contribution is a unified formulation of agentic communication for LLM-based recommendation, together with a training-free three-operator instantiation on top of it.

In standard acoustic anomaly detection (ASD), models are usually trained separately for each machine or domain using explicit metadata. However, in realistic deployments, machine identifiers are incomplete, unreliable, or simply unavailable. Thus, metadata-free universal ASD asks a single model to monitor mixed machine populations without machine IDs. In this paper, we argue that the main failure mode is not a weak detector family, but representation entanglement. The heterogeneous normal modes overlap in frozen pre-trained feature spaces, making anomaly scoring and approximate retrieval unreliable. We study this problem directly and propose HARMONY (Hierarchical Anchor Retrieval on Manifold for Oblivious-source acoustic aNomaly detection), an anchor-guided geometric repartitioning framework that reorganizes mixed-source features into compact Voronoi regions and reuses this structure for hierarchical retrieval. On dataset DCASE 2020 and MIMII, HARMONY improves mean AUC from 75.40% to 90.00% (+14.60% absolute improvement) over the strongest unified baseline and from 83.90% to 90.00% over a strong foundation-feature baseline. On a single CPU, its hierarchical retrieval achieves a 7.8× speedup over exhaustive search with only a 0.20% AUC drop. These results suggest that incorporating explicit geometric partitioning can improve both detection accuracy and retrieval efficiency for ASD. Code and data are available at the anonymous URL: https://anonymous.4open.science/r/HARMONY-26CC.

Understanding how aligned LLMs internally represent safety is critical for diagnosing alignment vulnerabilities, as it explains why jailbreaks succeed and informs the design of robust alignment strategies. Prior work shows that aligned LLMs encode harmfulness and refusal as separable directions in the residual stream at prompt-side token positions. We show that jailbreaks succeed at prompt encoding by suppressing either the refusal or harmfulness direction before any token is generated, with distinct attack classes occupying separable regions of the harmfulness-refusal plane. Extending the analysis to response-token positions, we find that the model recognizes harmful content while it is generating that content, even when it failed to recognize the input as harmful at the prompt side. Motivated by our findings, we introduce HARC (Harmfulness-And-Refusal Coupling), a fine-tuning method that pairs the two directions across both prompt and response positions. Since the intervention is confined to the harmfulness-refusal subspace, it leaves the rest of the residual stream intact and does not degrade general capability or inflate over-refusal. Across extensive experiments, HARC achieves the strongest robustness-capability-usability trade-off among six baselines spanning the major training-time and inference-time safety methods. The harmfulness and refusal directions at prompt and response positions transfer across the five model families and two scales we tested without architecture-specific tuning.

Large language model (LLM) routing aims to exploit the specialized strengths of different LLMs for diverse tasks. However, existing approaches typically focus only on selecting LLM architectures, treating candidate models as static black boxes while overlooking parameter adaptation, which can substantially affect task performance. In this paper, we introduce HAPS, a hierarchical LLM routing framework that jointly searches over model architectures and parameters. Specifically, a high-level router selects among candidate LLM architectures, while a low-level router generates input-conditioned LoRA parameters for the selected architecture. To couple these two decisions, we design a shared parameter generation mechanism that enables cross-level knowledge transfer between architecture routing and parameter adaptation. We further optimize the whole framework with a reward-augmented training objective. Extensive experiments show that HAPS consistently outperforms strong routing baselines. We further conduct broad ablations and analyses on parameter sharing, architectural choices, scalability, and efficiency, providing additional evidence for the effectiveness and practicality of the proposed framework. We have released our code at https://anonymous.4open.science/r/HAPS_private-68CB.

Physical reasoning about humans in videos has largely focused on everyday interactions, where contact, pressure, pose, and motion support embodied perception and contact-aware applications. Still, short and forceful properties like impact events remain relatively underexplored, despite their relevance to broader understanding of how abrupt physical events affect the human body. To fill this gap, we introduce human-centric impact localization, a new task that localizes target physical impacts temporally within a video and spatially on the human body. We further present HAPACT, a HumAn-centric imPACT dataset and benchmark built from movie clips, containing over 2.8K video-query pairs with more than 20K fine-grained human-annotated frames. We built HAPACT around viewer-grounded annotation, text-conditioned localization, and supervision at frame-level temporal and SMPL-X-based spatial granularity. To demonstrate the effectiveness of our benchmark, we conduct extensive evaluations of existing methods, employing video moment retrieval models for temporal localization and human-scene contact models for spatial localization. For the spatial task, we propose a text-conditioned baseline by reformulating localization as query-guided body-region selection. Benchmark results show that physical impact localization remains challenging for existing methods, motivating HAPACT as a dedicated benchmark to support physical impact reasoning in videos.

Supervised post-training, including supervised fine-tuning (SFT) and knowledge distillation (KD), is essential for adapting large language models (LLMs) to specialized domains. However, traditional supervised post-training methods often suffer from catastrophic forgetting, training instability, and prohibitive computational costs. While layer-wise strategies have emerged as efficient alternatives, they introduce a representation bottleneck, where constrained updates in early stages limit feature diversity and impair final generalization. To resolve these issues, we develop HALO (Homotopy-Augmented Layer Optimization), a homotopy-driven layer-wise strategy for LLM post-training. By introducing a continuous homotopy parameter and a sequence of monotonically increasing auxiliary functions, we formulate the post-training process as a continuous deformation of the optimization problem that progressively incorporates layers into the trainable set. Specifically, HALO unfreezes the model from the output back to the input in a structured sequence. We theoretically prove that this approach establishes a differentiable solution path, ensuring a stable transition from a restricted optimization state to a fully converged LLM. Extensive experiments across diverse model families and sizes demonstrate that HALO consistently enhances performance and generalization. Notably, HALO exhibits strong robustness in challenging scenarios, especially in low-resource settings, while significantly reducing convergence time and iteration counts. Our results position HALO as an effective and accessible framework for high-performing LLM specialization.

Visual geometry grounded transformers are strong feed-forward backbones for multi-view 3D reconstruction, but their global attention becomes the dominant inference bottleneck as views and tokens increase. Existing acceleration methods mainly focus on designing compression operators, while compression strengths are often assigned by uniform schedules or offline calibration, overlooking the heterogeneous sensitivity of layers and attention heads during inference. We therefore propose HALO-VGGT, a lightweight online compression allocator requiring no training or calibration for efficient VGGT-style reconstruction. HALO uses a sampled query-key margin proxy to estimate compression sensitivity at the layer and head levels, and assigns hierarchical compression ratios accordingly. Instead of introducing a new compression operator, HALO complements pruning, merging, and sparse attention mechanisms by deciding where and how aggressively they should be applied. HALO improves the trade-off between speedup and accuracy across datasets and backbones, achieving up to 4.20$\times$ speedup and further improving existing merging and sparsity methods by up to 1.46$\times$.

Multi-view bimanual robot world models must predict future observations while preserving scene identity, following two synchronized but non-exchangeable arms, and maintaining consistency between global and wrist-local views. Existing conditioning schemes often collapse left- and right-arm actions into a view-agnostic control signal, making long-horizon rollout prone to scene drift, wrong-arm responses, and cross-view inconsistency. We propose \textbf{HAI}, a \textbf{Hierarchical Anchored Interaction} architecture for controllable multi-view bimanual world modeling. HAI organizes prediction as a structured information flow. First, hierarchical action-view conditioning routes the bimanual action chunk to camera streams, injects coarse chunk-level intent, performs structured multi-view modeling on the coarse-modulated features, and then injects fine per-frame action control. Second, anchored dynamic generation fuses the fine action-conditioned descriptors with persistent per-view scene anchors before decoding future observations. Experiments on AgiBot and DROID show that HAI improves long-horizon rollout quality over the baselines. Ablations confirm gains in scene stability, wrist-view controllability, and cross-view consistency. Further experiments on policy-improvement diagnostics indicate more useful synthetic rollout signals for downstream VLA policies, achieving an averaging 67.2\% relative gain in success rate. The project page is available at \url{https://hai-anon.github.io/}.

Hierarchical 3D grouping aims to recover scene groups across multiple granularities, from fine object parts to complete objects, without relying on semantic labels or a fixed vocabulary. The main challenge is to transform 2D foundation-model cues into coherent hierarchy supervision and embed that hierarchy in a 3D representation. We propose H2G, a hyperbolic affinity field for hierarchical 3D grouping. Our method derives semantically organized tree supervision by interpreting foundation-model affinities through Dasgupta's objective for similarity-based hierarchical clustering. This supervision is distilled into a single Lorentz hyperbolic feature field, whose geometry is well suited for tree-like branching structures. A hierarchy-aware objective aligns the field with fine-level assignments, coarse object structure, compact feature clusters, and LCA (Lowest Common Ancestor) ordering. This formulation represents multiple grouping levels in one feature space, enabling semantic hierarchical grouping grounded in 2D foundation-model knowledge.

Parametric human models capture global pose but overlook fine-grained surface dynamics. Generic scene flow estimates dense motion but struggles with articulated humans and lacks 4D ground truth. To bridge this gap, we introduce H-Flow, a dense 4D motion representation that captures non-rigid deformations beyond skeletal kinematics. We estimate H-Flow from monocular video via a unified multi-head architecture that jointly perceives pose, depth, and flow. To overcome the scarcity of dense motion annotations, we propose a self-supervised learning paradigm that embeds geometric, structural, and biomechanical priors into optimization. These cross-modal constraints tightly couple pose, depth, and flow, so that improving any one modality simultaneously drives the others toward consistency. For evaluation, we present DynAct-4D, a high-fidelity synthetic benchmark providing dense 4D ground truth for complex human movements. Extensive experiments show that our method outperforms state-of-the-art scene flow baselines. Furthermore, H-Flow serves as an effective motion primitive, bringing substantial improvements to downstream tasks including action recognition and video generation. Models, code, and resources will be released upon publication.

Speculative sampling is a widely used method for losslessly accelerating Large Language Model (LLM) inference. State-of-the-art speculative sampling methods (e.g. EAGLE series) use a ``dynamic draft tree'' with drafter temperature implicitly set to $T=0$. Greedy drafts, however, perform poorly as temperature increases. Moreover, we show that naively instantiating the dynamic draft-tree with drafter temperature $T>0$ leads to an incorrect distribution. Fixes exist, but are computationally prohibitive as they require $O(|\mathcal{V}|^k)$ computation, where $\mathcal{V}$ and $k$ denote vocab and number of child nodes. We address these limitations by introducing Gumbo, a new speculative sampling algorithm designed for high-temperature settings that uses only $O(|\mathcal{V}|)$ compute. Gumbo is drafter-invariant in that it does not require the draft distribution during verification. Additionally, Gumbo increases the average acceptance length and speedup by expanding and reranking the dynamic draft tree based on Gumbel scores rather than draft probabilities. At the heart of Gumbo is a multi-draft generalization of the communication-free coupling principles and Gumbel sampling developed by Daliri et al., which is of independent interest. Replacing EAGLE-3's speculative sampler with Gumbo results in up to a +26.7\% increase in speedup at a temperature of 1.0 across four models and five datasets, without altering the draft model or requiring any other code changes.

Classifier-free guidance (CFG) has become a widely adopted and practical approach for enhancing generation quality and improving condition alignment. Recent studies have explored guidance mechanisms for unconditional generation, yet these approaches remain fundamentally tied to assumptions specific to diffusion models. In this work, we propose a spectrum weakening framework for visual autoregressive (AR) models. The method is training-free and condition-free by constructing a controllable weak model in the spectral domain without architectural modifications. We theoretically show that invertible spectral transformations preserve information, while selectively retaining only a subset of the spectrum introduces controlled information reduction. Based on this insight, we perform spectrum selection along the channel dimension of internal representations, which avoids the structural constraints imposed by diffusion models. We further introduce a spectrum-renormalization technique that maintains numerical stability during the weakening process. Comprehensive empirical and ablation studies confirm the effectiveness of our approach, including raster-scan, random-order, and scale-wise AR, spanning discrete and continuous modeling and covering class and text-condition scenarios, demonstrating high-quality unconditional generation and strong prompt alignment maintenance for conditional generation. Code will be made available.

Machine learning models often achieve high average accuracy in the training data while performing poorly on test data due to subpopulation shift. A common data-centric solution is to enforce balanced subgroup proportions, but full balance can be costly, infeasible, and not always necessary. In this work, we study whether balanced training data is the unique optimal configuration for robust subpopulation generalization. We analyze the theoretical performance across training subgroup distributions and show that multiple configurations, including substantially imbalanced ones, can achieve performance comparable to the balanced configuration. We further derive gradient-based directions for adjusting subgroup proportions to improve robustness more effectively. These directions can diverge from the direct path toward balance, suggesting that balancing is not always the most efficient data-allocation strategy. We validate the theoretical insights with controlled experiments on image and text benchmarks, showing that gradient-guided allocation can improve robustness more efficiently than directly enforcing balance. These findings suggest that strategically allocating data offers a more flexible and principled path to robust performance than simple balancing.

There is a growing need for understanding how trained machine learning models behave beyond standard predictive performance. With this work, we aim to understand trained machine learning models by questioning their data preferences. We propose a mathematical framework for guided data generation that allows us to produce fixed-label, prediction-risky, parameter-sensitive, or model-contrastive samples, among others. To showcase our framework, we pose these queries to a range of models trained on a range of classification and regression tasks, with answers in the form of generated data.

Genomic Foundation Models (GFMs) have become a promising paradigm for decoding DNA sequences, yet their privacy risks are critically underexplored.Trained on human cohorts, GFMs can unintentionally memorize sensitive genomic signatures, that enable membership inference attacks (MIAs) and expose individuals to long-lasting privacy harms.However, existing defenses often fall short in genomics: globally applied mechanisms can substantially degrade motif-sensitive utility, and model-agnostic post-processing provides limited insight into which genomic regions drive leakage. Through a systematic privacy audit, we find that membership leakage in GFMs is highly non-uniform, concentrating on a small subset of high-risk samples and sparse token spans corresponding to biologically meaningful patterns. Motivated by this finding, we propose GenoGuard, a model- and sample-adaptive defense that shifts from blanket protection to targeted, region-aware mitigation. GenoGuard (i) localizes privacy-leaking subsequences via gradient-based token attribution and performs token-level risk reduction through selective gradient routing, (ii) applies sample-level risk reduction with risk-adaptive label smoothing guided by a reference-calibrated loss gap score, and (iii) provides region-level attribution maps for privacy auditing and biological interpretation. Experiments across representative GFMs (Mistral-DNA, Nucleotide Transformer) demonstrate that GenoGuard consistently improves privacy against multiple MIA families while preserving strong fine-tuning performance and interpretability.