NeurIPS 2026 论文列表

Contextual Integrity (CI) defines privacy not merely as keeping information hidden, but as governing information flows according to the norms of a given context. As large language models are increasingly deployed as personal agents handling sensitive workflows, adhering to CI becomes critical. However, even frontier models remain unreliable in making disclosure decisions, and existing mitigation strategies often degrade underlying task performance. To overcome this privacy-utility trade-off, we propose SelfCI, a complementary self-distillation framework that decouples information suppression from task resolution. SelfCI jointly optimizes two independent reverse KL divergences over distinct teacher distributions derived from feedback: one encourages preserving task-relevant information for utility, while the other enforces minimal and appropriate disclosure. This complementary formulation induces a Product-of-Experts (PoE) target, aligning the policy with the intersection of capability and privacy requirements. Without relying on costly external supervision, empirical evaluations demonstrate that SelfCI consistently outperforms competitive baselines such as online reinforcement learning algorithms (e.g., GRPO). These trends further extend to out-of-domain settings involving agentic workflows and accumulated private context, suggesting that SelfCI provides a practical path toward CI alignment.

Neural posterior estimation has emerged as a powerful tool for amortized inference, with growing adoption across scientific and applied domains. In many of these applications, the conditioning variable is a set of observations whose elements depend not only on the target but also on unknown factors shared across the set. Optimal inference therefore requires treating the set jointly, which in turn requires training the estimator at the deployment set size---a regime where memory and compute quickly become prohibitive. We introduce a simple, theoretically grounded strategy that decouples representation learning from posterior modeling. Our method trains a mean-pool Deep Set on sets of size at most two, producing an encoder that generalizes to arbitrary set sizes. The inference head is then finetuned on pre-aggregated embeddings, making training cost essentially independent of the deployment set size $N$. Across scalar, image, multi-view 3D, molecular, and high-dimensional conditional generation benchmarks with $N$ in the thousands, our approach matches or outperforms standard baselines at a fraction of the compute.

Many sequential Bayesian methods maintain a whitened operator $A_w = L^{-1}ML^{-\top}$, where $L$ is the Cholesky factor of a kernel, precision, or curvature matrix. When the underlying matrix changes by a rank-$1$ perturbation, existing techniques propagate only one-sided quantities $L^{-1}B$; refreshing the sandwiched operator $A_w$ requires a full $O(r^3)$ re-factorization that dominates iteration cost, consuming up to 91% of per-step time in natural gradient methods. We show that this cost is unnecessary. For pre-whitened inputs, intermediate factors of $L$ cancel exactly from both sides of the augmented Givens rotations, and the updated $(L^+)^{-1}M^+(L^+)^{-\top}$ can be recovered in $O(r^2)$ time by applying the rotations double-sidedly to a zero-padded matrix, followed by a rank-1 correction. We provide a self-contained derivation, stability bounds, and implementations as a C++ extension, MLX kernel, and Triton GPU kernel. On streaming sparse GP regression, online Bayesian optimization on protein fitness landscapes, and natural gradient preconditioning, the method matches full re-factorization in accuracy with up to $5.4\times$ speedups. At $r = 1024$, training time drops from 1075s to 198s, making larger curvature charts affordable and producing strictly better test MSE.

Introduced is a methodology for adapting the topology of dense neural networks, enabled by isotropic activation functions. Achieved through prescribed reparameterisation symmetries and singular-value decomposition of affine maps, this diagonalises layers into one-to-one, ordered connections. This makes it simpler to assess the impact of individual connections on the function. Low-impact neurons can be removed (neurodegeneration), and a thresholded buffer of largely inactive 'scaffold' neurons is maintained (neurogenesis). These symmetry-led diagonalisation and structural changes are function-invariant, demonstrated to be computationally identical during neurogenesis, arbitrarily well approximated during neurodegeneration, and enable asymptotic 50% parameter sparsification of dense networks with identically preserved function. Thus, real-time restructuring of the architecture in response to task demands, task appending, removal or changes is shown. The approach is conceptually centred on primitive symmetry-prescriptions, through which isotropic functions are derived that feature explicit basis independence and a loss in the individuation of neurons implicit in typical elementwise functional forms. Hence, this allows freedom in the basis to which layers are decomposed and interpreted as individual artificial neurons, directly enabling this adaptive topology approach. Additionally, a new tunable model parameter, the 'intrinsic length', is introduced to improve this analytical invariance, alongside a generalised isotropic-perceptron architecture that enables parallel precomputation of all matrix-vector products and displays a nested functional class. Diagonalisation is suggested to offer new possibilities for interpretability and monitoring of isotropic networks.

Current sign language recognition and translation benchmarks are commonly collected in controlled settings using fixed or studio-style cameras, particularly limiting their ability to reflect how signers communicate in everyday mobile scenarios. As such, this gap becomes increasingly evident when considering the Arabic Sign Language (ArSL), which is a primary sign language used across Arab countries. In this work, we introduce Isharah-Selfie, a large-scale ArSL dataset collected entirely using front-facing smartphone cameras under unconstrained selfie conditions. The dataset comprises 8,997 video clips performed by 16 signers across 1,493 unique sentences, covering several domains such as healthcare, education, and transportation. Unlike conventional datasets dominated by fixed-camera, two-handed signing, Isharah-Selfie captures realistic mobile communication characterized by one-handed signing, close-range viewpoints, partial body visibility, camera motion, unconstrained environments, and heterogeneous device resolutions. Each video is annotated with both gloss sequences for continuous sign language recognition (CSLR) and Arabic sentence translations for sign language translation (SLT). Moreover, we define signer-independent and unseen-sentences evaluation protocols to assess generalization across unseen signers and novel sentence compositions. Extensive benchmarks using RGB-based and pose-based data show that current CSLR and SLT methods remain challenged by selfie-style signing conditions, particularly under unseen-sentence evaluation, where compositional generalization remains weak. By releasing Isharah-Selfie, we aim to support research toward robust, user-centered sign language technologies that operate reliably in realistic smartphone-based communication settings. The Isharah-Selfie dataset is available on https://www.github.com/.

The linear representation hypothesis (LRH) holds that high-level concepts are encoded as linear directions in a model's feature space, and grounds much of mechanistic interpretability, from probing to sparse autoencoders to steering vectors. We argue the standard test for it is too lenient. A high linear-probe accuracy shows that a concept is linearly readable from the representation, not that it is one of the representation's privileged directions of variance, which is what the LRH actually claims. These two readings come apart in any high-dimensional, richly structured representation, and the gap matters. The structure that linear probing ignores is precisely where compositional generalization, out-of-distribution behavior, and the validity of post-hoc feature decomposition live. We propose three asymmetric diagnostics that operationalize the strong reading: reverse predictivity, spectral concentration, and out-of-distribution generalization. We validate them in a fully observed simulation where ground truth is known, then apply them to pretrained vision encoders on SVG-World, our deterministic generative benchmark with paired visual styles. Pretraining produces structure beyond what a forward probe can certify, but the linear-feature reading systematically overstates how much. Probing on its own is not enough to license the mechanistic claims that rest on it.

Current Large Language Models (LLMs) are gradually exploited in practically valuable agentic workflows such as Deep Research, E-commerce recommendation, job recruitment. In these applications, LLMs need to select some optimal solutions from massive candidates, which we term as \textit{LLM-as-a-Recommender} paradigm. However, the reliability of using LLM agents for recommendations is underexplored. In this work, we introduce a \textbf{Bias} \textbf{Rec}ommendation \textbf{Bench}mark (\textbf{BiasRecBench}) to highlight the critical vulnerability of such agents to biases in high-value real-world tasks. The benchmark includes three practical domains: paper review, e-commerce, and job recruitment. We construct a \textsc{Bias Synthesis Pipeline with Calibrated Quality Margins} that 1) synthesizes evaluation data by controlling the quality gap between optimal and sub-optimal options to provide a calibrated testbed to elicit the vulnerability to biases; 2) injects contextual biases that are logical and suitable for option contexts. Extensive experiments on both SOTA (Gemini-\{2.5,3\}-pro, GPT-4o, DeepSeek-R1) and small-scale LLMs reveal that agents frequently succumb to injected biases despite having sufficient reasoning capabilities to identify the ground truth. These findings expose a significant reliability bottleneck in current agentic workflows, calling for specialized alignment strategies for LLM-as-a-Recommender. The complete code and evaluation datasets will be made publicly available upon acceptance.

Large language models (LLMs) provide a powerful reasoning backbone for speech understanding, but integrating continuous acoustic signals into a frozen LLM remains challenging. Existing speech-to-LLM interfaces typically operate at two extremes: either enforcing near-discrete token alignment, which benefits transcription but loses paralinguistic information, or learning unconstrained continuous representations, which can drift away from the LLM’s input space and degrade autoregressive decoding. In this work, we propose Convex Gate (C-Gate), a speech–LLM bridge that constrains all speech representations to lie within the LLM’s input embedding manifold with an architectural convex-hull constraint. Concretely, each frame is represented as a convex combination of token embeddings, ensuring compatibility with the pretrained LLM while preserving continuous expressivity. Across automatic speech recognition (ASR) and emotion recognition, C-Gate achieves strong joint performance, improving LibriSpeech WER by up to 48.7\% relative while matching or exceeding single-task emotion accuracy. Beyond performance, our analysis reveals a key insight: information is not carried by discrete token identities, but by time-resolved trajectories in the embedding space. Causal interventions confirm that both the trajectory structure and alignment to the pretrained embedding manifold are critical for performance. These results suggest that geometry, rather than token discreteness, is the fundamental design factor in speech–LLM interfaces, and provide a controlled regime for studying multimodal integration in frozen LLMs. We release the checkpoint, per-sample outputs, mechanism dumps, and intervention suite for replication.

Deep learning models are known for their ability to memorize training data. While memorization is often linked to risks such as privacy leakage and poor robustness, a highly influential claim by~\citet{feldman2020longtail} argues that memorization is actually \textit{necessary} for generalization. Their conclusion is based on the observation that removing points with high memorization scores reduces test accuracy. Upon closer inspection of their work, we uncover four critical flaws in the underlying methodology: \textbf{(1) sampling bias} in their approximation algorithm inflates memorization scores; \textbf{(2) high false positive rate} in their definition of memorization, leads to misclassification of non-memorized points as memorized; \textbf{(3) unprincipled thresholding} that resulting in an ill-posed problem; and \textbf{(4) data leakage} skews the test accuracy results. To address these limitations, we introduce a modifications for correctly identifying and evaluating memorization, including higher sampling rates, modifying the original memorization definition to reduce the false positive rates, proposing a method to identify a principled score threshold, and employing test datasets especially designed to avoid data leakage. Having accounted for these errors, our results show that, in contradiction to the original work, removing truly memorized points does not cause a drop in accuracy, and in most cases, improves test performance. These findings call into question the necessity of memorization in deep learning and highlight the importance of mitigating its risks.

Long-tailed out-of-distribution (LT-OOD) detection is often addressed with specialized training, including auxiliary out-of-distribution (OOD) data, abstention heads, contrastive objectives, energy losses, or gradient-conflict control. We show that these training mechanisms can obscure a simpler issue: frozen long-tailed representations may already contain useful OOD evidence, but raw Mahalanobis distance is distorted by frequency-coupled feature radius and poorly supported tail covariance. We propose \emph{Hyperspherical Pooled Mahalanobis} (\HPM), a post-hoc detector that normalizes features onto the unit sphere and replaces class-specific covariance with a pooled, ridge-regularized metric while keeping class means as semantic anchors. In CIFAR-LT experiments and an ImageNet-100-LT near-OOD boundary analysis, \HPM improves raw Mahalanobis scoring; for Prior-Calibrated ERM (PC-ERM), it raises AUROC from 46.49 to 85.67 on CIFAR-10-LT and from 50.40 to 78.35 on CIFAR-100-LT. This simple PC-ERM+\HPM pipeline also achieves the best Log Efficiency Score (LES; 3.08) on CIFAR-100-LT, retaining roughly 95\% of the best CIFAR-100-LT AUROC observed among the compared post-hoc scores at substantially lower training-time cost. These results argue for evaluating representation quality, detector geometry, and training complexity as separate factors in LT-OOD detection.

Backpropagation is the default learning rule for artificial neural networks and is often treated as the settled approach whenever differentiability is available. In this work, we revisit this convention from the perspective of sample efficiency. We introduce a unified vectorized feedback framework for loss-based and reward-based learning on computational graphs, in which synthetic gradients emerge as a natural alternative to backpropagation. We characterize the conditions under which synthetic gradients can achieve a lower gradient-estimation mean squared error than backpropagation. We construct examples illustrating that this sample efficiency advantage can be arbitrarily large. Experiments on contextual bandits and reinforcement learning tasks demonstrate the potential of our theoretical findings.

We ask whether agentic AI systems built for software engineering transfer to realistic hardware engineering. Existing hardware LLM benchmarks isolate sub-tasks but none jointly requires repository navigation, hierarchy-aware localization, Electronic Design Automation (EDA) executable verification, and maintenance-style patching. We introduce **Phoenix-bench**, a synchronized corpus of 511 verified Verilator instances from 114 GitHub repositories, each shipped with the developer patch, design-flow labels, fail-to-pass and pass-to-pass testbenches, and a Docker-pinned EDA environment so resolved-rate differences reflect agent behavior rather than toolchain availability. Using Phoenix-bench we run a uniform evaluation of four commerical agents and eight open-source agentic structures across four LLM backbones, plus two diagnostic interventions (file-level oracle localization and one round of testbench-log feedback). Three findings emerge. **Software and hardware are fundamentally different engineering tasks:** the same agent loses 37\% to 58\% from SWE-bench Verified to Phoenix-bench because hardware bugs propagate across parallel instantiated modules through signal flow rather than along a software-style call graph, and software-tuned agents stop at the symptom file instead of tracing back through the instantiation chain. **Failures concentrate on design control-flow / Finite State Machine bugs, verification testbench bugs, and hard cases** that demand cross-hierarchy signal-flow tracking and coordinated multi-file edits. **Localization granularity matters far more than localization itself:** a perfect file-level oracle yields only $+1.4$\% because the agent then breaks files that did not need editing, while a single round of test case feedback lifts resolved rate by $42$\% to $45$\% because the test case tells *where* the bug is and *what* the fix has to look like.

Reinforcement learning (RL) with verifiable rewards has driven recent gains in large language model (LLM) reasoning, but whether it creates new capabilities or merely sharpens existing ones remains debated, especially for compositional generalization, where models must combine learned primitives into novel multi-step functions. Using a controlled program synthesis domain over a typed domain-specific language (DSL) of 119 primitives with verifiable compositional depth and irreducibility, where the base model has zero capability prior to fine-tuning, we show on Qwen3-8B (and replicate within-family at 4B and 32B, and cross-family on Llama-3.1-8B) that compositional capability is established during supervised fine-tuning (SFT) and depends on irreducibility at two levels of SFT data: structural irreducibility of compositional exemplars and observational irreducibility, where prompt input-output pairs are chosen adversarially to rule out shallower programs. Neither alone is sufficient; even moderate contamination with structurally reducible exemplars substantially degrades performance, and full contamination yields zero pass@64. Given this SFT foundation, standard outcome-reward RL (GRPO, PPO) maintains but does not extend compositional performance, contrasting with prior results in settings with simpler nested function execution. Compositional capability is bottlenecked by structural properties of SFT data, before RL begins. Code release pending institutional IP review.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable effectiveness in general video understanding. Human motion understanding, a dominant topic in video analytics, however, still remains as a bottleneck due to the lack of explicit structural cues in 2D visual patches and the significant motion information loss caused by sparse visual temporal sampling. To address this, we propose Iris (Integrating representations of in-the-wild skeletons), a novel pose-augmented video MLLM tailored for human-centric motion understanding. Iris adopts an asymmetric dual-stream architecture, pairing the traditional sparse vision stream with a lightweight, high-frequency human pose stream to provide spatially structured and temporally dense kinematic signals. To enable pose-vision modality correspondence awareness, we design a cross-modal spatiotemporal binding mechanism, featuring a pose-structured 3D RoPE for precise temporal alignment and a person-grounded cross-attention module for explicit spatial visual grounding. Furthermore, to enable robust large-scale training, we construct an automated pose curation pipeline that organizes pose data into a delicate heterogeneous representation. Extensive experiments demonstrate that Iris achieves superior performance on multiple fine-grained motion benchmarks, while also maintaining highly efficient token computation overhead. Code will be available upon publication.

Text watermarking helps identify AI-generated content, but its effect on factual reliability remains underexplored. In this paper, we study watermarking hallucination: factual errors induced or amplified by watermarking even when the required evidence is present in the context and the unwatermarked model can answer correctly. Using a controlled RAG setting, we compare paired unwatermarked and watermarked generations under the same condition. Across representative watermarking methods, including KGW, SWEET, DiPmark, GumbelSoft, Gumbel-Max, and SynthID-style watermarking, we find that watermarking hallucination is widespread: watermarked outputs can remain fluent while introducing factual errors. We attribute this failure mode to two mechanisms: (1) direct token-level bias, which can suppress fact-consistent tokens, and (2) prefix-induced drift, which accumulates through autoregressive decoding and weakens later attention to factual context. Motivated by this analysis, we propose Fact-Preserving Token Intervention (FPTI) and Fact-Preserving Attention Intervention (FPAI), two plug-in interventions that can be integrated into existing watermarking methods to improve factuality. Our experiments show that combining FPTI and FPAI mitigates around 90\% of watermark-induced hallucinations while preserving fluency and comparable decoding efficiency. Overall, this work highlights factuality as a first-class criterion in watermark evaluation, alongside detectability and robustness, and calls for careful factuality validation before deploying watermarks in fact-critical applications.

We introduce **InvestigationWorlds**, an agentic environment for legal investigation. We build on an underused artifact of U.S. civil litigation: the summary judgment motion. This motion relies upon a record composed of real evidence exhibits, and results in a court-adopted hypothesis that is treated as ground truth for the purposes of deciding the motion. Each environment is built from a real U.S. Federal Court case retrieved from Public Access to Court Electronic Records (PACER) and augmented by an attorney-validated generation pipeline that synthesizes role-tagged documents around the original record. The resulting corpus admits multiple coherent factual readings, only one of which matches the court-adopted hypothesis. Evaluating on 100 cases, we find agents often commit to incorrect hypotheses despite retrieving relevant evidence, struggling to distinguish the court-adopted hypothesis from alternative hypotheses.

Humanoid teleoperation datasets are growing rapidly, with operator demonstrations retargeted onto a shared robot skeleton to train whole-body controllers and foundation models. We report a surprising property of this pipeline: retargeting normalizes body proportions but preserves operator movement dynamics (joint velocity profiles, ranges of motion, and coordination patterns shaped by the operator's physiology). On BONES-SEED (522 operators, 142K sequences), retargeted Unitree G1 trajectories support gender classification at 96.0% and operator re-identification at 97.2% Top-1; on operators never seen during training, gender holds at 83.4% and age and height regress within ±4.2 yr and ±5.7 cm. Partial correlation analysis reveals emergent, biomechanically interpretable structure: the signals are task-invariant across activity categories and hold across retargeting implementations. We introduce UNVEIL, a skeleton-aware spatiotemporal graph network to measure and interpret this effect, take initial steps toward operator anonymization, and ask the community: as teleoperation datasets scale, what should our data practices be? Code, models, and anonymized trajectories are at our project website https://project-unveil.github.io.

Post-hoc calibration aligns a classifier's predicted confidences with its empirical accuracy without retraining. An ideal calibrator should correct nonlinear miscalibration, scale gracefully to large label spaces, and preserve the original predictions; existing methods typically violate at least one of these properties---temperature scaling lacks expressivity, more flexible parametric alternatives introduce parameters that grow with the number of classes $C$, and other expressive methods do not preserve the rank ordering of class scores and may alter the predicted class. We propose \textbf{Invertible Logits Transformation (InvLT)}, which applies a learned scalar MLP $f:\mathbb{R}\to\mathbb{R}$ element-wise to the pre-softmax logits. Sharing $f$ across all logit dimensions makes the parameter count independent of $C$. Monotonicity of $f$---and hence preservation of the argmax prediction---is softly encouraged via a paired inverse network rather than enforced through the numerical integration required by prior monotone calibrators; this avoids their computational overhead while empirically preserving the original classification accuracy in every setting we evaluate. Across standard image classification benchmarks and a range of architectures, InvLT consistently outperforms a broad set of post-hoc baselines on standard calibration metrics.

*Steering vectors* (SVs) are widely used to influence the expression of concepts (e.g., truthfulness) in large language model outputs. A key assumption underpinning SVs is that they are linearly discriminative with respect to the concept: representations of texts that exhibit the concept are more aligned with the SV than those that do not, motivating shifts along the positive or negative SV direction to respectively promote or suppress the concept. In this work, we identify an *inverted detection-control* phenomenon in which some highly discriminative SVs that are aligned with positive representations can *consistently promote the opposite behavior*. We refer to such vectors as inverted-steering vectors (ISVs). We provide a geometric characterization of ISVs' effects, finding that steering along these directions systematically pushes representations in discriminative downstream heads as if the concept were absent, even prior to decoding. Motivated by this analysis, we propose an approach for distinguishing ISVs without requiring generation or associated response scoring. This enables targeted sign flips, which we use to improve a foundational detection-based steering pipeline via Inference Time Intervention (ITI). Our approach improves results in 27/30 experiments, ranging from +0.9\% to +138\%. We evaluate our findings on Gemma 3 12B, Qwen 2.5 14B, and Olmo 3 7B across 5 concepts. We will make our code publicly available upon acceptance.